War: An Efficient Pre-processing Method for Defending Adversarial Attacks

Deep neural networks (DNNs) have achieved extraordinary successes in many fields such as image classification. However, they are vulnerable to adversarial examples generated by adding slight perturbations to the input images, leading incorrect classification results. Due to the serious threats of adversarial examples, it is necessary to find a simple and practical way to defend against adversarial attacks. In this paper, we present an efficient preprocessing method called War (WebP compression and resizing operation) for defending adversarial examples. WebP compression is first performed on the input sample to remove the imperceptible perturbations from the adversarial example. Then, the compressed image is appropriately resized to further destroy the specific structure of the adversarial perturbations. Finally, we can get a clean sample that can be correctly classified by the model. Extensive experiments show that our method outperforms the state-of-the-art defense methods. It can effectively defend adversarial attacks while ensure the classification accuracy on the normal samples drops slightly. Moreover, it only requires a particularly short pre-processing time.