Enhancing adversarial robustness in image classification models through frequency mutual information

This paper explores the dependencies between inputs and outputs in deep neural networks under adversarial conditions from the perspective of information theory, particularly focusing on the dependencies on low-frequency and high-frequency patterns. We propose and validate a novel adversarial defense strategy that enhances the robustness of deep learning models by separating and optimizing the mutual information of low and high frequencies. Experiments demonstrate that this method effectively improves the model's defensive capabilities against complex adversarial attacks.