Adversarial Defense in Modulation Recognition via Diffusion and Segment-Wise Classification

With the development of wireless communication devices, due to limited spectrum resources and rising interference, the importance of effective spectrum sensing techniques has drawn much attention. Hereby, modulation recognition serves as a cornerstone for non-cooperative communications and anti-jamming operations. While deep learning becomes popular through autonomous feature extraction, its vulnerability to adversarial attacks poses critical security risks. To address this challenge, FlowSlicer is proposed based on diffusion models for the modulation recognition domain. Furthermore, we explore a segmented recognition strategy for communication signals and propose an aggregation algorithm to enhance the modulation recognition. Simulation results validate the robustness of FlowSlicer under various adversarial attack strategies.