Collaborative VQMAE: Defense Against Adversarial Attacks in Automatic Modulation Recognition

Deep neural networks (DNNs) offer advanced intelligent solutions for communications' automatic modulation recognition (AMR) tasks. However, AMR models are vulnerable to adversarial attacks, wherein carefully crafted perturbations are injected into input signals, causing the models to make incorrect predictions. Existing defense methods face challenges balancing recognition accuracy while effectively countering varied attacks. This paper proposes a dual-stage defense framework based on collaborative vector quantized masked autoencoder (VQMAE) to address this. In the first stage, collaborative VQMAE achieves feature compression and reconstruction through vector quantization. To further refine the adversarial signal by aligning features and minimizing the discrepancy between adversarial and original signals, feature alignment is employed to quantify their differences. In the second stage, a bidirectional Transformer is utilized to learn the prior distribution of tokens indices, thereby enhancing signal reconstruction accuracy. Blackbox and white-box attack-defense experiments conducted on the publicly available dataset RML2016.10a demonstrate that the proposed method significantly strengthens the model's resilience to adversarial perturbations while maintaining the classifier's original performance.