Titolo: Analysis of the vulnerability of AIbased classifiers against adversarial attacks

Scadenza: 28/01/2026

Struttura: Segreteria amministrativa Dipartimento di Ingegneria dell'Informazione e Scienze Matematiche

Descrizione: The goal of the research is to analyze the data distribution of widely used image datasets in AI - like MNIST, CIFAR-10, Food101 and possibly others - to understand structural and statistical properties that may influence model robustness. The focus will be on assessing how data geometry and highdimensional structure contribute to the emergence of adversarial examples. In a second phase the findings of the analysis will be evaluated by the light of existing theoretical work on the concentration of measure phenomenon, which suggests why, in high dimensions, small perturbations can significantly alter model predictions. Understanding how these theoretical insights manifest in real-world datasets can help identify intrinsic vulnerabilities in current AI models and guide the design of more robust learning systems. Eventually, the validity of the results of the theoretical analysis will be assessed experimentally, on a pool of classifiers trained on the datasets used for the analysis.

Durata: 12 mesi 0 giorni

Allegati:

• Avviso n. 1
  
• Scheda Attività Avviso n. 1
  
• richiesta partecipazione selezione