Activity: Talk or presentation types › Invited talk
Description
Abstract The increasing integration of AI systems into critical domains necessitates a focus on their trustworthiness, particularly in terms of security and privacy. This talk explores 2 pivotal aspects of Trustworthy AI: AI security and privacy. The security aspect will be investigated from both: i) Adversarial attacks with a focus on leveraging information theory to analyze and counteract these threats, and ii) Model hijacking attacks, introducing "Snatch-ML," a novel approach to repurpose and hijack ML models without training access. On the privacy front, the discussion will discuss the concept of Adaptive Differential Privacy, emphasizing its potential in safeguarding sensitive data while maintaining utility.