Does Semantic Similarity Provide Enhanced Physical Layer Security?

This paper investigates the security of semantic communication systems over Rayleigh fading wiretap channels. Unlike traditional systems, semantic communication focuses on recovering the intended meaning rather than exact bits. To cope with this inconsistency, we propose, for the first time, using the semantic similarity metric to enhance physical layer security within a realistic framework in which noise affects the transmitted messages at the bit level similarly to traditional systems. Semantic sets are defined via semantic (cosine) similarity thresholds and decoding performance is evaluated accordingly. To address complexity challenges, we segment the transmitted messages into chunks of decreasing semantic importance. We compare two strategies for the considered semantic similarity thresholds: (a) keeping them constant across all chunks and (b) linearly scaling them to adjust their value according to the importance. Results reveal a trade-off among semantic reliability, security, and complexity. Further, our results demonstrate that by adjusting the semantic similarity threshold, the size of the semantic set is controlled, which in turn affects both decoding success and vulnerability to leakage. Finally, we note that higher thresholds improve security, but reduce reliability, while lower thresholds enhance robustness at the cost of increased risk of semantic leakage.