GenFi: Enhancing WiFi-Based Human Activity Recognition to Unseen Scenario Via Feature Disentanglement and Meta-Learning

WiFi-based sensing technology has gained significant attention for its ability to enable pervasive human activity recognition (HAR) in indoor spaces. One challenge is that current WiFi HAR systems often experience performance degradation in unseen scenarios (e.g., new environments, people, and weather conditions). Some research has attempted to address this issue by extracting scenario-invariant features using deep learning (DL). However, discarding scenario-specific features brings about insufficient representation of task-related information, resulting in limited model adaptability. In this paper, we present GenFi, a robust WiFi HAR system that enhances model generalization by leveraging both scenario-invariant and scenario-specific features. To achieve this, GenFi first disentangles the raw input into these two types of features through adversarial learning and correlation analysis. Subsequently, GenFi uses meta-learning to self-optimize the fusion of these two features, leading to a generalized cross-scenario WiFi HAR system. Compared to state-of-the-art approaches, GenFi achieves the best trade-off between high performance and low complexity in diverse unseen scenarios, making it a promising solution for real-world deployment.