Toward Robust and Deployable Representation Learning for AI-Native Wireless Sensing

Abstract

Device-free Wi-Fi sensing offers a privacy-preserving approach to human activity recognition (HAR), yet practical deployment faces two key challenges: domain shifts across sensing environments and the gap between CSI compression and prediction in dynamic wireless systems. Existing Unsupervised Domain Adaptation (UDA) methods rely on labeled source data and fail in multi-user scenarios due to signal entanglement, permutation invariance, and severe class imbalance. AI-driven CSI feedback methods treat compression and prediction as independent problems, leaving channel aging insufficiently addressed. This thesis proposes two complementary frameworks. First, MU-SHOT-Fi is a Source-Free UDA framework for multi-user Wi-Fi sensing employing a permutation-invariant set-prediction architecture trained via Hungarian matching. It introduces an occupancy-weighted information maximization objective to handle class imbalance and integrates rotation-based spatial self-supervision to align cross-domain representations using the frequency-time structure of Channel State Information (CSI). A single-user extension, SU-SHOT-Fi, incorporates temporal self-supervision via Contrastive Predictive Coding (CPC). Second, a unified compression-prediction framework integrates CPC into a 3GPP-compliant CSI pipeline, jointly optimizing reconstruction fidelity and temporal predictive coherence to address channel aging without increasing feedback overhead. Two variants, CPC-before-Compression and CPC-after-Compression, offer different trade-offs between temporal modeling and device complexity. Evaluations on WiMANS and Widar 3.0 demonstrate consistent improvements over state-of-the-art baselines under cross-room and cross-frequency shifts. Experiments on Nokia, Oppo, and CATT datasets confirm favorable complexity-performance trade-offs, advancing AI-native wireless sensing toward robust, practical deployment.