DIBAD: A Disentangled Information Bottleneck Adversarial Defense Method Using Hilbert-Schmidt Independence Criterion for Spectrum Security

Automatic Modulation Classification (AMC) is crucial for monitoring the legitimacy of user frequency behavior and identifying potential sources of interference in spectrum monitoring. Deep learning-based AMC models have shown excellent performance, however, it has been proven susceptible to adversarial attacks. To address the problem, we propose a Disentangled Hilbert-Schmidt Information Bottleneck Adversarial Defense (DIBAD) method to enhance the adversarial robustness of AMC models. Specifically, we firstly analyze the task-relevant and task-irrelevant features in the intermediate representations of modulation signals from the perspective of mutual information theory. Secondly, a training framework consisting of a classification feature extractor, a supplementary feature extractor, and a classifier is designed. Under the information bottleneck constraint, the classification feature extractor and supplementary feature extractor are used to extract task-relevant and task-irrelevant features, respectively. The information bottleneck constraint is employed to reduce task-irrelevant features, thus improving the model’s adversarial robustness. Experiments on the RML2016.10a and DMRadio09.real datasets, along with comprehensive analysis, demonstrate the superiority of the DIBAD method terms of adversarial robustness.