Enhancing the Robustness of Deep Neural Networks by Meta-Adversarial Training

You-Kang Chang ； Hong Zhao ； Wei-Jie Wang

《International Journal of Network Security》 25卷1期 (2023/01) Pp. 122-130

https://doi.org/10.6633/IJNS.202301_25(1).14

摘要

Adversarial training can effectively defend against the impact of adversarial attacks on deep neural networks but suffers from poor generalization ability and low defense efficiency. To address this problem, this paper proposes a method combining meta-learning with adversarial training to enhance the robustness of deep neural networks. Firstly, a training dataset containing adversarial examples and clean examples is constructed, and conduct adversarial training on the deep neural network. Secondly, the features extracted from the adversarial training are learned using the meta-learning method, and the problem of the need to continuously input a large number of adversarial examples for training in adversarial training is solved by using the feature that meta-learning has strong adaptability in the face of new tasks. Experimental results show that this method can improve the robustness of deep neural networks and effectively resist standard classes of adversarial attacks.

關鍵字

Adversarial Training ； Adversarial Attack Defense ； Meta-learning ； Neural Networks ； Robustness Studies