PubDef: Defending Against Transfer Attacks Using Public Models

Adversarial attacks pose a serious threat to the reliability and security of machine learning systems. By making small perturbations to inputs, attackers can cause models to produce completely incorrect outputs. Defending against these attacks is an active area of research, but most proposed defenses have major drawbacks.

This paper (repo

adversarial adversarial attacks attacks learning systems machine machine learning major making paper plain english papers public reliability research security small systems threat transfer