An Information Theoretic Defense Algorithm against Adversarial Attacks on Deep Learning
Xinjie Lan, Samet Bayram, Kenneth E. Barner · 2021
Improving the adversarial robustness of Deep Neural Networks (DNNs) is an important topic in deep learning, and adversarial training is one of the most effective defense algorithm against adversarial attacks. Inspired by the mutual information regularization for solving the worst-case problem, this paper introduces an information theoretic defense algorithm for boosting the adversarial robustness of DNNs via incorporating a novel mutual information regularization into adversarial training. Simulations demonstrate that the proposed information theoretic defense algorithm outperforms the classical adversarial training based on multiple adversarial attacks for different network architectures on benchmark datasets.