Carlo Sansone

The expansion of solar photovoltaic capacity demands increasingly advanced operation and maintenance strategies. Conventional fault detection methods build their reference for expected production either from on-site instrumentation or from the design specification of the installation, and are exposed to sensor drift and to plant ageing. Machine learning offers an alternative route, in which the reference is learned from the operational record, but is constrained by data scarcity and the commercial sensitivity of production information. Federated learning addresses both constraints, yet its use in this domain has focused on power estimation.The present study draws on four years of data from ten commercial photovoltaic plants, comparing twenty-eight models and six collaboration regimes, two of which withhold the evaluated site to isolate the cold start of a new installation. Detection operates on a bidirectional residual indicator retaining both under- and over-production.Learned models estimate power roughly six times better than naive references; on detection the margin narrows, nine are equivalent within 0.02 ROC-AUC, and the difference from a performance ratio reference is not resolved. Sharing is most valuable where local history is absent: an unseen site receives a learned power reference from day one, at an error level an isolated site reaches only after six months, and there it outperforms training on pooled raw data. The advantage does not extend to detection, where no regime improves upon local training: collaboration supplies a power reference at commissioning rather than better diagnosis, and its bandwidth cost undercuts pooling only for the smallest families.

User authentication is more and more crucial in everyday life, with researchers developing sophisticated, yet easy-to-use, approaches to enforce secure authentication while providing a smooth user experience. This is particularly true in mobile authentication, in which the spread of AI-capable smartphones is opening up several biometric-based authentication methods, such as face and lips recognition. A more recently emerging setup consists of the matching of a spoken passphrase with lips movements. Despite being promising and well-suited for mobility authentication, the matching operation is not trivial, and the research around this topic is limited, as no publicly available dataset is available for experimental purposes. In this paper we thus introduce BioVid , a novel privacy-aware multimodal biometric dataset for user recognition from smartphone-recorded videos in real-world conditions. BioVid consists of 650 recordings from 43 participants, where each clip captures a user pronouncing predefined passphrases while simultaneously acquiring audio and lips movements. The dataset is designed to support dual-factor biometric authentication, combining speaker-unique identity (UID) and passphrase verification, and explicitly models realistic variability in device type, recording distance, background, and ambient noise. Together with the data, we provide a detailed statistical analysis of the linguistic, visual, and acoustic properties of the data, highlighting its diversity and suitability for robust biometric system evaluation. Moreover, to test the quality of the collected dataset, we organised an international challenge on multimodal user authentication leveraging BioVid . In this paper we thus also describe the challenge protocol, baseline systems, and evaluation metrics, while also summarising the best-performing submitted approaches as measured by Equal Error Rate (EER). BioVid and the accompanying challenge aim to establish a reproducible benchmark for multimodal biometric authentication on consumer-grade smartphones and to foster further research on privacy-aware, real-world biometric systems.

Artificial Intelligence (AI) models, particularly Deep Neural Networks (DNNs), have revolutionized the field of Computer Vision, leading to new approaches to complex problems, ranging from autonomous driving to biometric recognition. However, the shift toward the Machine Learning as a Service (MLaaS) paradigm, and more recently AI as a Service (AIaaS), has exposed the Machine Learning supply chain to critical vulnerabilities, such as backdoor attacks, raising concerns about model integrity and forensic analysis in practical deployment settings. While existing literature covers broad attack vectors, the rapid evolution of model architectures, especially in classification tasks related to Computer Vision, has left the scientific landscape fragmented.This study provides a systematic literature review and presents a comprehensive taxonomy of backdoor attacks targeting Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) in a classification setting, with implications for forensic understanding and model auditing. By analyzing research trends, we document a decisive shift in scientific interest toward transformer-based architectures, while, by providing a unified taxonomy, we highlight the structural vulnerabilities unique to each. In particular, the taxonomy presents 38 prominent attacks, categorized based on their trigger design, injection vector, and labeling strategy, illustrating how these threats have evolved from static pixel patches to sophisticated exploits of the self-attention mechanism.Furthermore, a replicability audit of four major open-source frameworks assesses the current state of attack benchmarking. The results reveal a significant maturity gap: while early CNN-based attacks are mostly well integrated and replicable, recent ViT-based threats frequently lack public implementations, hindering independent validation and limiting reliable forensic evaluation and comparison. This study serves as a foundation for future analysis of the backdoor attack landscape, providing a dual-architecture perspective to address both the unique and shared vulnerabilities of modern vision models in realistic and operational contexts. It further outlines practical considerations and preliminary guidelines to support the forensic auditing, evaluation, and benchmarking of potentially compromised models in real-world settings.

Video-based alarm notification systems play a critical role in safety-critical applications such as fire detection and pedestrian monitoring, because they allow for prompt intervention in the event of dangerous situations. However, existing approaches often struggle to balance high sensitivity and specificity with real-time performance, particularly in complex scenes. To address these limitations, we propose a novel method enabling Fast Alarm Notification with Task-Aware Spatio-Temporal Image Classification (FAN-TAST-IC). It is an innovative and efficient framework that combines a lightweight task-specific object detector with a pre-trained Vision-Language Model (VLM) encoder and a binary classifier. Unlike end-to-end multimodal systems, FAN-TAST-IC leverages the VLM solely as a frozen visual feature extractor, preserving its rich semantic knowledge while ensuring computational efficiency. The object detector performs a rough but real-time filtering of the temporal frames and spatial positions where objects can be. This filtering is then refined in two distinct ways, depending on the time constraints of the specific application, either to discard temporally incoherent detections or to confirm those with high confidence. The selection of such candidates drastically reduces the input space for the VLM and classifier to dubious detections only. Thus, the latter is trained on detector-guided positive and negative samples, enabling precise alarm validation, improving specificity, and preserving sensitivity without requiring extensive labeled datasets or fine-tuning the VLM. Experiments on fire and pedestrian detection tasks demonstrate the effectiveness of our method since FAN-TAST-IC consistently outperforms all the compared approaches, achieving superior F-scores and precision on challenging benchmarks while maintaining real-time capabilities.

In recent years, Artificial Intelligence (AI) has gained increasing traction in medical image analysis. Novel methodologies based on Deep Neural Networks (DNNs) have increasingly enabled the development of solutions that leverage data from multiple sources, including multimodal and synthetic data. Despite notable achievements, current AI systems often struggle with robustness and generalizability, particularly when deployed in real-world clinical scenarios characterized by data heterogeneity, incompleteness, and distribution shifts. This highlights the urgent need for resilient AI, that is, AI capable of maintaining consistent and reliable performance across diverse and imperfect conditions. In this paper, we explore medical image computing from a resilience-oriented perspective, focusing on two key sources of complexity: multimodal and synthetic data. For the former, we address two major challenges: (i) the design of fusion strategies that can effectively handle missing modalities; and (ii) the integration of complementary information to ensure robustness across heterogeneous subject populations. For the latter, we emphasize the critical role of biological plausibility in both the generation and evaluation of synthetic data, proposing physiologically informed pipelines that improve learning in data-scarce settings. We introduce innovative methodologies that advance the state-of-the-art, including a cross-modality calibration mechanism for incomplete acquisitions and a pharmacokinetics-driven synthetic augmentation framework. Our contributions aim to promote robustness, fairness, and generalization, ultimately fostering the development of more resilient AI systems in medical imaging.