Adversarial Samples Detection Based on Feature Attribution and Contrast in Modulation Recognition

Detecting adversarial samples is crucial for maintaining the security of automatic modulation recognition (AMR) systems, as adversarial attacks could severely compromise wireless communication. To address this threat, we propose a novel adversarial samples detection method named Null Feature Attribution Abnormality (NFAA), which leverages the target model’s interpretation of feature importance to distinguish between benign and adversarial signal samples. Furthermore, we propose the NFAA-TC method, incorporating a Triple Contrast (TC) approach to mitigate noise in signal data and enhance the performance of adversarial samples detection. Experimental results validate the effectiveness of the proposed method across various adversarial attacks and under different signal-to-noise ratio (SNR) conditions.