Meta Learning Enabled Adversarial Defense
Yiyi Tao
Abstract
Machine learning (ML) models have been found to be vulnerable against adversarial attacks recently. Such examples are carefully crafted instances with a small magnitude of perturbation added to mislead the target ML models. This has raised great concerns and different approaches have been put forward to defend against these malicious adversarial attacks, among which adversarial training so far has been shown to be very effective empirically. However, one major limitation of adversarial training is that it requires exploring diverse factors such as different attacks and datasets to achieve good performance, which significantly limits its practical usage. We propose a novel and effective meta-learning based adversarial training framework named Meta-Adv in this work. We leverage the knowledge obtained from past tasks that are trained on a subset of attacks and datasets and show that such knowledge is sufficient to improve the learning robustness against different unforeseen adversarial attacks. We conducted extensive experiments on different datasets including CIFAR-10 and ImageNet and show that the proposed method outperforms the state-of-the-art adversarial training methods significantly, including the winners in the CAAD adversarial defense competition. Our method provides a simple and effective baseline to train robust ML models against adversarial attacks.