Attention-Enhanced CNN Framework for WiFi CSI-based Cross-Domain Gait Recognition

WiFi channel state information (CSI) enables contact-free gait-based authentication but suffers from significant performance degradation across environments. This paper proposes an attention-enhanced convolutional neural network (CNN) framework for cross-domain gait recognition using commodity WiFi. A dedicated preprocessing pipeline with bandpass filtering, wavelet denoising, conjugate multiplication, and reconstruction independent component analysis is employed to obtain separated gait signals. We then construct a subcarrier-frequency map that exploits the multi-subcarrier structure of CSI while reducing sensitivity to walking speed and starting phase. These maps are fed into a CNN-Transformer backbone with an asymmetric tri-training strategy to learn discriminative and domain-robust features from labeled source data and unlabeled target data. Experiments show that the proposed method achieves over 90% accuracy in the corridor scenario and consistently outperforms several deep learning baselines in more cluttered environments.