MIMoAE: A Mutual-Information-Driven Mixture-of-Experts Autoencoder for Cross-Scenario Physical-Layer Secret Key Generation

Abstract

Physical layer secret key generation (PLKG) exploits wireless channel reciprocity to enable two legitimate parties to establish shared secret keys without relying on trusted third parties. However, most existing deep-learning-based PLKG schemes are designed or trained for specific channel environments, and their reciprocity enhancement capability can degrade significantly when deployed in unseen environments with different channel statistics, leading to severe performance loss under out-of-distribution (OOD) conditions. To address this issue, this paper proposes a PLKG scheme based on a mutual-information-driven mixture-of-experts autoencoder (MIMoAE). First, we design a mixture-of-experts encoder with specialized sub-networks and dynamic routing, enabling flexible adaptation to diverse channel environments. Then, a routing consistency loss is introduced to explicitly align the routing behavior between the two legitimate parties under distribution shift, preventing routing inconsistency. Finally, a nonlinear projection head is incorporated to stabilize mutual information estimation and provide more reliable gradient guidance. Compared with the state-of-the-art baseline, MIMoAE reduces the average key error rate by approximately 67% and achieves a raw key generation rate of about 3.85 bits/channel use on OOD test scenarios.