Clipped BagNet: Defending Against Sticker Attacks with Clipped Bag-of-features

摘要

Many works have demonstrated that neural networks are vulnerable to adversarial examples. We examine the adversarial sticker attack, where the attacker places a sticker somewhere on an image to induce it to be misclassified. We take a first step towards defending against such attacks using clipped BagNet, which bounds the influence that any limited-size sticker can have on the final classification. We evaluate our scheme on ImageNet and show that it provides strong security against targeted PGD attacks and gradient-free attacks, and yields certified security for a 95% of images against a targeted 20 × 20 pixel attack.