Potential Threat to Cognitive Radio Networks: A Black-Box and Label-Consistent Backdoor Attack Against Modulation Recognition

Deep learning-based automatic modulation recognition (AMR) technology provides critical support for dynamic spectrum management and access authorization in cognitive radio networks. However, signal acquisition in open electromagnetic space and the utilization of third-party training resources introduce potential security threats. A backdoor attack is among the most recent and severe threats, but its practicality is limited by mislabeled poisonous samples, the reliance on knowledge of recognition models, and the difficulty to adapt to dynamic channels. To reveal how adversaries can attack cognitive radio networks more stealthily and effectively within real-world constraints, we propose the first black-box and label-consistent backdoor attack (BBLC) against AMR. We formalize backdoor trigger generation as an optimization problem and derive an iterative projection method to establish constraints. Furthermore, we construct a shadow dataset and a surrogate model to solve. To accommodate channel fading and delay, we propose a novel attack strategy that simultaneously optimizes backdoor injection positions and adaptively adjusts the trigger scale. Extensive experiments conducted with various datasets and models demonstrate that BBLC can achieve an attack success rate that exceeds 90%, even with a poisoning rate as low as 0.5%. Furthermore, it maintains resilience against generic backdoor detection methods. The source code for BBLC is available at https://github.com/dzgx/BBLC-AMR.