SA-MTR: SNR-Aware Multi-Task Robust Wireless Signal Recognition

Wireless signal recognition (WSR) can simultaneously handle tasks such as automatic modulation classification (AMC) and signal type classification, offering higher efficiency and lower computational overhead compared with single-task approaches. However, such models often suffer severe performance degradation under adversarial conditions, making robustness enhancement particularly challenging. To address this issue, we propose a signal-to-noise-ratio (SNR)-aware multi-task robust signal recognition (SA-MTR) framework. First, we design an SNR estimation network that extracts SNR information from IQ signals in an end-to-end manner. This information is encoded into an SNR condition vector, which is injected into a lightweight multi-task learning (MTL) model to guide two tasks: AMC and signal type classification. Then, a collaborative adversarial training (AT) strategy is adopted to promote robustness of the MTL model. SNR estimation network is treated as a learnable conditioning branch and jointly optimized with MTL network under adversarial perturbations. Experiments on RadComAWGN show that SA-MTR outperforms existing MTL-based benchmark models in both clean-sample accuracy and adversarial robustness. Under FGSM adversarial training, it improves the average low-SNR adversarial accuracy by 7.94% to 27.53%. Additional results under multiple adversarial training strategies and on RadComDynamic further validate its robustness and generalization capability.