Cyber-Physical Systems (CPS) drive modern critical infrastructures by tightly integrating physical processes with computation and communication networks. This convergence exposes CPS to sophisticated cyber threats propagating across physical, control, and network layers, where stealthy and multi-stage attacks manifest through weak and distributed signals, challenging centralized intrusion detection systems and monolithic AI models that struggle to scale across heterogeneous subsystems and lack the interpretability required in safety-critical environments. This paper proposes a hierarchical multi-agent AI framework for automated cybersecurity assessment with human-understandable explanations in networked CPS. The framework dynamically instantiates a task-specific agent hierarchy from a natural language description of the target system, aligning the security analysis process with the underlying CPS architecture. Specialized agents perform fine-grained analysis of heterogeneous data sources, subsystem supervisors aggregate and contextualize local findings, and a deliberative round-table consensus mechanism enables cross-subsystem correlation for detecting coordinated and stealthy attacks. The framework is evaluated using four LLMs (Qwen 3 4B, Qwen 3 8B, Llama 3.1 8B, and Llama 3.3 70B) across three datasets (PicoDomain, CERT r5.2, and SWaT), achieving perfect recall for models with 8B parameters and above (i.e., zero missed attacks), while reaching an accuracy of 90.18% on PicoDomain, 92.77% on CERT r5.2, and 96.97% on SWaT. An ablation study confirms the effectiveness of the cross-subsystem consensus mechanism, demonstrating substantial precision gains when collaborative deliberation is enabled. Overall, this work establishes hierarchical multi-agent architectures as a scalable, interpretable, and structure-aware foundation for AI-driven cybersecurity in networked CPS.

Abstract Smart cities, as a constantly evolving field, offer numerous opportunities for advancement and innovation. In particular, more and more specific sectors within smart cities currently require increased attention and development. Notably, the integration of emerging technologies holds promise for significant progress in efficiency and management across various urban issues, with urban waste management being a key focus of this research. In this paper, we propose a new methodology for intelligent waste management combining graph-based modeling and advanced machine learning techniques. A case study of a municipality in the Melbourne area (Australia) is examined, where smart bins equipped with fill-level sensors have been deployed, and their data made publicly accessible. The core of this investigation lies in utilizing sophisticated machine learning models, including Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), for predicting bin fill levels. Additionally, we introduce a graph-based spatial representation in which bins are modeled as nodes of a weighted graph, and graph embedding techniques are employed to capture and encode spatial dependencies among them. The findings validate the effectiveness of the proposed methodology, achieving a Mean Absolute Error (MAE) of 0.16, a Root Mean Squared Error (RMSE) of 0.21, a Mean Squared Error (MSE) of 0.044, and an R 2 of 0.83, which outperforms the reference baseline. Our results highlight how the integration of spatial information through graph embedding holds promise for improving predictions in smart waste management. Furthermore, the study suggests a transferable methodological framework, applicable at the architectural level to smart city challenges where spatially distributed entities exhibit interdependent temporal dynamics, pending empirical validation across further domains and cities.

Nowadays, our security and privacy are strongly threatened by malware programs which aim to steal our confidential data and make our systems out of service, among other things. While traditional signature-based malware detection methods or statistical analysis have proven to be ineffective and time-consuming, recently data-driven Artificial Intelligence (AI) techniques, i.e. Machine Learning (ML) and Deep Learning (DL) approaches, have been successfully applied leveraging the behaviour of malware in terms of API calls, and achieving promising performances. However, their black-box behavior leads to a lack of explainability thus preventing their application in real world scenarios. In light of this, eXplainable Artificial Intelligence (XAI) methodologies and tools can be effectively embedded within an AI-based malware detection process in order to make more understandable the produced results. In this paper, we propose a XAI framework for behavioral malware detection problems and evaluate the usefulness of four XAI methods (SHAP, LIME, LRP and Attention mechanism) on three datasets with different size, sequence length and number of classes, by which we could evaluate the strengths and weaknesses – from effectiveness and efficiency point of views – of recurrent deep architectures (i.e. Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) models), and their applicability in the modern Cyber Security (CS) scenarios.

In this paper, we address issues associated with the effective management of handover events in satellite-enabled 5G network infrastructures. Namely, we devise a strategy for dynamically allocating 5G gNB available resources in the presence of a constellation of LEO satellites, based on several parameters collected and dispatched by an ad hoc orchestration platform. We propose to leverage a Combinatorial Multi-Armed Bandit approach to design a resource allocation game that helps make decisions over time under uncertainty conditions related to the incidence of several factors that can determine the quality of experience perceived on the user equipment side. The introduced approach does represent a pioneering one, since it allows us to model a joint optimization task as a competitive game in which agents typically share resources with other agents instead of occupying them exclusively. The designed task allows to dynamically enable and disable channels, taking care of the relationships with the lower layer, transparently managing the required handover operations, and considering possible interference with other channels. We discuss an implementation of the proposed solution in a simulated environment. We also analyze the performance it attains, by measuring both its efficiency and efficacy in a trial setup reproducing a scenario with near real-time temporal requirements. Results show that the proposed approach has a linear trend in terms of running time with respect to the number of user equipments and gNbs involved, while achieving a sub-optimal solution of the handover task in around 20–30 rounds.

Deep learning, owing to its transformative influence across a myriad of sectors, has recently made its foray into the legal domain, instigated by the surge in digitization. Among the multitude of applications in this space, legal document classification emerges as a pivotal yet complex undertaking. Legal texts, characterized by unique domain-centric semantics and intricate linguistic patterns, necessitate precision-driven classification systems for numerous practical implications. This paper illuminates the challenges and opportunities in automating the classification of European Union (EU) legal documents, emphasizing the interrelationships among statutes and the hierarchical nature of legal references. In this context, we introduce a novel graph data modeling technique that adeptly marries content-centric indicators with the relational dynamics inherent among diverse legal documents. Central to our approach is a framework that melds text embeddings with graph neural networks for the classification of legal documents aligned with their subject-based directories. Empirical evaluations on the EU law dataset underline the efficacy of our model across varying granularities, from general thematic categories to intricate subtopics. This endeavor not only augments the comprehensibility and accessibility of EU jurisprudence but also holds significant implications across regulatory compliance, legal research, and policy formulation, underscoring the potential of deep learning in reshaping legal paradigms.

In the railway domain, rolling stock maintenance affects service operation time and efficiency. Minimizing train unavailability is essential for reducing capital loss and operational costs. To this aim, prediction of failures of rolling stock equipment is crucial to proactively trigger proper maintenance activities. Indeed, predictive maintenance is a golden example of the digital transformation within Industry 4.0, which affects several engineering processes in the railway domain. Nowadays, it may leverage artificial intelligence and machine learning algorithms to forecast failures and schedule the optimal time for maintenance actions. Generally, rail systems deteriorate gradually over time or fail directly, leading to data that vary extremely slowly. Indeed, ML approaches for predictive maintenance should consider this type of data to accurately predict and forecast failures. This paper proposes a methodology based on Long Short-Term Memory deep learning algorithms for predictive maintenance of railway rolling stock equipment. The methodology allows us to properly learn long-term dependencies for gradually changing data, and both predicting and forecasting failures of rail equipment. In the framework of an academic-industrial partnership, the methodology is experimented on a train traction converter cooling system, demonstrating its applicability and benefits. The results show that it outperforms state-of-the-art methods, reaching a failure prediction and forecasting accuracy over 99%, with a false alarm rate of ∼0.4% and a mean absolute error in the order of 10−4, respectively.

Nowadays, really huge volumes of fake news are continuously posted by malicious users with fraudulent goals thus leading to very negative social effects on individuals and society and causing continuous threats to democracy, justice, and public trust. This is particularly relevant in social media platforms (e.g., Facebook, Twitter, Snapchat), due to their intrinsic uncontrolled publishing mechanisms. This problem has significantly driven the effort of both academia and industries for developing more accurate fake news detection strategies: early detection of fake news is crucial. Unfortunately, the availability of information about news propagation is limited. In this paper, we provided a benchmark framework in order to analyze and discuss the most widely used and promising machine/deep learning techniques for fake news detection, also exploiting different features combinations w.r.t. the ones proposed in the literature. Experiments conducted on well-known and widely used real-world datasets show advantages and drawbacks in terms of accuracy and efficiency for the considered approaches, even in the case of limited content information.

Hard disk drive failures are one of the most common causes of service downtime in data centers. Predictive maintenance techniques have been adopted to extend the Remaining Useful Life (RUL) of these drives, and minimize service shortage and data loss. Several approaches based on machine and deep learning techniques have been proposed to address these issues, mostly exploiting models based on Self-Monitoring analysis and Reporting Technology (SMART) attributes. While these models have proven to be reliable, their performance is affected by the lack of information about the proximity of disk failure in time. Moreover, many of these techniques are sensitive to the highly unbalanced nature of existing data-sets, in terms of good to failed hard disk ratio. In this article, we propose a LSTM (Long Short Term Memory)-based model combining SMART attributes and temporal analysis for estimating a hard drive health status according to its time to failure. Our approach outperforms state-of-the-art methods when evaluated on two data-sets, one containing hourly samples from 23395 disks and the other reporting daily samples from 29878‬ disks. Experimental results showed that our approach is well suited to data-sets with different sampling periods, being able to predict hard drive health status up to 45 days before failure.

Satellite communications (SatCom) have a role of advanced service enablers in the new virtual networks, following 3GPP specifications. Specifically, the satellite peculiar characteristics are of paramount importance to dynamically activate capabilities, such as multicast and broadcast channels, sudden traffic offloading, capacity bonding, and cost-efficient coverage of uncovered areas. A key aspect to achieve a seamless and efficient integration between satellite and terrestrial infrastructures is to make satellite resource management dynamic and with a centralised control of a single orchestrator that has visibility of the end-to-end network. Considering the adoption of Software Defined Networking (SDN) and Network Function Virtualization (NFV) paradigms and the upcoming 5thgeneration of mobile communications (5G), this paper presents the ongoing work within the European Space Agency (ESA) ANChOR project, whose main output will be a data-driven Network Controller and Orchestrator for SatCom networks. This tool will be based on Artificial Intelligence (AI) / Machine Learning (ML)-based techniques to properly allocate resources exploiting some feedback knowledge of the network and it aims to support different services over 5G integrated satellite-terrestrial networks.

Nowadays, Dynamic Contrast Enhanced-Magnetic Resonance Imaging (DCE-MRI) is increasingly succeeding as a complementary methodology for breast cancer, with Computer Aided Detection/Diagnosis (CAD) systems becoming essential technological tools to provide early detection and diagnosis of tumours. Several CADs make use of machine learning, resulting in a constant design of hand-crafted features aimed at better assisting the physician. In recent years, Deep learning (DL) approaches raised in popularity in many pattern recognition tasks thanks to their ability to learn compact hierarchical features that well fit the specific task to solve. If, on one and, this characteristic suggests to explore DL suitability for biomedical image processing, on the other, it is important to take into account the physiological inheritance of the images under analysis. With this goal in mind, in this work we propose "3TP U-Net", an U-Shaped Deep Convolutional Neural Network that exploits the well-known Three Time Points approach for the lesion segmentation task. Results show that our proposal is able to outperform not only the classical (non-deep) approaches but also some very recent deep proposal, achieving a median Dice Similarity Coefficient of 61.24%.