Archivio della ricerca della Scuola Superiore Sant'Anna
Not a member yet
26957 research outputs found
Sort by
How polluting is professional football? The environmental footprint of a football match calculated with a life cycle assessment
No full textIn-network AI Capabilities through Hardware Switches and Optical Computing
No full textIn-band telemetry could expand its functionalities through AI. However, practical limitations in hardware switches prevents this upgrade. Here, we discuss how an hardware-aware design can bring in-network AI through ASIC switches and optical computing
Le misure di interdizione e controllo giudiziale nella prevenzione della criminalità non mafiosa
No full textBehavior-Based Anomaly Detection in Access and Usage Control for Smart Home Environments
No full textThis paper proposes to enhance the security of smart home environments by integrating a behavior-based component in access and usage control systems to perform anomaly detection using machine learning. This component dynamically assesses each access request by assigning it an anomaly score based on its deviation from learned patterns of normal behavior. This enables context-sensitive and risk-aware policy enforcement, improving the system’s responsiveness to unusual or suspicious behavior. To train and evaluate anomaly detection models in the absence of real-world labeled datasets, we introduce an ontology-driven synthetic dataset generation method. This ontology encodes devices, contextual attributes, and subject behavior patterns to support scalable and customizable dataset creation across various domains. Based on this ontology, we generate different datasets of access requests for smart home scenarios and conduct an evaluation of standard performance metrics of both supervised and unsupervised machine learning models. Among the unsupervised models, Deep SVDD achieved the best results, with an accuracy of 88%, demonstrating strong generalization to unseen anomalous behavior. Supervised models, particularly SVM, reached 95% accuracy due to their training on a labeled dataset. While supervised models excel under controlled conditions, unsupervised models, especially Deep SVDD, proved more practical for real-world deployments where labeled anomalies are limited or unavailable. Our findings highlight the value of integrating anomaly detection into access and usage control systems and provide a reusable framework for detecting anomalous behavior patterns in smart environments
The Brussels Sphinx’s Riddle. What is a high-risk AI System?
No full textThe article examines the conceptual and normative “riddle” posed by art.
6 of the EU Artificial Intelligence Act (AIA) in defining “high-risk” AI
systems (h-AISs). It argues that the combination of a horizontal, tech-
nology-neutral framework with a risk-based classification generates sig-
nificant interpretative uncertainty and undermines legal certainty. After
situating the AIA within the broader EU product-safety regime and the
New Legislative Framework, the contribution meticulously examines in
detail the critical issues arising under Art. 6 AIA. These range from para.
2 recalling the Annex III list of high-risk AI systems, which does not rest
on an objective assessment of risk, to the exceptions in paras. 3 and 4, and
the cross-reference to Union harmonisation legislation in Annex I. Par-
ticular attention is paid to contested notions such as “safety component”
and “third-party conformity assessment required”, illustrated through
case studies (e.g. security mobile robots, humanoid robots, drone docking
stations). The article concludes that this unstable definitional architecture
undermines consistent application, equal treatment across sectors, and ef-
fective incentives for innovation
Dopo la dissoluzione della Duplice Monarchia austroungarica: la questione asburgica nell’Austria repubblicana e in Ungheria
No full textThe essay deals with the legal treatment of the legacy of the Habsburg monarchy after the dissolution of Austria-Hungary in 1918. In so doing, it develops a comparative analysis of how the Habsburg question was addressed in the Republic of Austria and in Hungary. In Austria, a defining feature of the constitutional order was a militant understanding of the republican principle, with constitutional and legislative provisions specifically aimed to prevent a restoration of the monarchy. In Hungary, the general attitude towards the Habsburg monarchy was rather ambiguous, as the form of state officially remained unchanged until 1946. By then, the Habsburg question had lost much of its relevance
Mitigating Noise Effects in Photonic Neural Networks Using Adaptive Quantization
No full textPhotonic neural networks offer energy-efficiency but suffer from noise-induced low precision. We propose AQ-PANN, which learns a quantization step size to mitigate noise. Experiments on SVHN show strong performance across bitwidths under different noise levels
Human-Centered Geodesics for Motion Planning
No full textThis paper addresses the challenge of designing human-like reference trajectories for exoskeleton-aided rehabilitation, with a focus on mimicking human joint coordination while addressing clinical requirements. Redundant kinematic chains in human biomechanics pose challenges to trajectory planning: state-of-the-art algorithms often do not explicitly address the problem of replicating natural movements nor do they provide a suitable performance over a wide range of human motions. To address this challenge, this paper proposes a geodesics-based computational method that incorporates joint-level constraints, in addition to energy and level of comfort criteria to solve the problem of redundancy and better emulate human movements. Using upper-limb data retrieved with an exoskeleton platform, the advanced method demonstrated significant performance gains over standard approaches like the minimum-jerk model and cubic polynomial planning, and leads to human-like trajectories, while closely aligning with human demonstrations, both at the configuration (joints) and task-space (hand) level. In particular, we provide detailed comparisons across various motion types and subjects, demonstrating the versatility of the proposed method and its strong potential for application in clinical and assisted living settings
