WalkAnon: Defeating Gait-Based Wi-Fi Identity Inference via Differential Privacy on Givens Rotation Angles

Recent work demonstrates that IEEE 802.11ac/ax Beamforming Feedback Information (BFI) frames — unencrypted, mandatory, and passively capturable — expose a biometric gait signature sufficient to identify individuals with 99.5% accuracy across 197 subjects using a BiLSTM classifier. We present WalkAnon, a firmware-level defense that applies ε-differential privacy directly to the Givens rotation angles (φ, ψ) encoded in BFI frames before transmission. WalkAnon introduces two mechanisms: (1) angle-domain DP noise — Wrapped Laplace for the circular φ domain and Truncated Laplace for the bounded ψ domain — and (2) per-frame temporal phase scrambling that destroys the stride-rate autocorrelation exploited by sequence models. The output is a standard-compliant unit-norm beamforming vector. Attack accuracy drops from 100% to 5% (= 1/N random baseline) at ε = 0.1. To our knowledge, temporal decorrelation of BFI sequences has not been proposed previously.