Cause of reduced recoverability after perturbation removal

Determine whether the loss of recoverability observed after a perturbation ends is caused by adversarial optimization rather than by perturbation duration and magnitude more generally, by comparing adversarial trajectories with matched non-adversarial controls.

Background

The paper studies whether adversarial visual patches have effects that persist after the patch is removed from a Vision-Language-Action policy’s observations. Prior work cited by the paper evaluates recovery after physical faults or half-sequence attacks, showing that longer perturbations can reduce subsequent task performance.

According to the paper, those studies do not distinguish whether reduced recoverability is specifically attributable to adversarial optimization or instead follows from generic factors such as perturbation duration and action-error magnitude. The paper addresses this distinction experimentally using matched controls for occlusion, action-error magnitude, and directional persistence, so the question is presented as an unresolved issue in prior work rather than as an open problem left unresolved by the paper’s own experiments.