Model-Reuse Attacks on Deep Learning Systems

Explore a 24-minute conference talk that delves into the security risks posed by malicious primitive models in machine learning systems. Examine a broad class of model-reuse attacks where crafted models manipulate host ML systems to misbehave on specific inputs. Investigate the attack setting, objectives, and techniques for generating semantic neighbors and finding salient features. Learn about case studies in multi-view autonomous steering, analyzing attack effectiveness and evasiveness. Gain insights into the potential threats of pre-trained models as building blocks in complex ML systems and understand the implications for security in our increasingly AI-driven world.