How Vulnerable Is My Learned Policy? Universal Adversarial Perturbation Attacks On Modern Behavior Cloning Policies
The study tested five behavior cloning algorithms against universal adversarial perturbations (UAP) across white-box, grey-box, and black-box attacks. Findings show most methods are highly vulnerable to these attacks, including cross-algorithm transfer. Implicit methods like IBC showed the highest resilience, while classical explicit methods suffered significant performance drops even with minimal noise.