Hybrid-DP Diffusion for Privacy-Preserving Wi-Fi CSI Sensing and Biometric Template Generation

Wi-Fi channel state information (CSI) enables device-free human activity recognition (HAR) and RF sensing but may expose user- and environment-specific information when shared or used for model training. Recent studies further show that CSI can leak sensitive user attributes, such as identity and physical characteristics, motivating the need for privacy-preserving sensing frameworks. We propose Hybrid-DP Diffusion, a selective differentially private framework for synthesizing privacy-preserving CSI. The framework combines nonprivate diffusion pretraining with selective DPSGD applied to identity-sensitive components, mitigating the fidelity degradation observed in prior differentially private generative models. Hybrid-DP is evaluated using both multilayer perceptron (MLP) and UNet backbones on UT-HAR and Widar 3.0 under matched privacy budgets. Results show that architectural inductive bias is critical: UNet-based diffusion achieves improved generative fidelity (lower Fréchet inception distance (FID), higher SSIM) and more stable optimization compared to MLP models under identical privacy constraints. Compared with Full-DP training, Hybrid-DP produces CSI samples with fewer artifacts while maintaining formal differential privacy (DP) guarantees and stable convergence behavior. Membership inference attacks (MIAs) remain close to random guessing, indicating limited exploitable leakage.