Cross-layer secure semantic communications with generative adversarial network-based semantic noise

This paper proposes a novel generative adversarial network (GAN)-based cross-layer secure semantic communication framework that jointly optimizes semantic representation, physical-layer security, and adversarial robustness against eavesdroppers. In contrast to conventional communication systems that focus on bit-level accuracy, we adopt a goal-oriented semantic communication paradigm that transmits high-level task-relevant features. To safeguard semantic information against unauthorized interception, we introduce a learnable semantic perturbation mechanism embedded within the transmitted features and optimize it via a GAN framework. The generator learns to generate perturbations that degrade the eavesdropper’s task performance while preserving semantic fidelity at the legitimate receiver. We further introduce a cross-layer semantic secure rate metric to jointly evaluate task security performance at the application layer and the secrecy capacity at the physical layer, which can provide a unified measure of end-to-end security performance. By jointly optimizing the parameters of semantic coding, GAN-based noise generation, and the number of semantic symbols, the proposed system can dynamically maximize the cross-layer semantic secure rate. Simulation results demonstrate the effectiveness of the proposed adversarial training scheme in terms of the cross-layer semantic secure rate while maintaining reliable semantic reconstruction for legitimate users.