Opportunistic Target Selection: Early Directional Commitment for Query-Efficient Black-Box Adversarial Attacks
Black-box attacks often waste queries by wandering aimlessly through feature space (class drift). This paper introduces Opportunistic Target Selection (OTS), a lightweight method that switches an untargeted attack to targeting the leading non-true class early on. This redirects the perturbation toward the correct decision boundary, significantly improving query efficiency without needing complex modifications.