Adversarial Robust Modulation Recognition Guided by Attention Mechanisms

Abstract

Deep neural networks have demonstrated considerable effectiveness in recognizing complex communications signals through their applications in the tasks of automatic modulation recognition. However, the resilience of these networks is undermined by the introduction of carefully designed adversarial examples that compromise the reliability of the decision processes. In order to address this issue, an Attention-Guided Automatic Modulation Recognition (AG-AMR) method is proposed in this paper. The method introduces an optimized attention mechanism within the Transformer framework, where signal features are extracted and filtered based on the weights of the attention module during the training process, which makes the model to focus on key features for the task. Furthermore, by removing features of low importance where adversarial perturbations may appear, the proposed method mitigates the negative impacts of adversarial perturbations on modulation classification, thereby it improves both accuracy and robustness. Experimental results on benchmark datasets show that AG-AMR obtains a high level of accuracy on modulation recognition and exhibits significant robustness. Furthermore, when working together with adversarial training, it is shown that AG-AMR effectively resists several existing adversarial attacks, which thus further validates its effectiveness on defending against adversarial sample attacks.