Physical Layer Security for Digital Semantic Communication: A Bit-Level Perspective
Abstract
Digital semantic communication requires reliable recovery of task-relevant information at a legitimate receiver while limiting source reconstruction by an eavesdropper. Uniform bit protection, however, does not account for the heterogeneous reliability and secrecy requirements of semantic bits. This paper proposes a bit-level physical layer security (PLS) framework that learns these requirements and incorporates them into physical-layer resource allocation. We first introduce a wiretap semantic feature channel whose two-stage training learns per-bit target bit-flip probabilities for the legitimate receiver (Bob) and the eavesdropper (Eve). These targets characterize the reliability requirements at Bob and the secrecy protection requirements against Eve. Based on the learned targets, we formulate a joint power and artificial-noise (AN) allocation problem for a multiple-input single-output wiretap channel under a worst-case angular-separation model. For a given target profile, the optimal information-AN power split admits a closed-form solution, and the per-bit power allocation is obtained by water-filling. We further develop a leakage-aware selection method among multiple learned-target profiles and extend the design to multiple-input multiple-output channels. Simulation results demonstrate that the proposed framework achieves a better trade-off between preserving reconstruction quality at Bob and degrading that at Eve than uniform power allocation and the evaluated learning-based secure semantic communication schemes.