Federated weakly-supervised representation learning for privacy-preserving human activity recognition using wearable sensors

Human activity recognition (HAR) plays a crucial role in wearable sensor-based applications, particularly in privacy-sensitive domains such as healthcare, fitness, and smart homes. Federated learning (FL), when combined with weakly-supervised representation learning, provides an effective solution to enhance HAR while ensuring privacy. This paper proposes a novel approach for privacy-preserving human activity recognition using wearable sensors, which integrates federated learning with weakly-supervised representation learning. The approach leverages contrastive learning techniques to enable feature extraction from weakly-labeled or unlabeled sensor data and employs multimodal sensor fusion to improve recognition accuracy. Experimental results demonstrate that the proposed framework outperforms traditional supervised learning models in terms of both accuracy and privacy preservation, making it suitable for scalable HAR applications. Our findings highlight the potential of this hybrid framework in advancing privacy-conscious systems for healthcare monitoring, fitness tracking, and smart home environments.