Search papers, labs, and topics across Lattice.
This paper introduces a personalized federated learning (PFL) framework for human activity recognition (HAR) on resource-constrained devices, addressing data privacy, non-IID data, and limited computational resources. The framework combines FedAvg training of a CNN with magnitude-based pruning, float16 quantization, and per-client fine-tuning. Results demonstrate that personalization mitigates performance degradation caused by non-IID data and pruning, while quantization reduces model size by 50%, enabling deployment on wearable devices.
Personalizing federated learning models after compression recovers accuracy lost from pruning, enabling efficient on-device human activity recognition.
Human Activity Recognition (HAR) using wearable sensors enables impactful applications in healthcare, fitness, and smart environments, but it also faces challenges related to data privacy, non-independent and identically distributed (non-IID) data, and limited computational resources on edge devices. This study proposes an efficient and personalized federated learning (PFL) framework for HAR that integrates federated training with model compression and per-client fine-tuning to address these challenges and support deployment on resource-constrained devices (RCDs). A convolutional neural network (CNN) is trained across multiple clients using FedAvg, followed by magnitude-based pruning and float16 quantization to reduce model size. While personalization and compression have previously been studied independently, their combined application for HAR remains underexplored in federated settings. Experimental results show that the global FedAvg model experiences performance degradation under non-IID conditions, which is further amplified after pruning, whereas per-client personalization substantially improves performance by adapting the model to individual user patterns. To ensure realistic evaluation, experiments are conducted using both random and temporal data splits, with the latter mitigating temporal leakage in time-series data. Personalization consistently improves performance under both settings, while quantization reduces the model footprint by approximately 50%, enabling deployment on wearable and IoT devices. Statistical analysis using paired significance tests confirms the robustness of the observed performance gains. Overall, this work demonstrates that combining lightweight model compression with personalization providing an effective and practical solution for federated HAR, balancing accuracy, efficiency, and deployment feasibility in real-world scenarios.
