2026 RoEduNet Conference: Networking in Education and Research

→ Europe/Bucharest
"Alexandru Ioan Cuza" University of Iasi and "Gheorghe Asachi" Technical University of Iași

"Alexandru Ioan Cuza" University of Iasi and "Gheorghe Asachi" Technical University of Iași

11 Carol I, Iasi, Romania, RO-700506
Nicolae ȚĂPUȘ (National University of Science and Technology POLITEHNICA Bucharest, Romania), Gheorghe DINU (Agency ARNIEC / RoEduNet), Lenuta ALBOAIE (Alexandru Ioan Cuza University of Iasi), Simona CARAIMAN (Gheorghe Asachi Technical University of Iasi), Florin Bogdan MANOLACHE (Carnegie Mellon University), Răzvan Victor RUGHINIȘ (National University of Science and Technology POLITEHNICA Bucharest), Paul GASNER (Agency ARNIEC/RoEduNet, Alexandru Ioan Cuza University of Iasi), Octavian RUSU (Agency ARNIEC/RoEduNet, Alexandru Ioan Cuza University of Iasi)
Description

25th RoEduNet Conference  
Networking in Education and Research

Nowadays, modern education and research activities are strongly dependent on a high-speed communication infrastructure and computer networks based on the newest technology. Design, implementation, management of such networks, together with development of new application fields are not possible without good knowledge of networking state-of-the-art.  

The 25th edition of Agency ARNIEC/RoEduNet's (Romanian Education Network) annual Conference is organised with the help of "Alexandru Ioan Cuza" University of Iasi and Gheorghe Asachi Technical University of Iasi as host partners, National University of Science and Technology POLITEHNICA Bucharest, Carnegie Mellon University, Pittsburgh, US, and IEEE Computer Society Romania Chapter, under the patronage of Ministry of National Education and Scientific Research of Romania, offers special opportunities for information exchange in computer networking: technical and strategic aspects, communication issues, and of course their applications in education and research.  

You are welcome to join in. Please register as a user on the conference website to receive updates on the 25th RoEduNet Conference.  

The conference will be organized as an in-person event with help from our partners and sponsors.

Thank you, and hope to see you soon at the anniversary edition of the RoEduNet Conference!!!

  • Thursday 19 November
    • Opening Session: Registration

      Opening Session

    • Opening Session: Welcome & Partners Keynotes

      Opening Session

      Conveners: Gheorghe DINU (Agency ARNIEC / RoEduNet), Dr Octavian RUSU (Agency ARNIEC/RoEduNet, Alexandru Ioan Cuza University of Iasi)
      • 1
        Welcome

        Organizers Welcome:
        RoEduNet Representative
        Alexandru Ioan Cuza University from Iasi Representative
        Gheorghe Asachi Technical University Representative
        Carnegie Mellon University Representative
        National University of Science and Technology POLITEHNICA Bucharest Representative

      • 2
        Bridging Central and Eastern Europe with Quantum-Secured Connectivity

        RoEduNet Keynote

        Speaker: Mihai CARABAȘ (Agency ARNIEC/RoEduNet, University Politehnica of Bucharest)
      • 3
        Building SOC maturity through SOCcare
        Speakers: Maria-Elena MIHAILESCU (National University of Science and Technology POLITEHNICA Bucharest, Romania), Stefan Dorin Jumarea (National University of Science and Technology POLITEHNICA Bucharest)
      • 4
        Security challenges in the post-Mythos era

        CISCO Keynote

        Speaker: George BOULESCU (CISCO)
      • 5
        Building the Modern Data Center for AI

        Dell Technologies Keynote

        Speaker: Daniel TANASE (Dell Technologies)
      • 6
        Smart protection in the new AI era

        FORTINET Keynote

        Speaker: Dan GRIGORICIUC (Fortinet)
    • 15:30
      Coffee Break
    • Network Security Room 1

      Room 1

      Network Security

      Convener: Mihai CARABAȘ (Agency ARNIEC/RoEduNet, University Politehnica of Bucharest)
      • 7
        Penumbra: An Adaptive Database Honeypot

        Honeypots designed for databases typically face a trade-off between being safe to run and being credible enough to engage real attackers, and few attempt to interpret what an attacker is doing in real time. Penumbra is a high-interaction adaptive PostgreSQL honeypot that addresses these limitations. It intercepts the PostgreSQL wire protocol, classifies every query in real time using a hierarchical Bernoulli Hidden Markov Model, and reacts by injecting SQL side-effects matched to the attacker’s inferred intent. The architecture is built as separated services connected over gRPC, with the actual database running inside hardened, ephemeral containers. The proof of concept reproduces a full exploit chain end-to-end, recording every step in a structured logging pipeline and capturing forensic artefacts when the session ends. Per-query latency overhead stays below 2 ms on average, keeping the honeypot indistinguishable from a direct database connection. The contribution shows that deep interaction, per-query adaptation and structured logging can be achieved together in one coherent and extensible system.

        Speaker: Andreia-Irina Ocanoaia (National University of Science and Technology POLITEHNICA Bucharest)
      • 8
        A Multi-Agent LLM and RAG Architecture for Evidence-Grounded Alert Triage in Security Operations Centers

        Security Operations Centers (SOCs) must triage large volumes of heterogeneous alerts while preserving decision quality, evidence traceability, and operational safety. This paper presents a controlled multi-agent architecture that combines large language models (LLMs), retrieval-augmented generation (RAG), evidence verification, and human approval for alert triage. The architecture separates alert normalization, contextual enrichment, retrieval, hypothesis formation, MITRE ATT&CK mapping, evidence verification, policy enforcement, and reporting, and deliberately augments the analyst rather than delegating autonomous containment or remediation. Retrieved documents and alert fields are treated as data rather than instructions, and claims must cite verifiable sources from a versioned, allowlisted corpus, mitigating prompt injection and RAG poisoning risks. We define four research questions concerning the effect of RAG, multi-agent verification, security controls, and operational cost, and specify a reproducible offline benchmark of 160 synthetic alert cases covering ten ATT&CK techniques, including dedicated prompt-injection and poisoned-RAG variants. The evaluation harness, benchmark integrity, and unit tests are validated end to end, and an executed severity-rule baseline yields chance-level performance (F1 = 0.50) on the paired test split, confirming that contextual evidence, not alert severity alone, is required for correct triage. The pre-specified protocol enables controlled comparison of single-LLM, RAG-augmented, and multi-agent verified configurations under identical conditions, with decision quality, evidence grounding, adversarial robustness, and latency or token cost reported together. The architecture and benchmark support transparent, reproducible evaluation of LLM-assisted triage in operational security contexts.

        Speaker: Mr Daniel Abotezătoaei
      • 9
        Phantom Peripherals: A Measurement Study of Leak Profiles and the Price of PCIe Device Impersonation

        PCIe hosts identify a peripheral by its configuration-space
        declarations, and the IOMMU authorizes its memory access unchecked.
        Off-the-shelf FPGAs can impersonate trusted peripherals; how far
        host-side telemetry pushes back is unmeasured. We build a ladder of
        progressively more capable imitations of one commodity network
        controller, scored by a behavioral detector over configuration
        semantics, MMIO timing, interrupt dynamics, and receive cadence, with
        a masked configuration hash and a platform-state control. The masked
        hash flags every rung while leaving genuine runs and cold-boot reads
        unflagged; no rung closes the configuration image, so the dynamic
        layer measures a structural residual. The strongest rung matches the
        genuine mean rates yet leaves a timing texture; the one band-matched loaded capture reads a
        straddle (0.63 and 0.68 on two passes). Masking the host-rewritten
        bytes moves the structural arm's false alarms to zero across the
        development-set cold-boot and relocation captures while every
        impostor stays flagged. In the
        matched-reference contexts the structural arm alone catches every
        impostor run at zero false alarms; the dynamic families need a
        condition-matched reference or a fresh profile, and under a fresh
        profile their per-family OR flags every relocated impostor at 4--5 of
        20 genuine alarms without a catch the structural term does not
        already make. The dynamic layer's measured deployment
        value is therefore non-positive, and its case rests on the
        hypothetical regime where the structural residual is closed, which
        this ladder never builds. Our profiles come from one unit per model,
        bounding the triage layer to deployments that profile devices at
        provisioning; the result prices this detection ceiling without
        reaching it and argues quantitatively for hardware-rooted measurement
        (CMA, SPDM, IDE, TDISP).

        PCIe security, device attestation, IOMMU, behavioral fingerprinting,
        FPGA emulation, hardware authenticity

        Speaker: Mr Rareș-Andrei Sărmășag (National University of Science and Technology Politehnica Bucharest, Faculty of Automatic Control and Computer Science)
      • 10
        Performance Evaluation of an Envelope-Encrypted Electronic Health Record Architecture with Managed Keys and Blind Indexing

        We investigate a reusable encrypted-record archi-
        tecture, using electronic health records as the implementation
        and evaluation case study. Electronic health record (EHR)
        platforms must protect sensitive fields across databases, logs,
        caches, and backup-oriented infrastructure without making or-
        dinary clinical transactions impractical. This paper presents
        and evaluates CryptoShred Health, a containerized EHR pro-
        totype that combines record-level AES-256-GCM encryption,
        per-entity envelope keys managed by a three-node HashiCorp
        Vault cluster, context binding through authenticated associated
        data, deterministic HMAC-SHA256 blind indexes for exact-match
        lookup, PostgreSQL ciphertext persistence, Redis read caching,
        and Kafka audit events. We separate isolated cryptographic costs
        from end-to-end behavior. On a Java 21 testbed, local AES-GCM
        processing averaged 22.18 μs per operation, whereas a full Vault-
        backed envelope operation averaged 10.68 ms. Short-duration
        k6 workload experiments across 1–500 virtual users produced
        96,820.5 requests/s for reads under an 85% Redis hit model at
        500 users (p50 4.92 ms, p99 12.40 ms), and 14,890.2 requests/s
        for encrypted ingestion (p50 31.80 ms, p99 84.10 ms). The
        results show that application-layer encryption and centralized
        key wrapping can coexist with low-latency reads when cache
        and key-management boundaries are explicit. We also identify
        the security leakage of deterministic indexes and plaintext caches,
        and delimit the conclusions imposed by synthetic data, brief
        measurement windows, and a prototype deployment.

        Speaker: Mr Robert HEVESI (Politehnica University Timisoara)
      • 11
        Design and Experimental Evaluation of a Low-Cost Virtualized Cyber-Range for SME Cybersecurity

        The digitalization of business worldwide is in full swing, bringing multiple undeniable benefits and at the same time major challenges, especially those related to cybersecurity. Multiple studies worldwide have assessed the cybersecurity challenges faced by small and medium-sized enterprises, the data shows a major impact, both in developed and developing countries. Although SMEs appreciate cybersecurity as very important for their business, the limited financial and human resources, legislation and fragmented information they hold, lead to insufficient implementation of security controls, which makes them vulnerable, unlike large enterprises that manage with complex information security management systems and constantly invest in the security of the digital environment. This limitation is addressed in the paper by proposing an Experimental Cyber Simulation Infrastructure (ECSI), based on a suite of open-source cybersecurity technologies such as Suricata and ELK stack, which would contribute to the quantitative assessment of security risk. The proposed network infrastructure is specific to SME network architectures, and has been segmented into several virtual networks: LAN, DMZ, SOC and Attackers, in order to test common SME cyber-attacks. We validated the infrastructure by testing various reconnaissance and flooding attacks at various levels of the OSI model. The experiments demonstrated that the infrastructure can identify distributed DDoS attacks and generate data for analysis and evaluation in real time. The proposed solution is low-cost and can be adopted by SMEs for cybersecurity testing and training.

        Speaker: Arina Alexei (Technical University of Moldova)
      • 12
        A Layerd Security Platform for Automated Incident Response in Kubernetes

        Kubernetes environments support scalable cloud applications, but introduce security challenges across infrastructure, workload deployment, runtime execution, service communication, and monitoring. Existing security solutions often address these domains individually, lacking an integrated framework that combines automated remediation with full-stack observability. This paper presents the implementation and evaluation of a layered DevSecOps security platform specifically for Kubernetes clusters in Amazon Web Services. The platform uses Infrastructure as Code and Amazon Elastic Kubernetes Service to provide reproducible infrastructure with isolated worker nodes. Open Policy Agent is used to enforce admission policies that reject privileged containers, root users, mutable image tags, and workloads without CPU or memory constraints. Falco monitors runtime behavior using the extended Berkeley Packet Filter driver and detects suspicious actions such as shell execution, sensitive file access, and Kubernetes service account token access. Falcosidekick forwards alerts to a custom response engine, which evaluates their priority and deletes affected pods when remediation is required. Linkerd secures service-to-service communication through workload identity and mutual transport layer security without requiring application code changes. External Secrets Operator integrates Kubernetes with Amazon Web Services Secrets Manager for external secret storage, while Prometheus and Grafana provide centralized observability. Controlled evaluation scenarios confirmed that insecure workloads were rejected, suspicious runtime activities were detected, security alerts were forwarded, affected workloads were remediated for selected alert priorities, and service communication was secured and successfully observed through the service mesh. These results demonstrate that integrating preventive controls, runtime detection, automated response, secret management, and observability provides practical defense in depth for Kubernetes environments.

        Speaker: Nicolae-Adrian Patrascu
      • 13
        Behavioural State Consistency in Horizontally Scaled Context-Aware Security Middleware

        Context-aware security mechanisms adapt their response according to the risk associated with each request, relying not only on immediate request-level signals but also on behavioral evidence accumulated over time, such as recognized device fingerprints, repeated failed authentication attempts, and dynamically generated blocklists. While this form of accumulated state is often kept in process memory, such an architectural choice is reliable mainly in single-process deployments and becomes increasingly fragile when the same middleware is horizontally scaled behind a load balancer to support production-level traffic. This paper investigates the security and performance implications of that mismatch between local state management and distributed execution. We show that the risk assessment performed by context-aware security middleware can be understood as the interaction of two components with distinct scaling properties: stateless per-request scoring, which remains independent of previous observations and can scale with the number of worker processes, and stateful behavioral accumulation, which becomes fragmented across workers when no shared coordination mechanism is used. As a result, when an attacker's requests are distributed across w worker processes, each process observes only approximately one w-th of the overall attack sequence, allowing malicious behavior to remain below thresholds that would have been exceeded in a single-process configuration. To evaluate this effect, we analyze an adaptive encryption middleware in three deployment scenarios: a single-process baseline, a multi-worker cluster that maintains behavioral state locally, and a multi-worker cluster in which behavioral state is externalized to a shared data store. Across configurations ranging from one to eight worker processes, and under attack traffic patterns designed to exercise threshold-based detection, we measure detection degradation, sustained throughput, and the latency overhead introduced by coordinated state access. The results reveal a practical trade-off that is frequently overlooked in the design of context-aware security systems: stateless risk scoring may be scaled with limited architectural friction, whereas stateful behavioral evidence requires explicit coordination whose performance cost must be incorporated into deployment decisions. Building on these findings, the paper offers practical design guidance for deploying context-aware security middleware in load-balanced production environments while preserving the consistency and effectiveness of behavioral detection.

        Keywords: context-aware security; horizontal scaling; distributed state; behavioral state consistency; adaptive encryption; web application security; security middleware

        Speaker: Mr Flavius Stasac (Doctoral School of Engineering Sciences, Department of Computer Science and Information Technology, University of Oradea, 410087 Oradea, Romania)
      • 14
        Detecting Illicit Activity in Bitcoin Cryptocurrency

        This study evaluates the statistical distinguishability
        of licit and illicit Bitcoin transactions using transaction-level fea
        tures and simple graph-based network indicators. The objective
        is to provide an interpretable evaluation based on statistical
        testing and effect-size measurement rather than to maximize
        predictive accuracy. A two-stage quantitative design combines
        Monte Carlo simulation with empirical analysis of the Elliptic
        Bitcoin dataset. A Mann–Whitney U test and rank-biserial
        correlation (RBC) were applied independently to each feature.
        The simulation recovered the predefined signal features under
        medium- and strong-separation conditions, including scenarios
        with added noise and class imbalance. In the Elliptic analysis,
        several transaction-level features produced strong separation
        between the licit and illicit classes. The transaction features
        had a mean absolute RBC of 0.2268, compared with 0.1223
        for the four graph indicators. Out-degree and fan-out showed
        small separation, with an absolute RBC of 0.190, whereas in
        degree and fan-in showed negligible separation, with an absolute
        RBC of 0.054. Temporal analysis showed that several transaction
        features maintained at least medium separation across most of the
        49-time steps. Within the evaluated feature sets, transaction-level
        features therefore showed greater observed statistical separation
        than the four simple graph indicators. However, these findings
        should not be generalized to richer graph representations or
        advanced graph-learning methods.

        Speaker: Ashenafi Abdi
    • Doctoral Symposium Room 2

      Room 2

      Convener: Florin LEON ("Gheorghe Asachi" Technical University Asachi of Iasi, Romania)
      • 15
        A comparison of Artificial Intelligence Risk Management Frameworks (AI RMF) for Privacy

        As the emerging technologies have moved fast in the past few years toward the adoption of Artificial Intelligence (AI) from software development to education, the existing Risk Management Frameworks (RMF) evolved as well. The gap created by the rapid adoption has generated a proliferation of Artificial Intelligence Risk Management Frameworks (AI RMF). Government bodies, standardization organizations, technology companies, cloud service providers, system integrators, cybersecurity companies are trying to provide a systematic approach to identify, assess, monitor, and reduce risks associated with AI systems. In this manuscript, we are looking at the most important AI RMF, their focus areas, and describe the coverage of each selected framework for data privacy and protection.
        Keywords— artificial intelligence (AI), risk management frameworks (RMF), cyber, security, privacy preserving, composable architecture.

        Speaker: Liviu-Mihai Popescu (Universitatea Lucian Blaga din Sibiu)
      • 16
        A Modular Open-Source Architecture for Intelligent Digital Mathematics Assessment Based on Moodle

        Digital mathematics assessment requires software ecosystems capable of integrating learning management systems, computer algebra systems, reusable educational resources, and automated assessment services into a unified, scalable environment. This paper presents the software architecture of an Intelligent Support System (ISS) for digital mathematics assessment, built on a Moodle-based ecosystem that integrates STACK, the Maxima Computer Algebra System (CAS), H5P, WIRIS, and a repository of Intelligent Digital Educational Resources (IDER). The architecture follows a decoupled, multi-tiered pattern that separates presentation, LMS orchestration, symbolic computation, and data persistence, and it supports parameterized item generation, Potential Response Tree (PRT) based automated grading, and standardized interoperability through Moodle XML and GIFT formats. The system was implemented on the MathMentor.ro platform and validated in authentic educational settings involving lower-secondary schools in Romania and a pilot deployment in the Republic of Moldova. The validation confirmed the technical feasibility of the architecture: average correction time per test dropped from about four hours to under 45 minutes, the experimental group improved by 1.02 points on average between pre- and post-test (0.47 points above the control group), and 92% of students together with all participating teachers rated the system as very useful. These results indicate that the proposed architecture provides a robust, reusable, and scalable open-source foundation for intelligent digital mathematics assessment.

        Speakers: Gabriela Cristina Branoaea (State University of Moldova), Prof. Andrei Braicov („Ion Creangă” State Pedagogical University of Chisinau)
      • 17
        An Open-Source Framework for Equal-Robustness Comparison of PID, RST, LQR, LQG and MRAC on a Fieldbus-Closed Stabilization Platform

        A beverage stabilization platform closes its attitude loop over FDCAN and reads a fused inertial measurement unit. Controller comparisons can confound structure with tuning and the robustness margin each design accepts. We constrain PID, RST, LQR, LQG and MRAC to an identical nominal peak sensitivity, leaving bandwidth free. Hardware identification uses swept-amplitude square waves, cross-validation and frequency-response screening. Simulation then includes the identified delay, conservative sensor noise and command limits. The 62.5 ms delay spans the bus round trip, drive loops and sensor fusion; control arithmetic is estimated at 0.1 µs. At matched peak sensitivity, pitch bandwidth spans a factor of 3.3. PID leads one of twelve cases and trails the case-best composite score by up to 66%. The constant-numerator model permits state reconstruction from measured and stored signals; estimator-free LQR achieves 35–37% more bandwidth than LQG. MRAC wins none of the twelve composite comparisons. The open-source toolchain reproduces identification, synthesis and simulation; the predicted controller ordering still requires hardware validation.

        Speakers: Mr Sebastian-Alexandru Matei (National University of Science and Technology POLITEHNICA Bucharest), Mr Valentin Pletea-Marinescu (National University of Science and Technology POLITEHNICA Bucharest)
      • 18
        A Dual-Path AI Architecture for Preliminary Lung Cancer Assessment from CT Images and Risk Factors

        Lung cancer remains a leading cause of cancer mortality, while timely assessment depends on access to imaging, specialized expertise, and efficient digital workflows. This paper presents a research prototype for preliminary lung-cancer assessment that combines two complementary machine-learning pathways within a modular web architecture. The first pathway classifies computed-tomography images as normal, benign, or malignant using an AlexNet-style convolutional neural network. The second estimates lung-cancer risk from questionnaire-derived demographic, behavioral, and symptom variables using logistic regression. The platform integrates an Angular client with a shared Flask model-inference service, separate Flask services for authentication, history, and notifications, PostgreSQL persistence, and Docker-based deployment. In the original proof-of-concept evaluation, the CT classifier achieved 98.8% accuracy with class-wise F1-scores of 0.98–0.99, while the questionnaire classifier achieved 94.4% accuracy and an F1-score of 0.95 for the lung-cancer class. The application reported model-processing time of approximately 150 ms. Because the original CT workflow used image-level splitting after augmentation, the image-classification result is treated here as internal proof-of-concept performance rather than patient-level clinical generalization. The work demonstrates how two heterogeneous AI pathways can share one model-inference service inside a modular, containerized platform while surrounding application services remain separately deployable. Patient-grouped validation, leakage-safe preprocessing, and external evaluation are priorities for future work.

        Speakers: Andrei Buhnici (National Univesity of Science and Technology POLITEHNICA Bucharest), Ms Denisa-Daniela Frimu-Pascu (National University of Science and Technology Politehnica Bucharest)
      • 19
        Audio Event Detection with Foundation Models: Zero-Shot and Fine-Tuned CLAP versus Template-Based Approaches

        Audio event detection often relies on labelled or cu-
        rated audio examples of the target event, which limits deployment
        when representative recordings are scarce or costly to obtain.
        We investigate whether foundation audio-language models can
        replace such audio references with natural-language descriptions,
        using cattle vocalization detection for welfare monitoring in
        Precision Livestock Farming as a case study. Using CLAP
        (Contrastive Language–Audio Pretraining), we score each 1-
        second window by its similarity to a positive description against
        negative descriptions of farm background, and then adapt the
        model with a language-preserving linear adapter that leaves the
        text interface intact. Both variants are evaluated within the
        detection pipeline of our previously published template-based
        detectors, on the same corpus and timestamps, with all
        thresholds selected using leave-one-animal-out cross-validation.
        Across 20 animals, zero-shot CLAP achieves 69.5% precision,
        66.7% recall, and an F1 score of 67.2%, compared with 63.7%
        for the audio-template YAMNet baseline; the paired difference is
        not statistically significant (p = 0.23). Thus, a natural-language
        description achieves comparable detection performance without
        requiring an audio exemplar of the target event. The adapter
        raises F1 to 70.6% by increasing precision to 81.8% at a
        small cost in recall, and is the only evaluated configuration to
        significantly outperform the baseline. Further experiments show
        that the tested bioacoustics-specialised backbone provides no
        improvement in this setting, while negative prompts carry most of
        the discriminative signal, highlighting background descriptions as
        an important target for prompt engineering in language-guided
        audio event detection.

        Speaker: Lavinia-Ștefana Dragne (National University of Science and Technology Politehnica Bucharest)
      • 20
        Toward a Digital Twin of an Industrial Cyber-Physical Sorting and Storage System for Engineering Laboratory Education

        Engineering education depends on laboratory work, yet access to modern industrial equipment is limited by acquisition cost, availability and safety constraints, so students rarely accumulate enough hands-on time with the systems they will encounter in industry. Digital Twin (DT) technologies offer a way to extend laboratory capacity, but the literature still treats digital modelling and real-time communication with physical equipment and pedagogical evaluation largely in isolation. Few studies report a complete collaborative system that is both connected to real hardware and assessed through explicit educational indicators.

        This work, part of an ongoing doctoral study, proposes a Digital Twin model for an industrial cyber-physical sorting and storage system. The reference platform built around a SIMATIC S7-1200 controller, three SINAMICS servo drives, toothed-belt and pneumatic axes, an IO-Link RFID identification system and a TP700 human-machine interface.

        The current prototype consists of a 3D geometric and kinematic model of the installation developed in Unity, together with a companion C# application that handles scenario logic and data exchange. Ongoing work addresses bidirectional communication with the physical controller through industrial protocol (OPC UA).

        Validation is in progress. The expected results include a replicable system architecture for connecting educational Digital Twins to industrial cyber-physical systems, and a comparative evaluation of physical, virtual and desktop-based training using task completion time, error rate, perceived usefulness and cognitive load as indicators.

        Speaker: Robert-Daniel Crivăți (Department of Automation and Industrial Informatics, National University of Science and Technology)
      • 21
        Design and Validation of an Extended FPGA-Based HC-85 Architecture for Computer Architecture Education

        This paper presents the design and validation of an extended FPGA implementation of the Romanian HC-85 microcomputer as a platform for computer architecture education. Implemented on a Digilent Nexys3 board, the system combines the open-source T80 Z80-compatible processor core with a 16 KB ROM containing the original BASIC interpreter, 48 KB of RAM, and control and peripheral modules written in Verilog. The design extends the original system with a UART interface operating at 115200 baud, allowing commands from a serial terminal to be processed through the keyboard matrix. Dual-port RAM provides independent processor and video access, removing the need to suspend processor execution during display memory reads. VGA output and a PS/2 keyboard interface support interaction with the reconstructed system. Validation proceeded from individual subsystem tests using a dedicated ROM to integrated testing with the original HC-85 BASIC interpreter. Tests confirmed successful startup, BASIC command execution, operation of all 40 keyboard keys and tested modifier combinations, serial input, and color display functionality. The implementation occupies 2,077 lookup tables and 668 registers, while using all 32 available block RAM resources. Its modular structure provides a concrete basis for laboratory exercises on processor buses, memory organization, interrupts, and peripheral protocols, connecting digital hardware design with the operation of a complete computer.

        Speaker: Nicoleta-Alexandra Trancă (Universitatea Națională de Știința si Tehnologie Politehnica din București)
      • 22
        MCU-MDT: A Lightweight, Portable UART Debugging Monitor for Resource-Constrained Microcontrollers

        Debugging resource-constrained microcontrollers often requires dedicated hardware probes or application-specific instrumentation, complicating development on low-cost embedded platforms. This paper presents MCU-MDT, a portable software debugging monitor that provides memory inspection and execution control through an existing UART connection without requiring a dedicated debug probe. The architecture separates a Python host application, responsible for command validation and device metadata interpretation, from a compact C firmware library that processes fixed-length, CRC-protected binary packets. A twelve-function hardware abstraction layer isolates platform-specific operations, allowing the core firmware to remain unchanged across AVR and STM32 targets. The monitor supports memory and peripheral-register access, cooperative software breakpoints, and masked software watchpoints. Cooperative breakpoints suspend application progress while preserving communication with the host. Evaluation covers ATmega328P, STM32F030F4, and STM32F103-compatible configurations through hardware testing and complementary simulation. With size optimization enabled, polling builds occupy 2,656–2,894 bytes of Flash and 248–263 bytes of static SRAM, without dynamic memory allocation. Reported memory-read round-trip latency is approximately 18.8 ms at 19,200 baud, indicating that serial transmission dominates transaction time. These results demonstrate the feasibility of a common debugging interface across constrained microcontroller platforms with a small memory footprint, supporting embedded development and teaching environments where dedicated probes are unavailable.

        Speaker: Nicoleta-Alexandra Trancă (Universitatea Națională de Știința si Tehnologie Politehnica din București)
    • Security & Resilience in Cyber-Physical Systems * Quantum TC & SANE Room 3

      Room 3

      UAIC
      Convener: Cristian-Ioan BABAN (Alexandru Ioan Cuza University from Iasi)
      • 23
        Analysis of vulnerabilities of automotive CAN networks

        The Controller Area Network (CAN) protocol remains the dominant in-vehicle communication backbone, yet its design---conceived in the 1980s without encryption, authentication, or source identification---exposes modern Software-Defined Vehicles to critical cyberattacks. Physical testing is costly, hazardous, and non-reproducible, while purely virtual testbeds lack real-time physical feedback. This paper presents CARLA-CAN, an open-source, hybrid Software-in-the-Loop (SiL) platform that bridges the CARLA 3D driving simulator with Linux SocketCAN virtual buses through a high-performance, multi-threaded C++ routing core. The platform enables closed-loop evaluation: CAN frames injected by an attacker propagate through the virtual bus, are decoded by the routing core, and instantaneously alter the 3D vehicle's physical dynamics---steering, braking, engine state---providing both network-level traces and visual confirmation of attack impact. We implement and validate a comprehensive suite of seven attack scenarios, including DoS flooding, stealth flooding, speed spoofing (chaotic and proportional), engine compromise, body system fuzzing, and a novel state-aware engine attack. Experimental benchmarks demonstrate median passthrough latencies of 355 us across 91,000 CAN frames and sub-1% packet loss at normal automotive frequencies (10--100 ms), confirming real-time fidelity. A systematic comparison with the VitroBench hardware testbed validates the platform's suitability for IDS training dataset generation, educational demonstrations, and rapid prototyping of automotive cybersecurity defenses

        Speaker: Radu Constantinescu
      • 24
        Protocol-Agnostic Virtual Testbed for Next-Generation Automotive Network Prototyping

        Modern vehicles combine communication protocols with different timing, bandwidth, and dependability characteristics, complicating early network prototyping. This paper presents a virtual testbed coupling the CARLA driving simulator with four firmware-defined electronic control units (ECUs). Each ECU executes MicroPython firmware on an emulated RP2040 microcontroller and communicates through an emulated CAN/CAN-FD controller and synchronized shared bus. A gateway translates vehicle state and driver commands into network messages and returns ECU reports to CARLA. The ECU roles mirror those of a separate Teensy-based physical reference platform, which is not evaluated here. The current executable communication implementation supports Classic CAN and CAN-FD; LIN, FlexRay, and Automotive Ethernet/TSN remain architectural extensions. Source-code inspection identifies the implemented communication paths and observation capabilities, including message injection, packet inspection, and waveform visualization. Preliminary runtime measurements demonstrate a frame throughput of 7,916 frames/s (0.88 Mbit/s equivalent processed bandwidth), zero frame loss, and a median write latency of 417.6 $\mu$s, alongside an emulation progression ratio of $\approx$0.02$\times$ real time across the RP2040 nodes. The result is an extensible research prototype for scenario-driven automotive network experimentation and subsequent validation.

        Speaker: Ciprian-Mihail Popescu (National University of Science and Technology POLITEHNICA Bucharest)
      • 25
        ZK-Whistle: A Decentralized, Trust-Minimized Architecture for Confidential Whistleblowing

        Whistleblowers who expose corruption, fraud, or abuse face retaliation that ranges from professional destruction to physical danger, yet the tools available to protect them remain weakest at the two moments that matter most: while sensitive evidence sits unpublished and vulnerable to seizure, and when a credible source wishes to be rewarded without revealing who they are. Centralized submission platforms concentrate exactly the trust that a threatened source cannot afford to grant, while naive on-chain approaches leak metadata that deanonymizes participants. This paper presents ZK-Whistle, a decentralized architecture that closes these gaps by composing four production-ready cryptographic primitives. A cryptographic dead man's switch releases withheld evidence automatically if a source goes silent, inverting the adversary's incentive to detain them. Threshold key management based on multi-party computation replaces trusted hardware as the custodian of decryption keys, so that no single party—and no compromised processor—can unlock a payload ahead of its condition. Zero-knowledge transport-layer-security (zkTLS) provenance lets a source prove a real-world credential without revealing their identity, and stealth addresses make financial rewards unlinkable. A central contribution is a critical re-evaluation of the trust model: we document why a Trusted Execution Environment reintroduces a hardware-rooted single point of failure and replace it with a threshold network whose security rests on mathematics. The system is realized as three audited-style smart contracts and a typed client, validated by 91 automated tests and deployed to a public test network.

        Speaker: Giorgiana Vlăsceanu (University Politehnica of Bucharest)
      • 26
        LLM-assisted bench penetration testing of aftermarket Android head units: a framework focused on the CAN-decoder boundary

        Abstract: Millions of vehicles use aftermarket Android head units exposing USB, Bluetooth, Wi-Fi, ADB, and CAN interfaces, weakly hardened and sitting one hop away from safety-sensitive buses. In this paper we report on a controlled bench penetration test of a popular aftermarket infotainment unit (Rockchip RK3399, Android 10) installed in a Ford Focus Mk2, using only commodity hardware: a Raspberry Pi implant, a software-defined Wi-Fi adapter, and a SocketCAN interface. Following a PTES-derived method aligned with ISO/SAE 21434 and scored with DREAD, we demonstrate a full attack chain that uses no zero-day: associating with the unit's Wi-Fi hotspot over the stock default password, gaining unauthenticated ADB root on a user debug build, and extracting plaintext credentials, including the previous owner's saved Wi-Fi passphrases and 59 cached Google OAuth tokens with mailbox, contacts, and location scopes. All findings follow responsible disclosure. We contribute a repeatable, low-cost method, a sixteen-finding risk register, and a human-supervised, locally hosted LLM-assisted testing framework, reported with both contributions and failure modes

        Speaker: Vasile-Catalin Mihacea (Lucian Blaga University of Sibiu)
      • 27
        One Ladder or Four Participation Contracts? Public Willingness for Networked Citizen Science in the EU27

        Research infrastructures invite citizens to learn, donate personal data, lend computing capacity, or join research, yet these activities are often treated as one engagement continuum. Using Eurobarometer 102.1 (2024; EU27 N = 26,500), this paper tests that assumption. Population-weighted willingness was 15.0% for data donation, 14.2% for a science and technology course, 11.3% for active co-research, and 7.1% for volunteered computing. The modes were not nested: 67.2% selected none, 22.2% one, and 0.8% all four; the largest conditional overlap was 40.8%. Adjusted source-profile models showed the strongest contrasts among scientific-journal users for course-taking (+11.0 percentage points) and co-research (+9.4), while social-network/blog contrasts remained below one point. Time, knowledge, information, opportunity, inclusion, and privacy were common barriers within the willing pools. The contribution is a service-portfolio model for NREN-linked citizen science: federated identity can reduce repeated onboarding, but each mode needs a distinct authorization, support, and evaluation contract.

        Speaker: Dr Simona-Nicoleta Vulpe (Faculty of Sociology and Social Work, University of Bucharest)
      • 28
        High-Throughput JPEG Restoration via Mean Flows

        Recent JPEG restoration systems achieve strong visual quality with large models, but their computational cost often limits efficient on-device deployment. We present a 65M-parameter generative restorer that improves LPIPS over three regression baselines in seven of nine held-out conditions on LIVE-1, Urban100, and DIV2K validation, while sustaining 8.05 images/s at 720p on a single RTX 3090. Trained from scratch, the model replaces learned VAE encoding and decoding with an exactly invertible two-level Haar transform. A rank-enhanced linear-attention DiT predicts clean wavelet-domain residuals and internally estimates compression severity. A MeanFlow objective enables inference with one or two network evaluations without distillation. Although large pretrained priors remain stronger on several quality metrics, our approach offers a compact, high-throughput alternative for resource-constrained JPEG restoration.

        Speaker: Mihai-Alexandru Radu (Faculty of Physics, "Alexandru Ioan Cuza" University of Iași)
      • 29
        Atoms Ground State Simulation Using Quantum Algorithms

        We analyze the ground state energy of atoms using quantum algorithms, primarily Variational Quantum Eigensolver (VQE), implemented on a quantum computer. The goal is to assess the accuracy of quantum algorithms in modeling electron correlation in many-electron systems, with a practical application on Helium.
        Our expected outcomes: a working VQE implementation that computes a good approximation of the helium atom’s ground state energy; insights into how well current quantum computers can handle electron correlation in simple systems.

        Speaker: Cristian-Ioan BABAN (Alexandru Ioan Cuza University from Iasi)
    • Welcome Dinner: Welcome Cocktail Sala Pasilor Pierduti

      Sala Pasilor Pierduti

      https://maps.app.goo.gl/oysnWRpnXrzTDAzJ7

  • Friday 20 November
    • Networking in Education and Research Room 1

      Room 1

      Convener: Emil-Ioan CEBUC (Agency ARNIEC/RoEduNet, Technical University of Cluj-Napoca)
      • 30
        Students' Perceptions of Teacher Support in Developing Digital Competences: An Exploratory Study Aligned with DigCompEdu Area 6 and DigComp 3.0

        Digital competence has become a strategic priority for higher education as the updated European Digital Competence Framework for Citizens (DigComp 3.0, 2025) explicitly integrates artificial intelligence, cybersecurity, data protection, and digital well-being. This exploratory study examines students’ perceptions of the support provided by university teachers in developing their digital competence, focusing on Area 6 (Facilitating Learners’ Digital Competence) of the DigCompEdu framework. A questionnaire originally developed on the basis of DigCompEdu/DigComp 2.2 and subsequently mapped onto DigComp 3.0 was administered to a convenience sample of 394 students from five higher education institutions in the Republic of Moldova between 2025 and 2026. Descriptive results reveal an uneven facilitation profile: comparatively strong support for information literacy, communication, and content creation (48–77% of respondents reporting high support), but markedly weaker support for collaborative content editing (29.7%) and engagement with artificial intelligence (31.7%) and augmented/virtual reality (5.1%). Student use of AI and teacher-led integration of advanced technologies were each reported by only around one third of respondents, although these indicators measure different sets of technologies and should be interpreted separately, while cybersecurity and data-protection concerns ranked among students’ top reported negative aspects of digital learning (19.0%). Given the convenience, interinstitutional sample and the unavailability of respondent-level data, the findings are interpreted descriptively and should not be generalised to the national student population. The study contributes an interinstitutional, Moldova-specific perspective explicitly correlating DigCompEdu Area 6 with DigComp 3.0, and outlines directions for strengthening AI literacy, cybersecurity awareness, and digital well-being in higher education pedagogical practice.

        Speakers: Prof. Andrei Braicov (Ion Creangă State Pedagogical University of Chișinău), Dr Dinu Țurcanu (Technical University of Moldova), Dr Ilona Popovici (Technical University of Moldova, “B. P. Hasdeu” University Center in Cahul)
      • 31
        ARGUS: Intelligent ocular analysis system for detecting fraud in online examinations

        This paper presents an intelligent software system for detecting potentially suspicious behavior during online examinations using video recordings captured with a standard webcam. The proposed approach applies computer vision techniques to extract behavioral indicators such as gaze direction, head pose, and blink rate. To improve robustness and reduce sensitivity to minor variations, the extracted features are preprocessed before behavioral anomaly detection.

        The system uses the Isolation Forest algorithm to identify time intervals in which a participant’s behavior deviates from their normal patterns. The results are presented through a web-based interface that provides a risk score and highlights potentially suspicious moments, allowing examiners to focus on relevant sections of the recording rather than reviewing the entire video.

        The system was evaluated using recordings collected in real examination settings during the AcadNet Applied Informatics Olympiad. The results show that the proposed approach can distinguish between normal and potentially suspicious behavior while maintaining a low false-positive rate during regular participant activity. Usability testing with users unfamiliar with the application also indicated that the interface is intuitive and easy to use. The findings demonstrate the potential of AI-assisted behavioral analysis as a practical decision-support tool for improving the scalability and efficiency of online examination review.

        Speaker: Cristiana Manole
      • 32
        A.R.G.U.S. Intelligent System for Detecting Plagiarism in Online Examinations

        The increasing adoption of online assessment has introduced new challenges for maintaining academic integrity, particularly with the widespread availability of generative AI tools, online resources, and digital communication platforms. This paper presents an intelligent multimodal system for detecting potentially fraudulent behavior during online multiple-choice examinations.

        The proposed solution analyzes both screen recordings and webcam video. Screen recordings are examined to identify access to unauthorized resources, such as AI-based tools, examination-related materials, and communication platforms, while webcam recordings are used to detect prohibited objects, including mobile phones and headphones, as well as the presence of other individuals. The screen-analysis component uses anomaly detection based on a KNN memory bank built with DINOv2 representations, while the video-analysis pipeline combines YOLOv8 for object and person detection with CLIP-based zero-shot classification.

        The system generates structured reports containing risk indicators and highlighted suspicious moments, supporting educators in the review process while leaving the final decision to the human evaluator. The solution was validated in national-level Informatics and Applied Informatics Olympiads involving more than 2,000 high-school students. The results demonstrate the effectiveness of the proposed approach in identifying anomalous and potentially fraudulent behaviors and highlight its potential as a scalable, AI-assisted solution for supporting academic integrity in online assessment.

        Speaker: Adrian-Ștefan Moroianu
      • 33
        An Auditable, Evidence-Based Multi-Level Framework for Automated Assessment of Cybersecurity Hands-On Labs

        Manual evaluation of terminal logs in cybersecurity labs is difficult for large cohorts, especially in the absence of a centralized infrastructure. Checking the presence of a command in the terminal history cannot confirm that the task was completed. We propose an auditable evaluation framework, based on 4 types of evidence extracted from logs: command, argument, verification and output. The evidence was combined cumulatively on 3 levels of confirmation: syntactic (L1), semantic (L2) and confirmed technical result (L3), corresponding to the log fragment. The framework was tested on 899 eligible logs obtained from 6 heterogeneous labs: network scanning, access control, cryptography, digital signatures, VPN and firewall. For the direct comparison, 550 logs containing the full session transcript were used, with a total of 7192 decisions made, with average scores of 85.34% (L1), 58.59% (L2), and 39.68% (L3). A difference of 45.66 percentage points and a very large effect size (Kendall's W=0.880) was recorded. Procedural tasks had higher scores than multi-stage ones. The framework is sensitive to partial log loss or truncation and invariant to ANSI formatting tests, after data normalization, processes 8.71 logs/second and supports automatic initial cohort assessment, in the absence of centralized collection infrastructure.

        Speaker: Arina Alexei (Technical University of Moldova)
      • 34
        Arena AcadNet - Isolated and Predictable Evaluation

        Arena AcadNet is the online judge of the Romanian National Olympiad of Applied Informatics, Software Interoperability section, a contest that hands the competitor a faulty or incomplete program to repair instead of a blank editor. The platform reported in earlier work was a server rendered monolith whose contest rules lived in its views and whose evaluator started a container per submission. It ran, but it had stopped growing: whole release cycles went into asset delivery rather than features, exports inherited a display bug because they scraped the rendered page, and the language field shown to problem authors had been decorative for three years, since the evaluator knew one compile command. Both halves were rewritten. The interface is now a single page application served same origin by a versioned JSON API, with verdicts and the leaderboard pushed rather than polled. The evaluator became worker processes running untrusted code in isolate, the sandbox built for the International Olympiad in Informatics: control group memory accounting, a process cap, an isolated network namespace, an unprivileged user, and CPU time rather than wall clock deciding a time limit, where the old one gave contestant code root, outbound network and a writable host directory. An unchanged wire protocol let the two halves be deployed independently. On the same host and problem the new evaluator sustains roughly three and a half times the throughput with a tail roughly five times tighter, and the pipeline is now language neutral: C++ and Python problems run in production.

        Speaker: George-Andrei Nita-Pirvu (Universitatea Politehnica Bucuresti)
      • 35
        A Bare-Metal Live OS for Secure Examinations

        Most tools that protect an online exam are applications running on an operating system the student controls. A lockdown browser or a remote virtual machine restricts only its own window, thus the host system underneath stays untouched, and a student stepping outside it leaves no trace. We drop that assumption rather than hardening it. The machine boots from a USB stick into a Debian live image with a read only root and an overlay discarded at power off. Therefore, the exam runs on bare metal and nothing survives a reboot. What it may reach is decided by a default deny kernel firewall, not a browser URL filter. One build tree yields two images: a restricted environment for written exams in a learning management system, and a supervised environment for practical Linux exams, where the student does get a shell, sudo and apt, but every privileged invocation is policy checked and confined by AppArmor, ensuring local root access cannot compromise the supervision stack. Each seat receives the same versioned workspace on joining, with per seat variants. Supervision is local first, the server governing an exam runs in the room and works without internet access, while a cloud console mirrors it read-only. Telemetry keeps timing and context rather than content, with a short list of exceptions disclosed to students beforehand, and it records no video, audio, screenshots or keystroke text.

        Speaker: George-Andrei Nita-Pirvu (Universitatea Politehnica Bucuresti)
      • 36
        Transform-Based Configuration Management on a 400-Switch Multi-Vendor Campus

        Our earlier framework configured a campus network by sending commands to switches task by task. It was slow, its dry run listed as changes lines the switch already had, and each switch was half-configured during a run. On the same campus of some 400 switches, we replace it with a configuration transform. Instead of trying commands on the switch, we record each platform's syntax in advance, read each switch's configuration and state, and compute its complete configuration locally. The transform rewrites the lines the site policy manages and copies every other line unchanged and in place, so every pending change can be reviewed as a diff. With configure replace, the switch computes the difference and applies it. On IOS from 12.2(55) and on NX-OS, the switch undoes the change if the session is lost, and a scheduled reload recovers a switch the change left refusing logins. A web application shows each pending change, ties approval to the exact file, and checks an edited line with the switch's parser without entering it. On Cisco switches the transform refuses to drop the management path or every login. Seven verification layers, including Lean and TLA+, check the transform and protocol. On ten leaf switches, the median read-only check falls from 227 s to 18.8 s, and the median real deployment from a modelled 592 s to 69.5 s. One rollout deployed 313 switches in 46 minutes without a failure. The cost is a platform table covering 23 hardware families and 19 software versions.

        Speaker: Andrei-Alexandru Bleorțu (National University of Science and Technology POLITEHNICA Bucharest)
    • Doctoral Symposium Room 2

      Room 2

      Conveners: Florin Pop (University Politehnica of Bucharest, Romania / National Institute for Research & Development in Informatics – ICI Bucharest, Romania), Vasile MANTA (Gheorghe Asachi Technical University of Iasi)
      • 37
        From Dark-Web Text to Threat Intelligence: What Works, What Fails, and What Translation Breaks

        Pipelines that turn security text into cyber threat intelligence (CTI) commonly translate non-English posts into English and then extract indicators with regular expressions and names with dictionaries. We show that the first step undermines the second. Machine translation silently alters file hashes, IP addresses and URLs while leaving them well-formed, so an extractor applied to translated text reports indicators that never existed. The effect appears in every open translation model we tested, from MarianMT to NLLB-200 and M2M100, is not fixed by larger models or beam search, and is only partly avoided by masking, whose success depends on the placeholder style the model happens to copy. Indicators therefore have to be taken from the source text. Using our own multilingual dark- web pipeline as the test case, we then measure the rest of the recipe against five CTI corpora with gold entity labels, under one protocol and on one GPU. Regular expressions are as good as anything we trained. Dictionaries of malware, tool and actor names miss most of what the corpora contain, and a small encoder fine-tuned for one or two minutes finds several times more entities than they do, while 8B language models sit in between at far higher cost. Models trained on one corpus lose much of their accuracy on another, because the corpora disagree about what a tool, a malware family or an indicator is. Keyword- based relevance filters look good against random Wikipedia text and fail against technical text about other topics, which a zero- shot entailment model handles without training. Along the way we restate our earlier headline figures on a defensible footing, including the overlap with the CISA KEV catalogue, which shrinks from a quarter to under three percent once the KEV feed itself is left out.

        Speaker: Traian Becheru (Faculty of Automatic Control and Computers National University of Science and Technology Politehnica Bucharest)
      • 38
        Does Prompt Design Affect MITRE ATT&CK Mapping from CTI?

        Large language models (LLMs) are increasingly used to map Cyber Threat Intelligence (CTI) text to MITRE ATT&CK techniques, but they can also return techniques that are not supported by the source text. We investigate whether prompt design can reduce these unsupported predictions. Qwen3-32B maps the same 228 annotated CTI paragraphs under three prompting strategies: direct mapping, evidence-constrained mapping that requires a verbatim supporting span, and behavior-first mapping that describes the adversary behavior before predicting the technique. The prompts are finalized on a validation subset before evaluation on the held-out test split. We find no statistically significant difference in gold-mapping recovery between the three strategies (exact McNemar test, p > 0.5 for all pairwise comparisons), while the unsupported prediction rate remains between 0.89 and 0.91. The evidence constraint is followed correctly: 43 of the 44 correct mappings under this condition cite text that appears verbatim in the paragraph, with a mean token Jaccard of 0.55 against the annotated evidence span. However, this additional grounding does not improve technique-level accuracy. Recall is instead strongly associated with how the technique is expressed in the text: lexically signaled techniques are recovered 2.3 to 3.4 times more often than behavioral mappings across all three prompts. Requiring evidence improves the verifiability of ATT&CK mappings, but does not by itself improve their correctness.

        Speaker: Ebru Resul
      • 39
        Catalogue-Scale, Style-Aware Furniture Recommendation for AR Replacement Using Vision-Language Models and Dual-Channel CLIP Retrieval

        Furniture replacement in augmented reality (AR) requires both isolating the target object within a photorealistic 3D reconstruction and retrieving visually and stylistically coherent replacements from a large catalogue of 3D assets. This paper presents a context-aware furniture recommendation system built on a curated catalogue derived from the 3D-FUTURE dataset. From the 16,563 models available, we construct a balanced catalogue of 2,400 furniture objects spanning six residential super-categories (Chair, Table, Sofa, Bed, Cabinet/Shelf/Desk, and Lighting), with 400 models per category. Each object is rendered from three canonical viewpoints at 512×512 resolution, and more than 40 heterogeneous style annotations are consolidated into nine canonical style labels to enable consistent compatibility reasoning. A task-specific evaluation of seven open-source vision-language models (VLMs) on closed-set interior style classification shows that all functional models perform within sampling uncertainty of one another, and Qwen2.5-VL-7B is selected for its lowest latency (1.53 s/image). The selected VLM generates natural-language captions for 2,347 catalogue objects, describing shape, material, style, and plausible interior context. Captions and thumbnails are indexed with CLIP ViT-L/14 into a dual-channel (text and image) embedding space. At inference, a user click in the reconstructed scene triggers a VLM query that identifies the room style, object category, and a replacement description, which is matched against the index with a tunable visual-channel weight and re-ranked under a hard category constraint and a soft style-compatibility penalty. The offline pipeline processes the full catalogue in under 15 minutes; the index achieves self-consistency Recall@1 of 1.0 (text) and 0.997 (image), and text-dominant retrieval reaches Category Precision@5 of 0.97. The results show that a catalogue-scale, style-aware furniture recommender for AR can be assembled entirely from publicly available 3D assets and open-source vision-language models.

        Speaker: Mr Dan Toderici (National University of Science and Technology POLITEHNICA Bucharest)
      • 40
        When Is a Network-Service Switch Complete? Continuity Failures in EU27 Bundle Migrations

        Provider switching is a planned transition between administrative and technical domains, but order completion can conceal service interruption, partial activation, and unresolved billing. This paper measures those failures among 4,638 EU27 households that changed a bundled-service provider within two years in Eurobarometer 94.3. Overall, 45.7% reported a named problem, 29.6% a service problem, and 20.0% a cost problem. Temporary loss for at least one day (17.3%) and services taking days to work fully (16.2%) align directly with continuity and recovery intervals. Adjusted service-problem probability did not rise monotonically with bundle breadth. Bundles containing television, music, or video had 4.2 percentage points more problems overall (95% CI 0.2 to 8.2), but no increase in the service composite; item analysis instead found 4.0 points more delayed activation (1.1 to 6.9) and 2.0 points less lost content (0.6 to 3.5). The data therefore reject component count as a risk score and locate the content-bundle signal in readiness delay rather than generic service failure. We translate the estimates into an instrumented two-provider handover with measurable cutover, recovery, and commercial-closure states, applicable to household access and institutional/NREN service migrations.

        Speaker: Anania Iordache (Doctoral School of Sociology, Faculty of Sociology and Social Work, University of Bucharest)
      • 41
        Nobody Read This Code: What Code Review Must Become When Software Is Written Faster Than It Is Read

        Code review assumes that changes arrive slowly enough to be read by another engineer before deployment. Large-language-model code generation can increase change arrivals without increasing review capacity. This position paper asks when human diff reading may cease to be the primary defect gate without weakening assurance. We synthesise evidence on modern and AI-assisted review, agent-authored pull requests, and production verification, separating three functions of review: defect detection, design deliberation, and shared awareness. The original contribution is a control-retirement protocol that integrates (i) a reviewer-capacity condition, (ii) a defect-class map from review concerns to candidate verification stages, and (iii) local comparative tests, process preconditions, and reversal triggers. The protocol requires measured generator–reviewer joint misses, establishes behavioural verification during progressive delivery before reducing human review, and externalises design rationale and ownership into versioned, testable artefacts. A qualitative maturity–cost–residual-risk map makes clear which replacements are established and which remain proposals. The result is a falsifiable, per-defect-class migration framework, not a claim that human review can already be removed universally.

        Speaker: Mr Vlad-Stefan Dieaconu (National University of Science and Technology POLITEHNICA Bucharest)
      • 42
        From-Scratch SVD Pipeline for Detecting Mispriced and Anomalous Properties in a Large Real-Estate Market

        We study whether a Singular Value Decomposition implemented from scratch using Householder tridiagonalization, Wilkinson-shifted QR iteration, Givens rotations and deflation can support large-scale real-estate analysis comparably to standard library routines. Using New York City property sales, we construct a 28,380 x 406 feature matrix combining numerical attributes with one-hot encoded categorical variables. The custom decomposition closely matches NumPy, recovering the same numerical rank of 387, with a relative reconstruction error of 5.2 x 10-11 and a maximum singular-value difference of 2.4 x 10-8, while requiring 28 s compared with 0.45 s for NumPy. Low-rank analysis shows that categorical one-hot features substantially reduce compressibility: 319 of 406 components are required to preserve 90% of the energy, compared with only 4 of 6 components for numerical features. SVD-based regression explains 72.7% of price variance, while Ridge regression on log-prices reaches an R² of 0.78. For anomaly detection, SVD reconstruction error and K-Means distance identify highly similar outliers, largely because both are strongly influenced by row norm, whereas Isolation Forest detects a substantially different subset. These results show that a from-scratch SVD implementation can provide both accurate numerical decomposition and useful insights into pricing structure and anomalies in large real-estate datasets.

        Speaker: Ms Maria-Alexandra Timbuș (University of Bucharest, Romania)
      • 43
        Benchmarking Classical Adversarial Attacks on Modern Vision Architectures

        Classical white-box adversarial attacks were designed and evaluated mostly on convolutional networks from 2015 to 2018. Modern architectures such as Vision Transformers and ConvNeXt have different inductive biases, and it is unclear whether the old attacks work equally well against them. We present a reproducible benchmark that runs FGSM, BIM, PGD and AutoAttack against four ImageNet-pretrained architectures (ResNet-50, EfficientNet-B0, ViT-B/16, ConvNeXt-Tiny) on Imagenette and report robust accuracy over six L-infinity budgets from 1/255 to 32/255, on 200 validation images for the gradient-based attacks and 500 for AutoAttack. Iterative attacks break every model: at 4/255, BIM and PGD leave at most 2.5% robust accuracy, and AutoAttack reaches 0% on all four architectures from 2/255 upward. Differences between architectures appear only at the smallest budget, where AutoAttack at 1/255 leaves ViT-B/16 with 9.8% robust accuracy while the other three stay below 1.5%. The single-step FGSM is the exception: on ResNet-50 and ConvNeXt its robust accuracy stays around 80% regardless of budget. We show that this is in part an evaluation artifact of restricting the 1000-way classifier to the 10 Imagenette classes. On the full 1000-class output, the same FGSM images at 32/255 score 24.0% on ResNet-50 instead of 80.5%, and 43.5% on ConvNeXt instead of 85.5%. This still leaves FGSM far weaker than iterative attacks with the same budget. Our from-scratch implementations of FGSM, BIM and PGD are verified against torchattacks, and all code, results and figures are released.

        Speaker: David-Ștefan Deonise (National University of Science and Technology POLITEHNICA Bucharest, Romania)
    • Open Source Education and Research & SNS Room 3

      Room 3

      Conveners: Adrian ISTRATE (Agency ARNIEC/RoEduNet, “Dunărea de Jos” University of Galati), Paul GASNER (Agency ARNIEC/RoEduNet, Alexandru Ioan Cuza University of Iasi)
      • 44
        AIMS: Articulated Intelligent Manipulation System

        Low-cost robotic manipulators are increasingly used in education and research, where affordability, portability and ease of deployment are important. Many existing platforms still depend on an external computer to perform essential functions such as inverse kinematics, trajectory planning and user interaction. This dependence introduces additional hardware and software complexity, requires continuous connectivity to a host system, and limits the portability and self-contained operation of the manipulator. These limitations highlight the need for affordable robotic platforms capable of performing core motion-control and Cartesian-planning tasks directly on embedded hardware. To address this need, we present AIMS, a low-cost six-degree-of-freedom robotic manipulator with fully integrated Cartesian control. The system runs entirely on an ESP32-S3 microcontroller, handling inverse kinematics, motion planning, calibration and teach-and-repeat functionality without requiring an external computer. A web interface enables wireless control from any browser, while a separate computer vision module based on OpenCV and YOLO performs object detection and workspace mapping. The robot uses stepper motors with belt transmissions and planetary gear reductions to achieve high resolution in a low-cost design. A real-angle joint model simplifies the kinematics, and gravity-induced deformation compensation improves accuracy under load. Experimental results indicate sub-millimeter repeatability and efficient vision-guided manipulation within a workspace of approximately 750 mm.

        Speaker: Andrei Cojocaru (National University of Science and Technology POLITEHNICA Bucharest)
      • 45
        A Dual-Domain CNN Framework for Steganography Detection and Generation in Web Applications

        Steganography conceals information within digital images through modifications imperceptible to the human eye, while steganalysis addresses the inverse problem of detecting such hidden content. Convolutional neural networks (CNNs) have been shown to outperform classical steganalysis techniques, yet these advances remain largely confined to academic research and are absent from accessible, publicly usable tools. We present a web platform that unifies steganographic generation and CNN-based detection across the spatial and frequency domains. Two CNN architectures were developed by adapting SRNet, ZhuNet, and YeNet to available hardware constraints: a spatial-domain network for PNG images and a frequency-domain network for JPEG images. Both were trained on datasets derived from ALASKA2 using four embedding algorithms, including a cascading-fallback pixel-value-differencing (PVD) method that resolves the boundary-overflow problem of the classical algorithm, and a texture-adaptive DCT method that concentrates payload in naturally noisy image regions to resist statistical detection. The spatial-domain network achieved detection rates of 96.49% on unmodified images, 98.53% on LSB-stego images, and 99.33% on PVD-stego images, while the frequency-domain network reached 85.24%, 96.12%, and 98.27% on unmodified, JSTEG, and adaptive-stego images, respectively. A usability study with five participants confirmed the platform is intuitive and inspires user confidence despite occasional false positives. These results demonstrate the feasibility of integrating deep learning-based steganalysis into a practical, user-facing detection platform.

        Speaker: Giorgiana Vlăsceanu (University Politehnica of Bucharest)
      • 46
        bhyve - libvdsk checkpoint functionality

        The bhyve hypervisor gained support for additional virtual disk formats through the vdsk library, enabling the use of image formats beyond the traditional raw format. Despite these improvements, bhyve still lacks a native checkpoint mechanism for virtual machines using advanced disk formats such as qcow2. This paper presents the design and implementation of a checkpoint solution for qcow2-based virtual machines built on top of the existing bhyve suspend and restore infrastructure. The proposed approach extends the vdsk abstraction layer and integrates qcow2 external snapshot capabilities to provide checkpoint creation, restoration, and management. The implementation reuses existing bhyve mechanisms whenever possible, reducing complexity while maintaining compatibility with the current architecture. The paper discusses the required modifications to both vdsk and bhyve, analyzes the implications of using external checkpoints based on copy-on-write techniques, and compares the proposed approach with checkpoint mechanisms available in other hypervisors such as QEMU/KVM.

        Speaker: Ionuț Mihalache (National University of Science and Technology POLITEHNICA Bucharest)
      • 47
        Analysis pipeline for mouse visually evoked defensive behaviours using DeepLabCut

        Automated analysis of defensive behavior requires a practical connection between pose estimates, experimental timing, and explicit behavioral definitions. This paper presents a Python pipeline built on open-source components for analyzing single-mouse looming-stimulus recordings. The workflow detects stimulus-related brightness changes in a mirror, extracts analysis clips, applies SuperAnimal-TopViewMouse inference with an adapted animal detector, filters coordinates, corrects perspective, and assigns freezing, fleeing, and shelter-occupancy labels. Different keypoints support different behaviors, and intermediate videos expose the effects of each processing stage. A descriptive audit of eight recordings contains 4,774 analyzed frames at 30 frames/s, with 24 predicted freezing intervals, 15 fleeing intervals, and one shelter-occupancy interval. One recording has a truncated output, and all four retained keypoints are unavailable in 20.55% of analyzed frames after post-processing. These results characterize the current implementation; they do not establish behavioral accuracy. The contribution is an inspectable, experiment-specific integration of established tools, with explicit operational definitions and a quantitative characterization of prediction coverage, interval structure, and processing limitations.

        Speaker: Remus Rughinis (Transilvania University of Brasov, Faculty of Medicine, Research and Development Institute, Brasov, Romania.National University of Science and Technology POLITEHNICA Bucharest)
      • 48
        Building a Statute Retrieval Evaluation Set for the Romanian Civil Code: LLM-Assisted Annotation and Expert Validation

        Retrieving the statutory provisions needed to assess a legal statement is a core task in legal information retrieval. For Romanian, existing resources provide legal examination questions and legislative corpora, but article-level relevance annotations remain a gap. We describe an LLM-assisted approach to constructing a statute retrieval evaluation set for the Romanian Civil Code from professional admission examinations.

        The source material consists of publicly available civil law questions from admission examinations for magistrates (INM) and lawyers (INPPA). Each question–answer-option pair forms a separate query, preserving the context needed to interpret the option. Official answer keys establish the expected assessment, taking account of the wording of each question. The annotation task is to identify the provision or provisions needed to support that assessment, including those needed to establish why a statement is false.

        Language models propose candidate article identifiers from the complete index of Civil Code article headings, avoiding dependence on a candidate pool supplied by a preselected retrieval system. Each proposed article must be accompanied by a verbatim passage, checked automatically against the corresponding corpus text. This check establishes textual traceability; legal relevance and completeness require separate assessment. Models are prompted independently, and their agreement is used to prioritize annotations for human review.

        Legal experts will independently evaluate a subset of the annotations for relevance, sufficiency, and omitted provisions. The evaluation will include both model-agreement and model-disagreement cases, allowing us to examine whether agreement is a useful indicator of annotation quality. We will measure inter-model agreement, quotation-verification failure rates, agreement with expert judgments, and inter-expert agreement, with results analyzed by examination source and statement truth value. An error taxonomy will characterize recurring annotation failures. The planned release comprises the annotation pipeline and an expert-reviewed evaluation subset, with disagreements adjudicated before release.

        Speaker: Dorin-Andrei Geman
      • 49
        Multi-modal Social Media Simulations in OASIS

        Large-scale LLM-based social media simulatorssuch as OASIS can reproduce phenomena such as informationspreading, group polarisation and herd effects, but their agentsread and write text only. Real platforms increasingly carry imagesand video, and misinformation in particular often pairs a claimwith a supporting image. We extend OASIS so that media isa first-class part of a post: images and videos are stored withthe post and passed directly to a vision-language backbone(Qwen3.8-27B) whenever an agent views its feed, and contentexposure is handled by the Gorse recommender engine. Wedemonstrate the extension by adapting the OASIS information-spreading experiment to image posts from the Fakeddit dataset,seeding one post from each of Fakeddit’s six categories intoa 111-agent network built from Twitter15 data and trackingpropagation scale, depth and breadth over 30 steps. Five of thesix posts produced substantial cascades. Satire/Parody spreadfastest, Manipulated and Imposter content formed the deepestrepost chains, and the False connection post barely spread. Withone post per category and one run per configuration, theseresults illustrate what the framework can measure rather thanestablishing differences between categories.

        Speaker: Teodor Milea (UNSTPB)
      • 50
        When Do Distilled Surrogates Preserve Algorithmic Amplification? A Simulation Study with LLM Agents

        Auditing algorithmic amplification on very largeonline platforms, as envisioned under Article 40 of the EU DigitalServices Act, is obstructed by two constraints: auditors rarelyhave white-box access to production rankers, and counterfactualinterventions cannot be run on a live platform. A naturalworkaround is to distill a surrogate of the ranker and studyit inside an agent-based simulation, but this is only soundif conclusions transfer from surrogate to target. We test thistransfer in OASIS, an LLM-agent simulation of X/Twitter whoseranker is TwHIN-BERT. We distill two surrogates under differentaccess assumptions, one reusing the target’s first four transformerlayers and one training a compact sentence encoder on thetarget’s embeddings, and run matched simulations in whichonly the ranker differs. When a single post spreads throughthe population of its real cascade, all rankers reproduce thescale of Twitter15/16 cascades equally well. When 100 postscompete for two recommendation slots, the surrogates share only9–29% of their first recommendations with their teacher, andoutcomes such as the share of reposts on false rumors varyacross rankers. A minor implementation change in the targetitself, its position-embedding handling, reduces agreement withthe deployed ranker to 7%. Embedding-level fidelity thereforedoes not guarantee fidelity of exposure outcomes, and single-postcascades are a weak test of a ranker. These preliminary, single-seed results motivate validating surrogate-based audits in settingswhere the ranker actually filters content.

        Speaker: Teodor Milea (UNSTPB)
    • 11:00
      Coffee Break
    • Cloud Computing and Network Virtualisation Room 1

      Room 1

      Convener: Laurențiu STOLERIU (“Alexandru Ioan Cuza” University of Iasi)
      • 51
        Versioning and Data Quality in MLOps Pipelines

        In today’s machine learning industry, a significant
        portion of technical debt accumulates due to the maintenance
        of complex data infrastructures. To address this, we propose a
        configuration-driven MLOps pipeline that unifies data validation,
        data versioning, experiment tracking, and drift reporting across
        structured and unstructured datasets. Rather than optimizing a
        single predictive model, this paper evaluates whether the same
        pipeline logic can be reused across data modalities with minimal
        source-code changes. The medical image and transcripts and the
        tabular experiments serve as two stress tests for the pipeline.
        They show that the same orchestration layer can train, track,
        and validate heterogeneous model families, while also surfacing
        limitations such as class imbalance. Ultimately, the framework
        demonstrates the feasibility of a reproducible MLOps design
        pattern managed entirely through a YAML-controlled execution
        layer.

        Speaker: Mr Cosmin Marian Zamfir (National University of Sciences and Technology Politehnica Bucharest)
      • 52
        Minimizing Active Hosts: Predictive Capacity Planning for Dynamic NFV Workloads

        Flexible network service deployment is made possible by Network Function Virtualization (NFV); yet, under severe resource restrictions and dynamic traffic, it is still difficult to determine how many Virtual Network Function (VNF) instances to run and where to locate them on shared infrastructure. We concentrate on \emph{host-level capacity planning} for NFV clusters in this study. We provide a hybrid predictive-optimization methodology that ensures that aggregate traffic demand can be met under CPU, memory, and network-interface capacity limits while minimizing the number of active hosts.

        First, we develop a tractable Integer Linear Programming (ILP) model that determines the number of instances to deploy for each kind of VNF and host; the model explicitly accounts for multi-resource host capabilities but does not account for service-function chains or per-flow routing. The Chronos-2 time-series model, which forecasts short-term aggregate traffic levels and permits proactive scaling decisions, is the basis for the forecasting layer that powers the ILP. A standard, robust Layer~3 control plane that uses OSPF with Anycast addressing and quick failure detection then enforces the consequent host activations and deactivations.

        In a time-stepped simulation, we test the framework on actual traffic traces from the CESNET dataset. Our results show that, for the examined NFV cluster sizes, prediction-driven ILP-based capacity planning may reduce the number of active hosts while meeting the anticipated aggregate demand in comparison to reactive baselines. These findings demonstrate that NFV deployment efficiency may be significantly increased without the requirement for fully complete learning-based controllers by combining short-term forecasts with even a host-level capacity planning layer.

        Speaker: Claudiu Traistaru (Universitatea din Craiova)
      • 53
        Edge-Fog-Cloud MLOps Architecture for Urban Traffic Video Monitoring with YOLOv5

        This article attempts to solve the constraints of tradi-
        tional urban traffic monitoring using a comprehensive three-level
        Edge-Fog-Cloud architecture. The huge amount of video data in
        smart cities creates bandwidth problems and the instability of IoT
        networks and data drift often cause a decrease in the effectiveness
        of machine learning models in production.
        The main contribution of this study is the development of this
        decoupled MLOps system that provides computational workloads
        to the edge, fog and cloud tiers. Targeted validations were carried
        out on critical data integrity, resilience and deployment proce-
        dures to thoroughly validate the viability of this architecture.
        To clearly isolate validated outcomes from architectural inten-
        tions, the YOLOv5 inference, Contract Testing, and retraining
        cycles were functionally validated on the UA-DETRAC dataset,
        deploying an improved training configuration that raised the
        mAP@0.5 score from 0.297 to 0.972. The cloud orchestrator was
        prototyped on the other side on a local dedicated server instead of
        a public commercial cloud. The Fog Store-and-Forward resilience
        and OTA Atomic Swap mechanisms are still at a simulated
        prototype level and need a future validation on physical IoT
        gear.

        Speaker: Mr Radu Cristian Popescu (National University of Sciences and Technology Politehnica Bucharest)
      • 54
        Sizing Shared Storage for High-Availability Virtualization Clusters: A Study of Queue Depth, Multipathing and Storage Tiering

        Shared storage is a critical part of high availability clusters, and any stall or performance issue on that path affects the entire cluster. Hardware specifications often present system-wide peaks, and give no information about what a host or a virtual machine will see. In this paper we analyze a cluster composed of several servers connected to a shared storage over Fibre Channel links, with multipath I/O enabled, and both NVMe and SAS disks. We measure throughput, latency distribution and CPU load from both the cluster hosts and the virtual machines running on the cluster, sweeping through queue depth, block size, read/write mix and multipath policies. NVMe volumes saturated at under 300k read an 200k write IOPS, after which latency grows linearly, as predicted by Little's law. At around 1000 outstanding I/Os, the host queue is filled with 250-280ms stalls. Round-robin multipath increases the average throughput by 15%, but widens the latency tail, compared to fail-over-only multipath. The throughput reaches up to 10GB/s for cached entries, and after the cache is exhausted it drops down to about 700MB/s for NVMe volumes and less than 50MB/s for SAS volumes. The overhead introduced by the virtualization layer and guest file system is measured from the difference between host and VM measurements.

        Speaker: Stefan Dorin Jumarea (National University of Science and Technology POLITEHNICA Bucharest)
      • 55
        Computing Infrastructure for support AI Education and Research

        In the paper described the development of the distributed computing e-infrastructure for research and education in the Republic of Moldova and its specialization for the support of artificial intelligence (AI) applications. The infrastructure unites the computing resources of several data centres of universities and research institutions over the RENAM network and includes a GPU segment operated at the Vladimir Andrunachievici Institute of Mathematics and Computer Science of Moldova State University (IMCS MSU) that is available for researchers and students from different institutions. The GPU-enabled infrastructure and the JupyterHub-based environment is used at IMCS MSU for development of AI applications: “HeDy platform for the digitization of historical documents, neural machine translation, text classification with large language models” and “AI-based Educational Platform Guguta for speech recognition and synthesis for Romanian, retrieval-augmented generation”. The use of the deployed resources for supporting the master programme in Artificial Intelligence and Data Science in MSU is also discussed.

        Speaker: Dr Tudor Bumbu (Vladimir Andrunachievici Institute of Mathematics and Computer Science of Moldova State University)
      • 56
        Using Emulation Software as Core Backbone in Real SDWSN Deployments

        Commercial OpenFlow solutions face the problem that deployments become too costly for edge Software-Defined Wireless Sensor Network (SDWSN) sites, whether industrial IoT, automotive sensors, campus networks, or other field deployments that coordinate sensors on a Local Area Network (LAN). We address this with a cost-effective OpenFlow backbone based on emulation software on a small single-board computer. Dual-homed virtual hosts belong to both the programmable SDWSN fabric and the physical device LAN, keeping the two networks separate instead of one flat address plane. Our prototype runs Mininet-WiFi on a Raspberry Pi with ESP32 nodes. From the dual-homed hosts, authorized LAN access to sensors is possible (read data, reconfigure, manage, and related operations) under OpenFlow control, while sensors are Layer 2 separated from the OpenFlow-based switch and controller. The result is a fully reusable cross-domain programmable SDWSN edge solution at a fraction of commercial costs, with increased security through dual-homing.

        Speakers: Adrian PECULEA (Technical University of Cluj-Napoca), Dr Bogdan IANCU (Technical University of Cluj-Napoca, Computer Science Department)
    • Doctoral Symposium Room 2

      Room 2

      Conveners: Florin Pop (University Politehnica of Bucharest, Romania / National Institute for Research & Development in Informatics – ICI Bucharest, Romania), Vasile MANTA (Gheorghe Asachi Technical University of Iasi)
      • 57
        Guarding Tool-Using LLM Agents Against Indirect Prompt Injection: A Benchmark and Three Runtime Filters

        LLM agents that browse, read email, edit files and run commands act on whatever text they read. This exposes them to indirect prompt injections: instructions hidden in a document, a web page or an email that hijacks the agent into doing something the user never asked for. We build a sandboxed agent with eleven simulated tools (file system, inbox and outbox, calculator, an allow-listed terminal and a local browser), served by open models through Ollama, and a suite of 51 injection scenarios that cover three attacker goals (data exfiltration, unauthorized action, destruction), four techniques (direct, role-play, encoded, hidden markup) and three delivery channels (email, file, web page). Three runtime filters sit between the agent and its tools: a keyword blacklist over tool arguments, an allow-list derived automatically from the entities named in the user's task, and an LLM judge that sees both the task and the proposed call. In the evaluation we separate compromise, where the agent attempts the injected action, from execution, where the action goes through, which credits a defense for stopping an already hijacked agent. On qwen3:14b, the undefended agent attempts the injected action in 10% of scenarios (28 of 204 attempts across all runs) and completes 78% of benign tasks; role-play injections are by far the most effective technique (32% of attempts), while encoded ones never succeed. The keyword filter stops only 2 of 9 attempted actions, the allow-list 6 of 7 and the LLM judge 5 of 7; they block at least one legitimate call on 12%, 12% and 10% of benign tasks, and the judge adds 1.2 s per task (1.1x latency). The attacks that get through are those that abuse the task itself, such as deleting a file the user named or leaking a secret inside an otherwise legitimate email. We release the benchmark, the harness and all traces.

        Speaker: Mr Costin-Alexandru Deonise (National University of Science and Technology POLITEHNICA Bucharest, Romania)
      • 58
        Stochastic Taylor Derivative Estimation of High-Dimensional Laplacians: memory footprint, variance structure and fault tolerance

        Laplacian operators drive the physics-informed training of neural PDE solvers, where exact second derivatives cost time and memory growing quadratically with the input dimension d. The Stochastic Taylor Derivative Estimator (STDE) replaces the Hessian trace by an average of N second directional derivatives, each from one Taylor-mode forward pass. We characterise STDE along three axes. Scalability: the dense Hessian of a multilayer perceptron exceeds 8 GB at d = 5.1e4, whereas STDE reaches d = 1.6e6 within 3.2 GB, with cost linear in d, orders of magnitude below exact evaluation, and sparse directions three to eight times cheaper than Hutchinson probes at equal accuracy. Variance: the estimator is unbiased for sparse and dense directions; its variance, computed from the Hessian, predicts the measured dispersion within 4% with no fitted parameter, and analytic test functions at d = 1e6 confirm the 1/sqrt(N) law with relative errors from 1e-7 to 1e-1 depending on how the directions match the Hessian structure. Fault tolerance: when worker processes are killed at Weibull-distributed times, the pooled estimate stays unbiased and its scatter follows the surviving-sample fraction through sqrt(1/|A| - 1/N), with an observed-to-predicted ratio of 1.06. STDE is a practical Laplacian primitive for failure-prone compute environments.

        Speaker: Mr Costin-Alexandru Deonise (National University of Science and Technology POLITEHNICA Bucharest, Romania)
      • 59
        Fake News Detection in Romanian Using Language Models

        In this paper, we present a system for fake news detection in Romanian, a low-resource language with few manually annotated news articles and where real news prevails among available data. The task is classification of news articles into four classes - real news and three types of misinformation - through three iterations. The first iteration fine-tunes Romanian BERT encoder models with convolutional classification heads on about 760 COVID-19 articles from the FakeRom corpus, manually labeled by domain experts, resulting in 84.2% accuracy and weighted F1 of 0.83. In the second iteration, we go to a much larger corpus and a new four-class taxonomy, use generative large language models (LLMs), RoMistral-7b and RoLlama2-7b, and fine-tune these models with Low-Rank Adaptation (LoRA). We evaluate how 4-bit quantization and prompt phrasing impact our results. RoMistral-7b gets to the weighted F1 of 0.85 and quantization incurs penalty of six to seven points in both models. The third iteration returns us to an encoder with a fully fine-tuned XLM-RoBERTa-large classifier, which we evaluate using ten configurations with only one variable changed each time - the backbone model, loss functions for handling the class imbalance, two types of data augmentations, external data from SaRoCo satire corpus, and whether class descriptions are included in the input. The best configuration reaches a weighted F1 of 0.88 while training in under 15 minutes, compared to two to three hours for the generative models on the same hardware. Another comparison is made between the above configuration and the best commercially available
        LLMs such as GPT, Gemini, and Claude, which are prompted instead of fine-tuned on the four-class task. These models have not been exposed to the corpus or adapted for Romanian misinformation categories, and this comparison seeks to determine if the general-purpose scale holds its own.

        Speaker: Mr Eduard Petcu (National University of Science and Technology POLITEHNICA Bucharest)
      • 60
        An IoT-Based Context-Aware Architecture for Real-Time Assistance in Business Metaphor Learning

        Business metaphors are widely used to describe economic and organizational phenomena, but their meaning often depends on the context in which they occur, making them difficult to understand and learn. This paper proposes an IoT-based, context-aware architecture for real-time assistance in business metaphor learning. The IoT component acquires and transmits information about the current learning context and relevant interaction events, providing contextual input that is combined with linguistic context, a structured metaphor knowledge base, and learner-state information. A large language model (LLM) uses this information to generate assistance adapted to the current learning situation. Learner responses are assessed and used to update the learner state, while subsequent interactions initiate a new acquisition of contextual information, creating a continuous assistance cycle. This cycle is formally represented using a Generalized Stochastic Petri Net (GSPN) and analyzed in GreatSPN. Reachability and boundedness analysis, together with CTL model checking, are used to examine the progression and continuation of the modeled cycle. Finally, an illustrative business-metaphor learning scenario shows how the proposed components interact during the learning process.

        Speaker: Liana Boca (Transilvania University of Brașov & Technical University of Cluj-Napoca)
      • 61
        Multi-scale Digital Twin Using Emerging Technologies in Computing Continuum

        The concept of the Digital Twin originally emerged in the aerospace industry, but has in recent years been extended to domains of personal interest, such as medicine and industry. However, the specialized literature remains rather limited, presenting solutions at a single scale level and from a single data source. In parallel, recent research shows that Physics-Informed Neural Networks and hybrid surrogate models are being created by combining physical laws with the speed oered by neural networks; however, these approaches are typically validated on relatively simple problems, involving a single dierential equation and are not integrated into distributed architectures capable of simultaneously handling heterogeneous data sources. Classical numerical simulations require HPC resources and execution times on the order of hours or days, which are incompatible with the real-time decisions required by critical medical or industrial applications. The general objective is to research, design and evaluate the performance of a distributed and scalable platform for building multi-scale Digital Twins, which combines articial intelligence, parallel and distributed systems, numerical methods, heterogeneous hardware infrastructures and Edge Computing, capable of near-real-time simulations and predictions.

        Speaker: Taisia-Maria COCONU (National University of Science and Technology POLITEHNICA Bucharest)
      • 62
        Attention-Based Multi-View Fusion for Interpretable Abnormality Detection in Radiographs

        Musculoskeletal abnormality classification in radiographs is a challenging task that requires integrating information across multiple views while ensuring that model predictions are based on clinically meaningful evidence. This work proposes an interpretable multi-view architecture that combines a shared convolutional feature extractor with attention-based aggregation for study-level classification of musculoskeletal abnormalities. The attention mechanism learns view-specific importance weights, enabling both improved prediction and direct interpretability. Experiments on the MURA benchmark show that attention-based fusion substantially outperforms score-level aggregation, improving AUC by approximately 0.23. To analyze model behaviour, we leverage Grad-CAM visualizations in conjunction with attention weights, revealing reliance on non-anatomical artifact regions in trained models. To address this problem, we propose an OCR-based masking pipeline that suppresses text-related confounders, reducing artifact dependence with only a small performance drop. Our results show that multi-view attention, interpretability analysis, and artifact masking provide an effective framework for transparent radiographic classification.

        Speaker: Mr Andy Catruna (National University of Science and Technology POLITEHNICA Bucharest)
    • Technologies for Future Internet & HPC Room 3

      Room 3

      Convener: Prof. Răzvan Victor RUGHINIȘ (National University of Science and Technology POLITEHNICA Bucharest)
      • 63
        A personalised Framework for Residential Energy Consumption Estimation Using Electricity Bill Calibration

        Monitoring residential energy consumption at the appliance level remains a challenge, as consumption data is typically available only in aggregated form. This paper proposes a methodological framework for personalized energy consumption estimation that combines visual recognition based on the Google Gemini Vision multimodal model with a calibration mechanism grounded in actual consumption history extracted from energy bills. Individual estimates are adjusted using the calibration coefficient α=Creal/∑Ex, thereby reducing discrepancies between estimated and actual consumption. The framework is validated through the EnergyScan AI system, which integrates automatic appliance identification with energy bill processing. A technical evaluation on a set of 15 appliances achieved high accuracy in category classification, with rated power estimates falling within ±20% for 40% of cases. A usability study conducted with 20 participants yielded an average score of 4.9/5 and a correct appliance identification rate of 82%, demonstrating the feasibility of the proposed approach for personalized energy consumption monitoring.

        Speaker: Ms Cristina-Gabriela Racoviţă (Faculty of Computer Science, University “Alexandru Ioan Cuza” of Iași)
      • 64
        A Client-Centric Architecture for Metadata-Minimizing End-to-End Encrypted File Storage

        Existing End-to-End Encrypted (E2EE) File Storage solutions do not encrypt all file system metadata, failing to ensure full confidentiality, or do so by sacrificing concurrency or by opting for object storage instead of traditional file system semantics.

        We present Blobkeep, a client-centric architecture for metadata-minimizing End-to-End Encrypted file storage. The server stores abstract blobs which contain encrypted pages. Clients decrypt the blobs and operate on the file system in a concurrent and distributed manner, synchronized by the server without exposing decrypted content to it. Authenticated encryption is used to ensure confidentiality and tamper resistance, even in a worst case scenario where an attacker gains complete control of the server. Generation numbers are used to guard against data races without the overhead imposed by locking. File system in Userspace is used to serve a UNIX file system to the end user.

        We provide a description and in-depth security analysis of the blob architecture, as well as a functional proof of concept implementation.

        Speaker: Călin Ioncioaia (Faculty of Computer Science)
      • 65
        Analysis and Mitigation of a Localized Optical Loss in an OPGW Link Supporting SCADA Communications: A Field Case Study

        Abstract—This paper presents a field-validated approach for identifying and mitigating a dominant localized optical loss in an optical ground wire (OPGW) link supporting supervisory control and data acquisition (SCADA) communications. The approach integrates optical time-domain reflectometry (OTDR) event-level analysis, correlation with physical route documentation, targeted maintenance, and post-intervention verification. The investigation was conducted on a 27.98-km operational OPGW link using OTDR measurements at 1310 nm. A dominant localized loss of 0.98 dB was identified at 11.95 km and correlated with an intermediate optical distribution frame (ODF) connection. The connection was replaced with a direct fusion splice, after which the recorded event loss decreased to 0.09 dB, corresponding to an improvement of 0.89 dB and a 90.8% reduction in localized loss. The contribution of the targeted event to the cumulative recorded localized event loss decreased from approximately 57.0% before intervention to 11.25% after intervention, while the remaining recorded events exhibited changes no greater than 0.02 dB. The results demonstrate that event-level OTDR analysis combined with physical-route correlation can identify dominant localized losses, guide targeted corrective actions, and quantitatively verify their optical effect in operational OPGW infrastructure supporting SCADA communications.

        Speaker: Mr Vladimir Mirovski (Universitatea Tehnică a Moldovei)
      • 66
        Digital Twin and MLOps for IoT

        Residential heating is one of the most significant
        energy-consuming tasks in a household; however, in most homes
        it is still managed by static, rule-based thermostats that ignore the
        data generated by smart home sensors. This paper investigates
        whether a reinforcement learning based control policy used to
        control a thermostat, evaluated end-to-end via a reproducible
        MLOps pipeline and utilizing the Digital Twin concept to simulate
        a 1R1C smart home calibrated on the dataset, can produce
        measurable benefits compared to a simple rule-based solution.
        The main contribution lies in the development of such a system
        as well as the creation of a rigorous and transparent testing
        protocol, in which every controller is scored by one shared reward
        on a leakage-safe held-out split across three windows and five
        seeds, with confidence intervals. The main finding is that a PPO
        policy ties a well-tuned scheduled thermostat in-distribution and
        reduces demand-response exceedance by 41 percent, but does not
        beat it out of distribution; a serve-time filter shows that parity is
        reached by deferring 97 percent of decisions to the rule. Dropping
        a local large language model and an idealized model-predictive
        oracle into the same residual slot shows that a real headroom of
        roughly 8 percent over the rule exists, but is not realizable by a
        controller that must learn from a single short trace.

        Speaker: Mr Denis-Jenica Baston (National University of Sciences and Technology Politehnica Bucharest)
      • 67
        On the Reliability of SQL Performance Tests for OLAP Loads in Private Clouds Managed with OpenStack

        A broad spectrum of architectures, platforms, and tools is available for data persistence and processing. Because SQL is the most popular language for processing small, medium, and large volumes of data, most setups for comparing the OLAP performance of different tools rely on executing SQL query sets on established database benchmarks such as TPC-H or TPC-DS. Building the testing setups on premises (local hardware and software resources) is costly, so hiring cloud providers’ resources is a natural choice.

        Speaker: Prof. Marin Fotache (Al.I. Cuza University of Iasi)
      • 68
        An Evaluation Protocol for Learned Network Digital Twins in Congestion Control

        Learned network digital twins promise what-if analysis and safe policy training without touching production networks, yet they are typically validated by how accurately they reproduce observed behaviour. Accurate reproduction does not establish that a twin answers counterfactual questions correctly, that it encodes the physical properties of the path, or that controllers tuned inside it still perform well on the network it models. We propose an evaluation protocol that measures these properties separately. In a single-bottleneck congestion control setting, the sender observes only acknowledgement-derived telemetry, while link capacity, buffer depth, base round-trip time and cross-traffic are hidden but fully controlled, and any what-if scenario can be replayed exactly from the same seed. We assess twins built from recent world-model architectures on four axes: prediction accuracy over horizons measured in round-trip times; recovery of hidden path parameters, compared against minimum round-trip time and packet-pair estimators; agreement with the simulator on the effect of alternative sending rates; and the performance of rate controllers trained entirely inside the twin. We hypothesize that prediction accuracy is a weak indicator of the other three, and that controllers trained inside a twin exploit regions where it is optimistic, such as underestimated queueing delay. Initial experiments use a fast queueing simulator, and the simulator and evaluation suite are released for reuse.

        Speaker: George-Andrei Nita-Pirvu (Universitatea Politehnica Bucuresti)
    • 13:00
      Lunch
    • Cloud Computing and Network Virtualisation Room 1

      Room 1

      Convener: Vlad Rădulescu (Faculty of Computer Science)
      • 69
        Evaluating the Privacy–Utility Trade-off in Tabular Data Anonymization: A Comparative Study of k-Anonymity, l-Diversity, and t-Closeness

        Publishing tabular data containing quasi-identifiers involves a trade-off between protection against disclosure and preservation of data utility. This study compares k-anonymity, distinct l-diversity, and t-closeness using 32,561 records from the UCI Adult Census Income dataset. Twelve configurations were evaluated, corresponding to k values of 3, 5, 10, and 20 for each model, with l=2 for l-diversity and t=0,2 for t-closeness. The evaluation combines the suppression rate, Discernibility Metric (DM), accuracy loss across ten socioeconomic queries (QAL), and exploratory predictive accuracy. k-anonymity resulted in suppression rates of 0,01–0,12% and QAL values of 0,01–0,15%, l-diversity yielded values of 1,57–2,22% and 1,09–1,58%, respectively, while t-closeness yielded values of 12,66–12,75% and 5,62–5,71%. Exploratory predictive accuracy differed from the baseline by no more than approximately 0.22 percentage points for k-anonymity and l-diversity, whereas t-closeness recorded reductions of approximately 0.98–1.96 percentage points. The joint analysis of suppression, generalization, and DM shows that retaining more records does not automatically imply superior structural preservation. Under the conditions analyzed and at the operational threshold of 5%, l-diversity exhibited the most balanced profile among the models that explicitly protect the sensitive attribute.

        Speaker: Ms Nelea Manin (Technical University of Moldova)
      • 70
        Bridging FPGA-Accelerated Inference and Cloud-Native Environments: A CPU-Driven Approach

        The use of ML has grown considerably over the last period, leading to multiple levels of services offered on a cloud platform. An example is Inference as a Service, which offers a pre-developed tool that can directly do predictions based on a pre-trained model. Based on our previous work related to the Cloud-Native (CN)-like attachment of FPGA, we analyze the performance obtained by relying on a simpler methodology to bridge FPGA-accelerated Gradient Boosted Decision Trees (GBDT) inference and CN-like environments to enable the creation of ML inference services in small and cost-constrained bare-metal cloud deployments. This paper aims to expose the advantages, disadvantages and the performance obtained by using such a simplified bridging methodology, based on an FPGA-based inference and translation between the CPU subsystem and the FPGA region without relying on DMA-based transfers.

        Speaker: Alin-Tudor Sferle (Technical University of Cluj-Napoca)
      • 71
        Whose Cloud Is Your Country On? Single-Provider Dependencies of Romanian Critical Services

        Digital-sovereignty policy usually emphasises data location, whereas public-service availability also depends on the organisations operating authoritative DNS, edge or content-delivery networks, and hosting; configured mail and certificate issuance expose related dependencies. We map these five layers for a purposive sample of 192 Romanian public-facing services in ten sector groups derived from the NIS2 annexes. A Bucharest measurement, compared with a non-Romanian cloud vantage, used low-rate DNS queries, one certificate read, and one landing-page request chain per service. The original contribution is a service × layer × operating-organisation audit unit that separates web reachability, configured mail, and certificate issuance; applies the same sole-observed-provider rule to state and commercial operators; and reports modelling and scope sensitivities. Cloudflare was the sole observed provider at one or more web-reachability layers for 53 services (27.6%). Romania’s Special Telecommunications Service ranked second for web reachability and was concentrated in public administration. Sixty of the 62 services with an identifiable edge used a US-operated provider, and Cloudflare and Microsoft exposure sets each spanned all ten sector groups. The dnsc.ro worked example connects the snapshot to the provider’s 18 November 2025 outage without treating a measurement taken almost ten months later as an outage reconstruction. The results show that jurisdictional sovereignty and independently operated redundancy are distinct, measurable properties.

        Speaker: Mr Vlad-Stefan Dieaconu (National University of Science and Technology POLITEHNICA Bucharest)
      • 72
        eBPF-Centric Security Observability and Investigation in Kubernetes

        Kubernetes environments introduce dynamic networking, short-lived workloads, and multiple enforcement layers that make application behavior difficult to inspect and validate through static configuration alone. Existing cloud-native security tools provide network visibility, runtime monitoring, and policy enforcement, but they are commonly operated independently, resulting in fragmented operational and security evidence. The proposed solution, Isolens, integrates these capabilities into a configurable Kubernetes environment centered on eBPF-based observability and investigation. Terraform provides reproducible infrastructure and resource provisioning, while Cilium and Hubble expose network flows and enforcement verdicts, and Tetragon captures process-level runtime activity associated with Kubernetes workloads. Kyverno complements runtime visibility with admission-time policy enforcement. Applications are deployed into isolated namespaces and can be configured according to the required topology, exposure model, and communication constraints. The platform also enables operators to define custom validation scenarios by configuring how applications are deployed, exposed, interconnected, and constrained by security policies, providing a controlled and repeatable workflow for exercising and investigating application behavior. Amazon EKS is used as the current reference implementation, while the architecture separates provider-specific infrastructure from the Kubernetes-oriented observability and policy layers. Validation shows that network behavior, runtime activity, and policy decisions can be correlated across multiple evidence sources, providing a reproducible basis for application debugging, security validation, and Kubernetes workload investigation.

        Speaker: Mihai Alexandru Andrei
      • 73
        Air-Quality Observations and XGBoost Predictor Sensitivity During the August 2026 Serbian Wildfire-Smoke Episode in Timișoara, Romania

        The present article presents the effects of the smoke produced by a wildfire situated in Serbia on the sensor networks and an XGBoost based air quality predictor deployed in the city of Timisoara, Romania. The study covers a period ranging from the 20th to the 22nd of August 2026. The analysis is founded on regulatory measurements, Copernicus Atmosphere Monitoring Service (CAMS) estimates, historical low-cost-sensor data, wind direction analysis, and predicting services' outputs. During the day of 21 August, TM4 pollution monitoring station measured an 88.17 μg/m3 daily PM10 mean and a 378 μg/m3 hourly maximum, while the CAMS daily mean was 23.10 μg/m3. Compared to the measurements in the days before the wildfire (13–19 August), CAMS means increased by 103% for PM2.5, 95% for PM10 and 18.6% for CO during 20–22 August. Moreover, all 24 hours on 20 August and 16 of 24 on 21 August presented southern winds, PM10 values significantly decreasing after the 22 August wind direction reversal in the city of Timişoara. In controlled tests, a CAMS-like data injection increased the PM10 predictions at horizon H+1 from 3.30 to 19.26 μg/m3, while a TM4-like data injection increased it to 68.66 μg/m3. However, PM10 inputs ranging from 100 to 600 μg/m3 produced predictions of 69.61 μg/m3 PM10, demonstrating ulterior decision tree saturation and the system's reducing effects of upstream processing on readings above 200 μg/m3. The study provides a cross source event validation framework for sensor networks and machine learning predictors and emphasizes the importance of distinguishing sensor failure from real exceptional pollution episodes.

        Speakers: Mr Sebastian-Ioan Petruc (Politehnica University Timisoara), Razvan Bogdan (Universitatea Politehnica din Timisoara), Prof. Marius Marcu (Politehnica University Timisoara)
      • 74
        Cloud-Based Solution for Advanced Network Traffic Analysis and Visualization

        Cyber threats are harder to catch than they used to be, and organizations need monitoring that handles high traffic volumes without losing the rare malicious flow in the noise. Standard tools tend to fail especially when we talsk about scaling and real-time response. In papaer we proposes a cloud-native architecture for network traffic analysis built entirely in SAS Viya. Loading the full dataset on conventional hardware kept crashing the JVM on memory, so a stratified downsampling step was written that drops most benign background flows but keeps every malicious signature, leaving a working set of about 1.5 million flows. On it, the system goes past binary detection with two models. A logistic regression baseline gives fast, interpretable scores. A multi-class random forest then attributes traffic to specific families such as DoS Hulk and PortScan, so an incident responder gets something more useful than a blanket “malicious” label, and a gradient-boosting model is used alongside it as a cross-check.

        Speaker: Bogdan-Costel Mocanu (University Politehnica of Bucharest)
    • Pervasive Systems and Computing & Network Management Room 2

      Room 2

      Convener: Paul GASNER (Agency ARNIEC/RoEduNet, Alexandru Ioan Cuza University of Iasi)
      • 75
        Traceable LLM-Based Automated Anamnesis with Clinician-Guided Refinement

        Large Language Models (LLMs) can be used to automate parts of the medical history-taking process, but their use in clinical workflows requires generated information to be easy to verify and the system to follow the structure expected by medical professionals. This paper presents an LLM-based automated anamnesis pipeline that conducts a patient interview and generates clinical summaries for physician review. Each medical statement in the generated summary is linked to the patient message from which the information was extracted, allowing the physician to quickly check the source of the generated information. The traceability mechanism was evaluated on 1,634 clinical statements from 66 generated summaries, and every statement contained a valid link to a patient message. We also evaluate a clinician-guided approach for improving the system. A medical doctor reviewed the initial version and identified problems in the interview protocol and generated reports. This feedback was translated into changes to the LLM prompts and output schema, after which a revised version of the system was evaluated. For the reviewer who assessed both versions, the overall rating of the structured report increased from 1.70 to 4.27 on a five-point scale. The evaluation was performed using synthetic patients across multiple clinical scenarios and assessments from three medical doctors. The results provide preliminary evidence that explicit source linking and clinician-guided prompt and schema refinement can improve the practical use of LLM-generated anamnesis reports.

        Speaker: Mr Stefan Tanasescu (National University of Science and Technology POLITEHNICA Bucharest)
      • 76
        Beyond "Let It Crash": An Empirical Study of the Reliability of Supervised Recovery in Elixir/OTP

        Fault tolerance is a fundamental requirement of highly concurrent and distributed software systems. Elixir/OTP approaches this problem through process isolation and supervised recovery, following the "let it crash" philosophy inherited from Erlang. Rather than attempting to prevent every process failure, failed components are terminated and restarted by supervisors. While this model is widely used, the operating conditions under which repeated crash-and-restart recovery remains reliable are less clearly documented.

        This work proposes an empirical investigation of the reliability limits of supervised recovery in Elixir/OTP. A fault-injection test environment is developed around a stateful transaction-processing workload in which controlled process failures are introduced under varying frequencies and system loads. Recovery is evaluated not only through process restart time, but also through application-level metrics including throughput, tail latency, request backlog, recovery time, transaction completion and correctness.

        The study aims to identify the transition between successful recovery and degraded operation, as well as the conditions under which repeated failures prevent the system from returning to a stable state. Particular attention will be given to interactions between failure frequency, workload intensity, recovery cost and application state. Where limitations of conventional supervised recovery are observed, complementary recovery mechanisms are evaluated to determine whether the stable operating region can be extended.

        By experimentally documenting this recovery boundary, the work seeks to provide a clearer understanding of both the strengths and practical limitations of crash-oriented fault recovery and to create useful design guidance for resilient Elixir/OTP applications.

        Speakers: Mr Eriko-Laurentiu Nitu (University POLITEHNICA of Bucharest), Paul Burcă (University POLITEHNICA Bucharest)
      • 77
        Network Management and Security in Public Schools: a Romanian-German Comparative Analysis

        Creating and maintaining a school’s network infrastructure has been a challenge for many professionals, while the advancement of attacks and the addition of new smart devices amplify the need for a centralized, security first and management software suite for system administrators. Some critical points and services including Remote Desktop Protocol or the connection with the Internet Service Provider are not secure by default and new entry points for attacks appear and are exploited far more often today. This study analyses multiple schools located in Romania and Germany, their own approach of managing network devices, cybersecurity measures and protocols. The assessment also includes the balance between costs and usability, a main problem for public schools. The comparison reveals significant differences between the two nations. The proposed solutions rely on primarily Open-Source applications, while unreplaceable paid key systems are analyzed and compared with free variants. The school’s networks were analyzed in 2026, providing insightful information about the current challenges in an AI-focused environment. During the study, multiple attacks were observed, mitigated and documented. The problem of managing IT-connected devices is addressed through leading open-source solutions, offering public institutions the same capabilities as commercial alternatives without the associated licensing costs. In the end, the authors offer a resolution from start to finish, assessing both countries’ strengths and weaknesses in the infrastructure, providing key employees with the latest advancements in technology.
        Keywords— school IT infrastructure, cybersecurity of public institutions, open source software, real-time monitoring, remote management (key words)

        Speaker: Victor-Emanuil Nitu (UNSTPB)
      • 78
        Modeling the E-Commerce Delivery Layer: From Empirical to Automated Routing

        In recent years, e-commerce systems have developed considerably, and with them the number of deliveries to customers. These deliveries form the final step of the logistics chain, known as last-mile delivery. Within this layer, turning orders into vehicle routes still rests on human expertise: consultations with dispatchers and drivers of courier companies and retailer-owned fleets confirm that routes are built manually and empirically, relying on years of experience. This paper models that decision computationally, formalizing what an automated planner must reproduce from current practice. The underlying problem is the Multi-Trip Heterogeneous Fleet Capacitated Vehicle Routing Problem (MT-HFCVRP), NP-hard as an extension of the Traveling Salesman Problem, which becomes a mixed-integer nonlinear programming (MINLP) problem once compactness, a property of the stop-to-vehicle assignment, is included. Dispatchers achieve compactness intuitively but explore few alternatives, while well-known solvers optimize distance, time and capacity across a vast search space, yet have no native way to express compactness. A hybrid heuristic is therefore proposed that refines an OR-Tools solution through a compactness-preserving local search, under a monetary cost objective calibrated on national transport rates. The approach is evaluated on anonymized real deliveries from an operational week of the largest DIY e-commerce retailer in Romania, with scalability assessed on synthetic instances, against the dispatchers' allocations, a distance-optimal reference and established solvers.

        Speaker: Cezar-Marian Papară (1. Faculty of Computer Science, Alexandru Ioan Cuza University, Iasi, Romania, 2. Faculty of Sciences, Vasile Alecsandri University, Bacau, Romania)
      • 79
        ECHO-Edge: Ontology-Centered Plan Admissibility for Temporary Human Groups in Shared Edge Environments

        Shared pervasive environments combine personal devices, edge nodes, embedded controllers, and cloud services, but adaptation decisions are often formulated around tasks and resources rather than the temporary groups affected by those decisions. Technical feasibility is therefore insufficient because an executable plan may still contravene participant requirements, environment policy, identity constraints, processing-purpose restrictions, or limits on context disclosure and retention. We present Ephemeral Context and Human Objectives for Edge Intelligence (ECHO-Edge), an ontology-centered admission layer that evaluates complete adaptation plans before downstream optimization. Its Web Ontology Language/Resource Description Framework model separates persistent Person entities from time-bounded ParticipantSession entities and represents temporary groups, environments, context, requirements, data flows, execution bindings, and heterogeneous edge resources. The Shapes Constraint Language validation checks structural conformance, while an application-level three-status evaluator materializes per-requirement outcomes, typed explanations, and provenance. The semantic evaluator agreed with an independently implemented procedural test oracle across 46 controlled cases, while 10,000 generated boundary-focused scenarios produced no implementation-level mismatches. On a Raspberry Pi 5, evaluator p50/p99 latency was 0.64/0.89 ms and the complete measured semantic path 3.28/17.71 ms. These results support plan-level semantic admission as a lightweight gate for human-aware shared edge environments.

        Speaker: Gabriel-Catalin Nedianu (Univeristy of Craiova)
      • 80
        Cloud vs. Local LLMs for Privacy-Aware Notarial Document Processing

        The processing of administrative and legal documents increasingly relies on cloud-based artificial intelligence services, raising concerns regarding the handling of sensitive personal information outside institutional infrastructure. This paper investigates an architecture for automated processing of Romanian notarial documents that combines document understanding, structured information extraction, and multimodal Large Language Models (LLMs), with particular emphasis on the trade-off between cloud-based and locally deployed inference.

        The proposed experimental framework addresses two complementary tasks: automated completion of notarial document templates and structured information extraction from Romanian identity cards. Cloud and local models are evaluated under comparable conditions using real notarial templates and synthetically generated identity cards, avoiding the use of real personal data. The evaluation considers extraction accuracy, key-value mapping performance, inference latency, and sensitivity to different prompting strategies.

        Experimental results show that the cloud model achieves a pair F1 score of 91.5% for document completion, compared with 74.4% for the best locally deployed model. However, for multimodal identity-card extraction, both cloud and local models achieve the same mean field accuracy of 87.3%. Prompt adaptation is also shown to substantially improve the performance of smaller local models, with pair F1 increasing from 15.4% to 44.5% in the most significant case.

        The results indicate that locally deployed language and vision models can support privacy-oriented document processing architectures, particularly for sensitive identity information, while introducing measurable trade-offs in accuracy and inference latency. The study provides practical insights into the design of hybrid cloud-local AI infrastructures for document-centric institutional applications.

        Speaker: Mr Laurentiu Neagu (Universitatea Națională de Știință și Tehnologie POLITEHNICA București)
    • Data FAIR in Science & NER & SN & WMN Room 3

      Room 3

      Convener: Florin Bogdan MANOLACHE (Carnegie Mellon University)
      • 81
        Homomorphic Encryption in AI Training Using Mental Health Data

        Medical systems require privacy of the data they process. In the context of training and generating predictions for specific affections, one could reasonably expect that personal physiological indicators would never be exposed or fed in plaintext to an AI algorithm. The real solution in this case lies in using Fully Homomorphic Encryption (FHE), which preserves the data in ciphertext format during all training phases, that conventional pipelines treat as an inevitable cost of computation, but it is eventually eliminating the plaintext exposure.
        This article attaches a layer of complexity on the previous article “Multimodal Burnout Prediction Using Social Media Analysis and Physiological Biomarkers” [1] that created an experimental dataset in order to identify burnout link between speech and physiological state. At this point, the goal was to apply a FHE approach on the same WESAD and NLP dataset and to analyze the performance with different encryption libraries, techniques, depths and using bootstrapping or not. Along with this analysis, we run in parallel another mental health dataset, that contains a different source of data, brainwaves input with an affect-derived burnout-risk index and we compared the time complexity during the training process, for both sets of data.
        The results of the plaintext versus ciphertext were very similar in terms of accuracy metrics. Encrypted computation reproduced plaintext accuracy to within 10⁻⁶–10⁻¹⁴ across every tier and depth tested, confirming that encryption changes computational cost, not predictive accuracy, on either label type. On top of the presented results, we introduced two classification models - Naive Bayes and Decision Tree, a novel approach to adapt tree algorithms to fit the CKKS and preserve a Fully Homomorphic Encryption schema, using the guidelines described theoretically in the “Homomorphic Encryption for Machine Learning inMedicine and Bioinformatics” article [2].
        The ultimate goal of this work was preparing the best encryption practices for a real-world medical application that continuously monitors and predicts mental health risk - burnout, chronic stress, and related conditions - directly on encrypted patient data, from wearable and physiological sensors, without the raw signal ever being visible to the server processing it.

        [1] M.-A. Călin, M. Frunzete, and V. Chetroi, "Multimodal Burnout Prediction Using Social Media Analysis and Physiological Biomarkers," in 2025 24th RoEduNet Conference: Networking in Education and Research (RoEduNet), Chisinau, Moldova, Sep. 17–20, 2025, doi: 10.1109/RoEduNet68395.2025.11208428, https://ieeexplore.ieee.org/document/11208428
        [2] Wood, A.; Najarian, K.; Kahrobaei, D. “Homomorphic Encryption for Machine Learning in Medicine and Bioinformatics”. ACM Comput. Surv. 2020, https://doi.org/10.1145/3394658.

        Speaker: Mrs Madalina-Andreea Calin (Politehnica University of Bucharest)
      • 82
        A Privacy-Preserving Rule-Based Architecture for Personalized Patient Education Guides

        One-size-fits-all patient-education leaflets often ignore care phase and literacy needs, yet unconstrained personalization can over-collect clinical data and weaken clinician oversight.
        This paper presents a privacy-preserving system design and a runnable prototype for composing personalized patient-education guides from version-controlled, clinician-approved content blocks, without generative models at runtime.
        A purpose-limited education profile drives a deterministic IF-THEN rule engine that selects, orders, and excludes modular content.
        Each composition produces a guide draft, a dual (patient- and clinician-facing) composition record, and a recorded delivery role check.
        Role separation, consent-aware notifications, secure delivery, and auditability are design requirements for the networked educational content-delivery service.
        We verify the prototype on ten synthetic dermatology scenarios (psoriasis pathways plus one high-risk atypical-nevus case) through policy walkthroughs, dual composition records, and a guide-size comparison with an all-valid-block baseline (5.5 vs. 8.7 blocks).

        Speaker: Adrian Alexandrescu (Gheorghe Asachi Technical University of Iasi)
      • 83
        Adaptive Control Strategies for Nondeterministic Unstable Systems

        Many real-life systems have nondeterministic behav-
        ior and significant domains of instability. While simple control
        methods can be applied to these systems, in most cases they
        will not achieve the expected results. This paper exemplifies how
        to remap and isolate unstable regions, and explores strategies
        that automatically adapt control parameters to specific situations.
        The strategies are designed to be robust to the system’s nonde-
        terministic response. While the same ideas can be applied to a
        quite large variety of systems, like server load sharing, network
        traffic shaping, or portfolio allocation, the paper exemplifies
        the strategies on a geometrically intuitive robot propulsion
        driving in adverse conditions and using the minimum amount
        of information. A highly unpredictable car robot is controlled to
        exhibit a relatively smooth, stable behavior.

        Speaker: Florin Bogdan MANOLACHE (Carnegie Mellon University)
      • 84
        Gateway Selection in LoRa Networks for Energy Cost Optimization

        LPWANs (low-power wide area networks) are designed for low power, low data rate communication over long distances and have been heavily integrated in the IoT landscape. LoRa is a leading LPWAN technology deployed worldwide, with hundreds of millions of devices operating around the world. LoRa supports hundreds of use cases, including agricultural, forest, assets, and ship monitoring.
        The introduction of LR-FHSS (Long Range-Frequency Hopping Spread Spectrum) further extends LoRa’s applicability and coverage, allowing terrestrial networks to connect with low-earth orbit satellites thus enabling global low-power connectivity.
        In LoRa networks, all end devices communicate directly with a gateway, or a small set of gateways, making the gateway a single point of failure, whether due to battery life or to cybersecurity attacks such as jamming.
        In this work, we focus on selecting two gateways, so that an energy cost function is minimized. We take into account three types of distances: the distance from the other IoT (Internet of Things) devices to the nearest gateway, the distance between the two gateways, and the distance from a gateway to the nearest low orbit satellite, with the later two distances possibly weighted to reflect different power / energy levels for data transmission.
        We then define the objective function and provide the first results on this important localization problem in wireless networks. Due to the relationship to other known site selection problems, we infer that our results will be difficult, if not unlikely to improve.
        The algorithms we present are practical, as they can be implemented and deployed on existing LoRa networks.

        Speaker: Dr Ovidiu Daescu (University of Texas at Dallas)
      • 85
        Beyond RSSI: Disruption-Aware Client-Side Network Quality Assessment for Wi-Fi Access-Point Selection

        SSI-only Wi-Fi selection cannot distinguish a weak
        radio link from a failing local forwarding path or a shared
        upstream outage; it may therefore retain an unusable access
        point or roam without benefit. We present Beyond RSSI, a
        client-only Linux advisory service that combines radio telemetry,
        bounded gateway and service probes, route-aware fault attribution,
        and traffic-aware candidate discovery. A single radio
        cannot test an unassociated candidate’s end-to-end path, so the
        service combines recent scan visibility with time-decayed path
        quality learned while connected. Hysteresis and fault attribution
        gate an explainable Switch/Stay notification, and an accepted
        action is revalidated against fresh association, diagnosis, visibility,
        and trust state. Six deterministic policy scenarios pass, while
        a kernel-backed equal-RSSI two-AP experiment shows scan
        deferral during interactive traffic, discovery of an alternative,
        and recovery from an AP-specific backhaul fault after an exact-
        BSSID switch. Across 75 physical scan windows and 75 no-
        scan controls at fixed and five-location checkpoints using Realtek
        RTL8852CE and Intel AX210 adapters, scans raised pooled
        median gateway RTT from 3.48 to 8.03 ms and the median per-
        window maximum reply gap from 209.9 to 304.1 ms; pooled
        loss was 0.394% versus 0.467% in controls. Four natural BSSID
        changes included three in-window events with 1.06–1.24 s reply
        gaps; one deliberate reassociation caused a 4.36 s gap. In these
        measurements, discovery was a seconds-long operation with
        mostly subsecond packet-level effects, whereas reassociation was
        the larger interruption risk. The contribution is a deployable
        policy that coordinates diagnosis, evidence freshness, acquisition
        timing, and user consent rather than another AP-selection metric.

        Speaker: Cezar Zlatea (Universitatea Politehnica din Bucuresti)
      • 86
        Liquid Neural Networks for Adaptive Difficulty Selection in Personalized E-Learning

        Traditional teaching methods struggle to accommodate differences in prior knowledge, personality, and motivation. While suitable for mass instruction, they cannot match the benefits of individualized tutoring—a gap documented by Bloom's 2-sigma problem. This work leverages artificial intelligence to bring some of these benefits to students at scale. We present AdaptEd, an adaptive learning system for programming in which lessons are structured as directed acyclic graphs. At each branching point, a Liquid Time-Constant (LTC) neural network selects the difficulty level for the next segment based on a 19-dimensional context vector combining psychological traits, session state, and recent performance. The classifier contains only 3,427 parameters and trains in approximately 15 seconds. On our test set, the class-balanced LTC model achieves 75.3% accuracy with a macro-F1 of 0.692. Though a rule-based baseline attains higher overall accuracy, the LTC network generalizes considerably better on the rare but pedagogically important HARD class (F1 of 0.636 versus 0.444). A complementary simulation over 25,000 episodes comparing five decision policies reveals that profile-aware policies produce the highest gains in competency. In contrast, a reward-maximizing Deep Q-Network avoids difficult content and produces the least learning—highlighting the risks of naive reward maximization in adaptive education.

        Speaker: Mrs Sonia Grigor (Technical University of Cluj-Napoca)
    • Conclusions: Keynote after the event
      Convener: Dr Octavian RUSU (Agency ARNIEC/RoEduNet, Alexandru Ioan Cuza University of Iasi)
    • Special Dinner: Special Networking Dinner

      based on invite or for a fee