Answers are grounded in this paper and cite their sources.
Black-box adversarial attacks pose a practical concern for machine learning models, yet many existing methods rely on confidence scores and excessive queries, limiting their use in restricted-access settings. Most prior work focuses on deep neural networks, with less attention to conventional, non-differentiable models widely used in practice. This paper presents a unified hard-label black-box attack framework that targets both neural and non-differentiable models under strict access constraints, requiring only top-1 label outputs. The approach combines lightweight surrogate guidance with iterative perturbation refinement. Specifically, auxiliary one-vs-rest logistic regression models estimate perturbation directions, which are refined using an FGSM-like update and binary search to minimize perturbation magnitude. We evaluate the method on 13 models, including MLPs, a CNN, and traditional classifiers such as Random Forest, XGBoost, SVM, and kNN, using MNIST as a controlled benchmark. The attack achieves high success rates (e.g., 63–85% on single-layer MLPs) with low query counts (typically under 28), while exposing vulnerabilities across model types. Although results demonstrate the framework’s effectiveness in a unified setting, validation on more complex datasets remains future work.
No takes yet. Share an insight, caveat, or question.
Sharif et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: