Taking the Advantages of User Diversity to Enhance WiFi-based Gesture Recognition

WiFi-based gesture recognition enables intuitive human-computer interaction but struggles with user diversity, where variations in body shape and behavior reduce generalization. Existing domain-invariant methods often discard user-dependent features, overlooking their potential for personalization. Domain-adaptation approaches rely on additional user data but lack mechanisms to mitigate the negative influence of user-dependent features. This paper introduces UserFi, a framework that disentangles user-independent and user-dependent features, reintegrates them for prediction, and applies a joint fine-tuning strategy in a federated setting to balance generality and specificity. Experiments show that UserFi effectively transforms user diversity from a limitation into a benefit for WiFi-based gesture recognition. It achieves 84.6% accuracy on unseen users, 8.3% higher than the representative baselines.