With the rapid development of deep learning in wireless communication for signal detection, target recognition, and parameter estimation, the vulnerability of models to adversarial examples poses a critical challenge to system robustness and security. Existing adversarial attacks mainly focus on single-task models, limiting their applicability in multi-task scenarios. To address this issue, we propose a Multi-Teacher Distillation-guided Multi-Task Attack (MTDMA) framework. It integrates three high-performance teacher models—direct sequence spread spectrum (DSSS) detection, modulation recognition, and direction-of-arrival (DOA) estimation—to train a unified student model through hybrid knowledge distillation. By fusing soft targets from multiple teachers, the student jointly learns discriminative features of multiple tasks within a shared representation space. Furthermore, we design a Dual-Representation Momentum Attack (DRMA) that generates perturbations in both amplitude and IQ feature spaces with a momentum mechanism, improving attack transferability and stability. Experimental results demonstrate that MTDMA achieves high attack success rates on three task models and transfers effectively to heterogeneous architectures, outperforming FGSM, BIM, PGD, and MIM. This work extends adversarial attacks from single-task to multi-task settings, providing new insights into robustness evaluation and multi-task adversarial sample generation for wireless communication models.

Abstract Automatic modulation classification models based on deep learning models are at risk of being interfered by adversarial attacks. In an adversarial attack, the attacker causes the classification model to misclassify the received signal by adding carefully crafted adversarial interference to the transmitted signal. Based on the requirements of efficient computing and edge deployment, a lightweight automatic modulation classification model is proposed. Considering that the lightweight automatic modulation classification model is more susceptible to interference from adversarial attacks and that adversarial training of the lightweight auto‐modulation classification model fails to achieve the desired results, an adversarial attack defense system for the lightweight automatic modulation classification model is further proposed, which can enhance the robustness when subjected to adversarial attacks. The defense method aims to transfer the adversarial robustness from a trained large automatic modulation classification model to a lightweight model through the technique of adversarial robust distillation. The proposed method exhibits better adversarial robustness than current defense techniques in feature fusion based automatic modulation classification models in white box attack scenarios.