SFDT: A Signal Frequency-Domain Backdoor Attack With Sample-Specific Trigger Against Automatic Modulation Recognition
Gan Xu, Hongjun Wang, Xinhao Li, Guan Gui · IEEE Wireless Communications Letters · 2025
Deep learning (DL)-based automatic modulation recognition (AMR) plays a crucial role in contemporary wireless communication systems. However, collecting data in open electromagnetic space and relying on third-party training resources can introduce significant security vulnerabilities, notably through backdoor attacks. Existing backdoor attacks on AMR function primarily in the time domain, constrained by identical trigger designs, conspicuous embedding artifacts, and static activation positions. To reveal more stealthy and severe backdoor threats, we propose the first Signal Frequency-Domain backdoor attack with sample-specific Trigger (SFDT) against AMR. The SFDT integrates a controller and synthesizer to dynamically generate a trigger specifically adapted to individual signal sample, allowing stealthy embedding in the frequency domain. This work highlights previously unexplored risks and underscores the critical need for security measures in AMR systems.