PatchGuard: Provable Defense against Adversarial Patches Using Masks on Small Receptive Fields.

Localized adversarial patches aim to induce misclassification in machine learning models by arbitrarily modifying pixels within a restricted region of an image. Such attacks can be realized in the physical world by attaching the adversarial patch to the object to be misclassified. In this paper, we propose a general defense framework that can achieve both high clean accuracy and provable robustness against localized adversarial patches. The cornerstone of our defense framework is to use a convolutional network with small receptive fields that impose a bound on the number of features corrupted by an adversarial patch. We further present the robust masking defense that robustly detects and masks corrupted features for a secure feature aggregation. We evaluate our defense against the most powerful white-box untargeted adaptive attacker and achieve a 92.3% clean accuracy and an 85.2% provable robust accuracy on a 10-class subset of ImageNet against a 31x31 adversarial patch (2% pixels), a 57.4% clean accuracy and a 14.4% provable robust accuracy on 1000-class ImageNet against a 31x31 patch (2% pixels), and an 80.3% clean accuracy and a 61.3% provable accuracy on CIFAR-10 against a 5x5 patch (2.4% pixels). Notably, our provable defenses achieve state-of-the-art provable robust accuracy on ImageNet and CIFAR-10.