AutoDA: Automated Decision-based Iterative Adversarial Attacks

Adversarial attacks can fool deep learning models by imposing imperceptible perturbations onto natural examples, which have provoked concerns in various security-sensitive applications. Among them, decision-based black-box attacks are practical yet more challenging, where the adversary can only acquire the final classification labels by querying the target model without access to the model's details. Under this setting, existing works usually rely on heuristics and exhibit unsatisfactory performance in terms of query efficiency and attack success rate. To better understand the rationality of these heuristics and further improve over existing methods, we propose AutoDA to automatically discover decision-based iterative adversarial attack algorithms. In our approach, we construct a generic search space of attack algorithms and develop an efficient search algorithm to explore this space. Although we adopt a small and fast model to efficiently evaluate and discover qualified attack algorithms during the search, extensive experiments demonstrate that the discovered algorithms are simple yet query-efficient when attacking larger models on the CIFAR-10 and ImageNet datasets. They achieve comparable performance with the human-designed state-of-the-art decision-based iterative attack methods consistently.