English

Sparse-RS: a versatile framework for query-efficient sparse black-box adversarial attacks

Machine Learning 2022-02-09 v3 Cryptography and Security Computer Vision and Pattern Recognition Machine Learning

Abstract

We propose a versatile framework based on random search, Sparse-RS, for score-based sparse targeted and untargeted attacks in the black-box setting. Sparse-RS does not rely on substitute models and achieves state-of-the-art success rate and query efficiency for multiple sparse attack models: l0l_0-bounded perturbations, adversarial patches, and adversarial frames. The l0l_0-version of untargeted Sparse-RS outperforms all black-box and even all white-box attacks for different models on MNIST, CIFAR-10, and ImageNet. Moreover, our untargeted Sparse-RS achieves very high success rates even for the challenging settings of 20×2020\times20 adversarial patches and 22-pixel wide adversarial frames for 224×224224\times224 images. Finally, we show that Sparse-RS can be applied to generate targeted universal adversarial patches where it significantly outperforms the existing approaches. The code of our framework is available at https://github.com/fra31/sparse-rs.

Keywords

Cite

@article{arxiv.2006.12834,
  title  = {Sparse-RS: a versatile framework for query-efficient sparse black-box adversarial attacks},
  author = {Francesco Croce and Maksym Andriushchenko and Naman D. Singh and Nicolas Flammarion and Matthias Hein},
  journal= {arXiv preprint arXiv:2006.12834},
  year   = {2022}
}

Comments

Accepted at AAAI 2022. This version contains considerably extended results in the L0 threat model

R2 v1 2026-06-23T16:32:53.432Z