让下一篇重要论文，不再错过。

原文摘要

Encrypted semantic communication systems protect transmitted messages by inserting key-conditioned modules into a deep-learning semantic communication pipeline, but existing designs rely on a fixed adversarial weight and operate mainly at the semantic-feature level. We propose two changes to the encrypted semantic communication system (ESCS): a key-conditioned signal-domain perturbation layer, called learned adversarial perturbation injection (LAPI), and an information noise-contrastive estimation (InfoNCE)-guided adaptive training rule that {adds an adaptive leakage penalty alongside the fixed adversarial objective} using an InfoNCE-based proxy for mutual information leakage. Experiments on the Europarl corpus over additive white Gaussian noise (AWGN) and Rayleigh fading channels show that InfoNCE-guided training is the primary source of eavesdropper suppression, reducing the eavesdropper's bilingual evaluation understudy (BLEU) score to approximately 0.07 at high signal-to-noise ratio (SNR). Adding LAPI to the InfoNCE-guided objective further lowers the InfoNCE-estimated leakage by 5-9% with low measured inference overhead, while its impact on legitimate-receiver BLEU depends on channel conditions. The two mechanisms operate on different parts of the pipeline and can be jointly applied.