Hindering Adversarial Attacks with Multiple Encrypted Patch Embeddings
This work addresses security for machine learning systems by improving defenses against adversarial attacks, though it is incremental as it builds upon existing key-based methods.
The paper tackles the vulnerability of previous key-based defenses to adaptive attacks and training difficulties on large datasets like ImageNet by proposing a new defense with efficient training and optional randomization using multiple encrypted patch embeddings. The results show it achieves high robust accuracy and comparable clean accuracy on ImageNet against state-of-the-art attacks.
In this paper, we propose a new key-based defense focusing on both efficiency and robustness. Although the previous key-based defense seems effective in defending against adversarial examples, carefully designed adaptive attacks can bypass the previous defense, and it is difficult to train the previous defense on large datasets like ImageNet. We build upon the previous defense with two major improvements: (1) efficient training and (2) optional randomization. The proposed defense utilizes one or more secret patch embeddings and classifier heads with a pre-trained isotropic network. When more than one secret embeddings are used, the proposed defense enables randomization on inference. Experiments were carried out on the ImageNet dataset, and the proposed defense was evaluated against an arsenal of state-of-the-art attacks, including adaptive ones. The results show that the proposed defense achieves a high robust accuracy and a comparable clean accuracy compared to the previous key-based defense.