Opportunistic Target Selection: Early Directional Commitment for Query-Efficient Black-Box Adversarial Attacks
Accepted at CAp 2026 (Conférence sur l'Apprentissage automatique), poster presentation
Black-box adversarial attacks that minimize only the ground-truth confidence can suffer from class drift: perturbations make diffuse progress without committing to a specific adversarial class. Opportunistic Target Selection (OTS) first lets an attack explore in untargeted mode, then switches to a targeted objective against the leading non-true class.
The method requires no architectural modification, gradient access, or prior target-class knowledge. Across three score-based attacks and five ImageNet classifiers, OTS improves drift-prone random-search attacks by up to 27 percentage points in success rate and reduces censored-mean iterations by 43% on ResNet-50.