Workshop Program
Where: University of Cagliari - Facoltà di Ingegneria e Architettura - Room TBD When: Monday, February 9th, 14:30-17:30
Schedule
14:30-14:45
Workshop Opening and Greetings
14:45-15:45
Invited Guest Talk
Speaker: Giulio Rossolini
Title: “Are Adversarial Attacks Still Useful for Safety-Oriented Robustness Evaluation of AI Systems?”
Abstract: Beyond adversarial machine learning in secure AI, adversarial attacks are commonly treated as worst-case tests for safety-oriented robustness evaluation of machine learning models. However, in practical AI systems, safety is governed not only by the existence of extreme failure cases, but also by their probability and system-level impact. This talk challenges the implicit equivalence between adversarial robustness and safety. Building on the well-known observation that adversarial attacks and stochastic perturbations probe distinct failure mechanisms, we show that analyzing the relationship between adversarial and perturbation robustness—through the use of probabilistic risk metrics—leads to a more nuanced interpretation of robustness evaluation. In this perspective, we highlight that adversarial attacks remain valuable for exposing local model fragilities, but they should not be treated as universal worst-case safety oracles. This view motivates a shift toward perturbation-aware AI robustness evaluations and countermeasures in cyber-physical systems, moving beyond simple adversarial attack success rates.
15:40-16:00
Paper Presentation
Title: “The Feature-Space Illusion: Exposing Practical Vulnerabilities in Blockchain GNN Fraud Detection”
Authors: Francois Frankart, Thibault Simonetto and Maxime Cordy.
Presented by: François Frankart
16:00-16:15
Break
16:15-16:35
Paper Presentation
Title: “Enabling Group Relative Policy Optimization for Vulnerability Detection”
Authors: Marco Simoni, Aleksandar Fontana, Giulio Rossolini and Andrea Saracino.
Presented by: Aleksandar Fontana
16:35-16:55
Paper Presentation
Title: “On the Robustness of Tabular Foundation Models”
Authors: Mohamed Djilani, Thibault Simonetto, Karim Tit, Florian Tambon, Salah Ghamizi, Maxime Cordy and Mike Papadakis.
Presented by: Mohamed Djilani
16:55-17:15
Paper Presentation
Title: “WiP: Robust Domain Adaptation through Teacher-Guided Adversarial Training”
Authors: Stefano Bianchettin, Giulio Rossolini and Giorgio Buttazzo.
Presented by: Stefano Bianchettin
17:15-17:30
Workshop Closing Remarks