Towards Eavesdropper Awareness in CSI-Based Key Generation

Abstract

Physical-layer key generation derives shared bit strings from reciprocal wireless measurements, but conventional feature selection often optimizes Alice-Bob agreement without explicitly assessing whether the selected features are also observable by a passive eavesdropper. This paper presents a compact channel-state-information (CSI) key-generation pipeline that selects discrete-cosine-transform (DCT) coordinates using an eavesdropper-aware score and compares it with a capacity-matched eavesdropper-unaware control. Both variants share the same 96-dimensional CSI feature contract, 36-coordinate budget, guarded 2-bit Gray quantization, reconciliation, confirmation, Toeplitz hashing, grouped split, and development-derived extraction accounting. The study uses simultaneous recordings from two passive eavesdroppers over static, walking, endpoint-motion, line-of-sight (LOS), non-line-of-sight (NLOS), and placement conditions. Across 50 held-out capture sets, the eavesdropper-aware selector yields 26,971 final bits at 35.961 bit/s and a maximum post-generation Eve correlation of 0.003956; the matched control yields 27,873 bits at 37.164 bit/s and 0.005341. The results show a measured rate-exposure trade-off: Eve-aware selection lowers the evaluated exposure indicator while retaining the same number of confirmed sessions.