Despite the pervasive use of deep neural networks (DNN) in numerous domains, their training and deployment mainly resides on computationally intensive pipelines. In turn, this restrains the use of DNN for autonomous robots due to insufficient computational power and the need to regulate power consumption for computation as well as actuation purposes. To address these limitations, this work presents the LEASARD project which aims to increase navigation autonomy while preserving energy autonomy of aerial drones for search-and-rescue applications, that need to leverage DNN for scene perception and drone control. In detail, we advocate the enhancement of sensing and processing tasks through low-energy hardware such as event cameras and Field-Programmable Gate Arrays (FPGAs), so as to increase the efficiency of obstacle-avoidance and robot vision algorithms in a way that capitalizes DNN architectures that are adapted to this new hardware. Overall, the contributions of the project span over algorithm, software and hardware design for improving on-board, drone processing capabilities in adverse scenarios.
Financial regulatory documents are characterized by their fine-grained complexity and pronounced heterogeneity, featuring specialized domain-specific content and diverse structural formats that vary across different regulatory frameworks and jurisdictions. These characteristics challenge modern Question-Answering (QA) systems, which often suffer from limited domain adaptation, poor interpretability, and hallucinatory problems. This work was conducted within a banking and software company, where such challenges directly impact regulatory compliance efforts. Our goal is to introduce RegulQA, a hybrid QA system that can extract accurate, logical, and comprehensible answers from unstructured regulatory documents. RegulQA integrates knowledge graph reasoning, semantic search, and retrieval-augmented generation using large language models. Experimental evaluation shows that RegulQA improves QA performance and significantly reduces hallucination rates. The baseline model employing only the LLM demonstrated a hallucination rate of 24%, whereas the proposed approach, combining knowledge graph reasoning and semantic retrieval with LLM reasoning, effectively reduced the hallucination rate to approximately the half. This integrated approach also yielded the best balanced overall scores across key qualitative attributes, including coverage, non-redundancy, readability, and response quality.
As autonomous robots increasingly operate in private spaces, software-based privacy protection remains opaque and vulnerable to reconstruction attacks. We propose HuLePA-RoV, a human-legible, privacy-aware robotic vision framework that enforces privacy at the physical sensor layer. By default, a mechanically actuated optical element limits perception to low-frequency visual information, enabling navigation and coarse object classification while remaining physically incapable of capturing identifiable facial details. High-resolution sensing is enabled only through authenticated events and visible mechanical reconfiguration, providing users with an immediate and verifiable indication of the robot's privacy state. This approach shifts privacy from invisible software enforcement to observable hardware control, fostering trust and transparency in human-robot interaction.
Rapid digital healthcare transformation adds complex new technological role expectations to nursing, in addition to extensive clinical and holistic care expertise. Nursing education must expand to accommodate substantive new technologies involved in contemporary healthcare services, but knowledge of digital preparedness and impacts on nursing identity and leadership is lacking. This study explores the potential of a playful reflective workshop to explore innovative pedagogical techniques' ability to improve nursing learners' attainment of educational goals and satisfaction, based on a pilot study among 16 nursing students. The findings indicate that participants found digital transformation to be a useful support mechanism for nursing care and education, with high scores for participation, creativity, and depth of reflection. The majority found the workshop meaningful and useful for improving their confidence in using tools/approaches, feeling more prepared, and having improved understanding, with less support for its positive role in leadership identity.
Modern information systems consist of mutually dependent services, and traffic control actions taken during security incidents can contain attacks but may also propagate availability degradation and disrupt business operations. Because the appropriate balance between containment and availability depends on incident context and organizational policy, operators should be supported with multiple response plans rather than a single automatically fixed plan. This paper proposes a decisionsupport method that generates multiple traffic control plans by combining local probabilistic exploration on a service dependency graph (SDG) with multi-objective optimization. Starting from suspected compromised nodes, the method simulates attack propagation using Monte-Carlo trials, aggregates attack paths into a Trie structure, and computes edge-level metrics to select candidate edges for control operations. It then applies NSGA-II over discrete operations on these edges to obtain a Pareto set of plans with respect to residual risk and business impact, where residual risk is estimated by re-weighting the aggregated attack Trie and business impact is estimated by propagating availability degradation along service dependencies. Representative plans, including containment-oriented, availability-oriented, and structurally distinct plans, are selected from the Pareto set and presented with quantitative metrics. A prototype implementation and evaluations on synthetic networks show that the proposed method efficiently generates interpretable plans that visualize containment-availability trade-offs while significantly reducing computation time compared with exhaustive search.
Early detection of systemic financial stress is challenging in fast-moving, nonlinear environments. Traditional early warning systems rely on low-frequency indicators and linear models, limiting their real-time relevance. This paper develops a high-frequency machine-learning framework to monitor systemic financial stress in Europe using daily financial, banking, macro-financial, and sentiment indicators. Diverse data sources are integrated in a strictly time-ordered pipeline to avoid look-ahead bias, with systemic financial stress states defined by threshold exceedances of the ECB's Composite Indicator of Systemic Stress. Feature selection and evaluation are conducted using stability-based regularization, time-series cross-validation, and rare-event metrics. Models combining financial and sentiment indicators outperform linear benchmarks, particularly during periods of intensifying stress. The results suggest that high-frequency machine-learning models can improve the timeliness of systemic stress monitoring and complement existing macroprudential tools.
The RBAC model eases the administration of complex organizations and supports the implementation of basic security principles such as least privilege and separation of duty (SoD). In this work, we present PP-SSP, a Post-Processing Heuristic for the Static Separation of Permissions Problem. The heuristic has been designed to decouple role discovery from constraint enforcement phase, so that any existing rolemining algorithm can be selected at the beginning, making then the framework flexible and easily adaptable to different mining techniques. We evaluate the heuristic against a number of public datasets, report the results obtained, and discuss its performance in terms of different metrics, showing its efficiency and effectiveness.
Modern cyber threats are increasingly multilayered, propagating from identity compromises to network-level anomalies. Traditional SOC tools often fail to detect correlated events across domains. This research presents a GenAI-inspired Agentic AI framework integrating user behavioral analytics and network fabric telemetry into a unified anomaly detection and mitigation system. Using PCA and SVM for dimensionality reduction and classification, this framework autonomously detects anomalies, correlates cross-layer signals, and applies mitigation policies. The system demonstrates improved detection, proactive enforcement, and visual analytics support for forensic analysis.
Software testing often depends on a test oracle to determine whether or not an output is correct. However, many real-world systems lack reliable or affordable oracles. Metamorphic testing (MT) alleviates this problem by using metamorphic relations (MRs), which examine the correctness of relations between outputs produced from related inputs across multiple executions. While the effectiveness of MT depends heavily on the quality of MRs, existing MR evaluation is typically guided by informal or study-specific criteria, limiting standardization and repeatability, especially for complex and AI-integrated systems. This paper proposes a structured MR-evaluation rubric that operationalizes key criteria to support consistent and repeatable assessment. We conducted an empirical study on nine systems under test (SUTs) spanning three levels of complexity, using fresh sets of LLM-generated MRs for each SUT. We then compared evaluations from human experts against two widely used large language models (LLMs), namely ChatGPT and Gemini, by configuring these models as MR evaluators using the same rubric. The results show that human and LLM evaluators are more consistent when assessing simpler SUTs, but their judgments increasingly diverge as system complexity grows and evaluation criteria become more varied. The study further reveals clear differences in LLM-evaluation behaviors, indicating that evaluator choice can affect MR assessment outcomes in complex settings. Overall, the proposed rubric provides a standard framework for MR evaluation, and supports a systematic comparison between human and LLM evaluators, offering empirical evidence on when LLM-based MR evaluation is more reliable, and where additional care is needed as SUT complexity increases.
Accurate prediction of drug sensitivity in cancer cell lines is vital for precision oncology and patient-specific therapies. However, many computational approaches fail to integrate multi-modal biological and chemical features and often struggle with high-dimensional, imbalanced pharmacogenomic data, limiting predictive accuracy and interpretability. To address these challenges, we developed a machine learning framework that integrates pharmacogenomic profiles-including mutation status, copy number alterations, and microsatellite instabil-ity-with molecular fingerprints and descriptors of 85 anticancer drugs, generated using PaDEL from SMILES strings. Data from 40 breast cancer cell lines in the Genomics of Drug Sensitivity in Cancer (GDSC) dataset were employed. A threestage feature selection strategy combining Boruta, mRMR, and XGBoost was applied to reduce drug feature dimensionality while retaining 130 cell line features. Multiple models were trained, and LightGBM, optimized with grid search, class weighting, and 3-fold cross-validation, demonstrated superior performance in handling severe class imbalance (233 sensitive vs. 3167 resistant samples). LightGBM achieved training AUROC $=0.9455$, AUPRC $\boldsymbol{=} \mathbf{0. 5 1 4 8}$, Accuracy $\boldsymbol{=} \mathbf{0. 8 4 1 5}$, F1-score = 0.4481, Recall = 0.9409, and MCC = 0.4732, underscoring its suitability for sparse biomedical datasets. Model interpretation with SHapley Additive exPlanations (SHAP) highlighted BRCA-related features, identifying cnaBRCA25 (not mutated) as a resistance marker and cnaBRCA47 (mutated) as a context-dependent biomarker, consistent with their roles in DNA repair pathways. Overall, this framework demonstrates the value of multi-modal integration and interpretable machine learning in pharmacogenomics. While results are promising, validation on larger and independent cohorts is essential to establish clinical relevance.
Energy community simulations based on optimization models are increasingly used to evaluate energy sharing strategies among prosumers. However, integrating such computationally intensive models into user-facing applications presents significant challenges, including limited solver concurrency, long execution times, and the need for scalable and secure infrastructures. This paper presents a scalable architecture for delivering energy community simulations as a service. The proposed system integrates optimization models defined in Pyomo and solved using Gurobi within a distributed environment based on remote compute servers. To address bounded solver concurrency and avoid blocking in the API layer under concurrent requests, a job-based execution model is introduced, enabling asynchronous task handling and controlled concurrency. The architecture is implemented using a containerized approach, combining an API layer for request handling with a distributed execution layer for optimization tasks, and infrastructure as code to ensure reproducibility and security. Experimental evaluation demonstrates the system's ability to manage concurrent simulation workloads while maintaining responsiveness and respecting solver limitations. The results highlight how appropriate system design enables the practical deployment of optimization-based energy simulations, bridging the gap between research models and real-world applications.
This research develops a web-based visualization system, named OSSim (Operating System Simulator), for operating system (OS) resource management. OSSim is designed as an interactive platform that visually demonstrates the execution of various OS resource management algorithms. The visualization and interaction design are structured to support systematic algorithm analysis and instructional demonstration. OSSim also serves as an aid for instructors to easily demonstrate diverse examples, configure parameters during lectures, and visualize algorithm execution results in real time for more effective instruction. This paper further enhances the system architecture and interaction design to support comparative algorithm analysis and multi-modal execution control. We implement additional algorithm execution modes and advanced visualization mechanisms, such that users can execute algorithms step by step, revert to previous execution states, or observe an animated, automatically advancing execution process. Throughout execution, the system explicitly visualizes evolving system states and corresponding results at each stage, facilitating a deeper understanding of algorithm behavior. Moreover, for a given workload, OSSim simultaneously presents the execution results of multiple algorithms, allowing users to conveniently compare and analyze their differences. Enhanced OSSim was applied in an undergraduate OS course. A pre-test-post-test experimental design was adopted to evaluate learning effectiveness. Experimental results show that students in the experimental group who used OSSim achieved greater score improvements and exhibited a 14.6% higher score increase rate compared to the control group that did not use OSSim. As a result, the use of OSSim demonstrated a positive instructional trend and suggests the practical value of structured multi-modal execution control with comparative visualization in supporting systematic algorithm analysis.
Traffic congestion significantly impacts safety and urban livability in smart cities, motivating the development of accurate Traffic Flow Prediction (TFP) systems. Traditional deep learning approaches typically rely on centralized training, which is difficult to scale in distributed Internet of Things (IoT) environments. To address these limitations, decentralized paradigms such as Local Learning (LL) and Federated Learning (FL) enable on-device training and collaborative model updates while preserving data locality. For real-time TFP, the inherently non-stationary nature of traffic data necessitates continuous model adaptation, making online federated learning essential for scalable and collaborative deployment. However, this setting remains challenging because traffic data are typically non-IID across clients, with local patterns varying significantly across locations and devices. This paper presents an exploratory study of online federated learning for TFP on resource-constrained IoT devices. Using a GRU-based network as a common backbone, the Online Federated Learning paradigm is benchmarked relative to Centralized and Local Learning as reference baselines. A performance evaluation was conducted by evaluating RMSE and MAE on the PEMS-BAY dataset. Robustness to non-IID data is further assessed using FedProx and SCAFFOLD. Results show that LL achieves the lowest prediction error, whereas FL degrades as the number of local epochs increases, and non-IID mitigation strategies provide limited improvements under low-latency constraints. Overall, online federated learning is a viable approach for real-time TFP, but its performance is highly sensitive to client heterogeneity.
Personalization is essential for effective conversational Artificial Intelligence (AI). However, older adults remain underrepresented in existing research of AI on personalization. In this study, we introduce ElderBench, a synthetic benchmark dataset of 523 samples spanning 21 need categories for olderadults, and a two-level evaluation framework for personalizing senior-focused large language models (LLMs). We fine-tune five open-source LLMs (Llama-2-7B, Llama-3.1-8B, Qwen2.5-7B, Gemma-2-9B, DeepSeek-R1-1.5B) using GraphRAG-style simulated retrieved context. Level 1 evaluation reports referencebased metrics (BLEU, METEOR, ROUGE, BERTScore). Level 2 uses GPT-4o as an LLM judge to score relevance, context usage, empathy, helpfulness, coherence, and overall quality. Results show that fine-tuning substantially improves weaker baselines at senior-focused personalization, whereas stronger baselines exhibit a trade-off: higher context usage can slightly reduce relevance and helpfulness. These findings provide practical guidance for selecting and adapting open-source LLMs for personalization for older adults.
Access to culture and childhood literacy in Brazil faces linguistic and technological barriers that limit contact with international works. This paper presents an automated system for the translation and reconstruction of comic books, which integrates speech balloon segmentation, Optical Character Recognition (OCR), visual reconstruction via inpainting, and text-to-speech synthesis. The primary objective is to democratize access to reading and promote multimodal literacy. We conducted a study with ten early childhood education teachers in Brazil, simulating scenarios that ranged from individual use on tablets to collective mediation via a television connected to a computer. The results indicated excellent usability, achieving a System Usability Scale (SUS) score of 89.5. User experience, evaluated through the UEQ-S, also reached the highest classification of excellence across all dimensions. Furthermore, specialists confirmed the system's strong pedagogical alignment with the National Common Curricular Base (BNCC), particularly in fostering imagination and cultural diversity. Therefore, the proposed technology configures itself as a promising tool for reducing educational inequalities, aligning with the United Nations Sustainable Development Goals (SDG) 4 and 10.
The increasing complexity of cyber threats and the growing autonomy of digital infrastructures demand security solutions that are adaptive, trustworthy, and resilient by design. This paper presents the outcomes of the PANACEA project, a PRIN 2022 initiative aimed at advancing the state of the art in model-based self-protecting systems. PANACEA integrates dataefficient intrusion detection, decision-theoretic intrusion response, and cyber-resilient autonomic management within a unified architectural framework. On the detection side, the project introduces adaptive models capable of operating under delayed and partially available labels, as well as a transformer-based semi-supervised learning approach that leverages self-supervised pretraining to significantly reduce reliance on labeled data while preserving competitive performance. On the response side, the project formalizes mitigation as a risk-aware decision process grounded in Timed Competitive Stochastic Games automatically synthesized from augmented Attack-Defense Trees (ADT), enabling cost-sensitive and explainable action selection. To address the vulnerability of the autonomic manager itself, PANACEA proposes a permissioned blockchain-based coordination layer that ensures consensus on system state and response decisions, tolerating the compromise of a bounded subset of components. All contributions are validated within a programmable cyber-range supporting controlled experimentation and end-to-end resilience assessment. The results demonstrate the feasibility of integrating adaptive learning, formal response planning, and distributed trust into a coherent self-protecting framework.
AI-based perception systems for smart-city applications require not only accurate models but also dependable execution of the perception pipeline. While the Robot Operating System 2 (ROS 2) middleware is widely adopted for its flexibility, its asynchronous and event-driven execution model introduces non-deterministic behaviour and limited temporal and integrity guarantees, which complicate dependability in safety-relevant scenarios. This paper presents a work-in-progress toward a dependable middleware that preserves ROS 2 abstractions while providing deterministic and time-bounded execution, explicit data-flow control, and integrated observability. We describe the middleware architecture and report initial experiences from deploying a multimodal YOLO-based people-detection application on both ROS 2 and the proposed middleware, together with a preliminary characterization of baseline timing and resource behaviour.
Logistics Cyber-Physical Systems (LCPS) generate large volumes of regulatory and operational texts that encode early signals of safety risks. Converting short, noisy, and domain-specific records into actionable intelligence is difficult due to industrial semantic drift and the limited auditability of black-box predictors. This paper proposes Neuro-Symbolic Logistics Risk Awareness (NS-LRA), a dual-channel framework that integrates lightweight semantic perception with constraint-aware topological reasoning. NS-LRA first maps raw texts to a standardized schema of $K=20$ risk nodes using a dual-weighted embedding mechanism that combines TF-IDF and Word2Vec to mitigate short-text sparsity. It then constructs a directed risk graph by fusing co-occurrence evidence with a domain constraint mask, and derives hierarchical propagation via ISM level partitioning with deep-driver identification via MICMAC analysis. We evaluate NS-LRA on $N=8,435$ records, validated against an annotated subset $(\mathcal{D}_{\text{ann}}=1,500)$) with Fleiss' $\kappa=0.82$ and an expert-defined gold graph. NS-LRA achieves Micro-$\mathrm{F} \mathrm{1}=\text{0. 8 7 9}$ for risk mapping and approximately $22 \times$ lower perrecord CPU latency than fine-tuned BERT on the same test split under the same environment. For topological inference, NS-LRA reports $\text{E P}=\text{0. 9 2 4}$ and $\text{T C}=\text{0. 9 5}$ against the gold graph. These results indicate that NS-LRA can provide an efficient and traceable pipeline for proactive risk governance in LCPS.
