Watermarking-based Defense against Adversarial Attacks on Deep Neural Networks

The vulnerability of deep neural networks to adversarial attacks has posed significant threats to real-world applications, especially security-critical ones. Given a well-trained model, slight modifications to the input samples can cause drastic changes in the predictions of the model. Many methods have been proposed to mitigate the issue. However, the majority of these defenses have proven to fail to resist all the adversarial attacks. This is mainly because the knowledge advantage of the attacker can help to either easily customize the information of the target model or create a surrogate model as a substitute to successfully construct the corresponding adversarial examples. In this paper, we propose a new defense mechanism that creates a knowledge gap between attackers and defenders by imposing a designed watermarking system into standard deep neural networks. The embedded watermark is data-independent and non-reproducible to an attacker, which improves randomization and security of the defense model without compromising performance on clean data, and thus yields knowledge disadvantage to prevent an attacker from crafting effective adversarial examples targeting the defensive model. We evaluate the performance of our watermarking defense using a wide range of watermarking algorithms against four state-of-the-art attacks on different datasets, and the experimental results validate its effectiveness