Secure Semantic Communication With Non-Targeted and Target-Oriented Stealthy Semantic Protection

Semantic communication (SC), as an emerging communication paradigm, has garnered increasing attention due to its unique advantages in robustness and bandwidth savings in extreme environments, but the open nature of wireless channels and the reliance on deep learning expose SC systems to serious security threats. Existing defenses primarily focus on degrading eavesdropper reconstruction but neglect stealthiness. In this paper, we propose a lightweight stealth-aware secure SC (LS-SSC) framework for resource-constrained image transmission under eavesdropping threats. Built upon a channel-spatial attention-enhanced reversible residual network (RevNet), LS-SSC utilizes the forward process to inject stealthy non-targeted adversarial attacks (NTAA) and controllable target-oriented adversarial attacks (TOAA) into the transmitted semantic features, effectively misleading eavesdroppers’ downstream AI tasks. Conversely, the legitimate receiver leverages the invertibility of RevNet to eliminate these perturbations and recover the original semantic information. Furthermore, to balance the conflicting objectives of adversarial intensity and visual stealthiness, we propose a piecewise approximation dynamic-weighting strategy guided by validation indicators. We conduct extensive simulations for LS-SSC under NTAA and TOAA. Experimental results demonstrate that LS-SSC reduces parameters by 71% and FLOPs by 41% compared to baselines, while achieving superior attack efficacy, robust stealthiness, and reliable legitimate reconstruction.