STROKE: Simple Sparse Structured Occlusions for Black-Box Adversarial Attacks

The researchers propose STROKE, a query-efficient, score-based black-box adversarial attack method that constructs perturbations using simple targeted rectangular occlusions and optimizes them through a lightweight greedy "add—refine" procedure. On a variety of image classifiers, STROKE achieves high attack success rates with fewer queries under a nominal 1-2% area setting, and in several configurations performs better relative to existing patch-based attacks. Comparative experiments also show that its attack time is substantially shortened.