Accuracy Optimization of Lightweight Home Health Models via Fog Computing, Federated Learning and Source-Free Domain Adaptation

Aiming at the dual constraints of data silos and privacy protection in home health monitoring scenarios, a lightweight model accuracy optimization method integrating fog computing architecture, federated learning mechanism and source-free domain adaptation technology is proposed. The adaptive aggregation strategy operates on encrypted gradient parameters rather than raw feature statistics, and the distribution matching is performed on aggregated prototype vectors at the cloud server rather than individual client data, preventing lifestyle pattern leakage. An adaptive aggregation strategy is deployed on edge fog nodes to enable cross-domain knowledge transfer via feature alignment and distribution matching, which improves the generalization ability of target domain models without accessing source domain data. Experimental results showed that this method compressed the model parameters to 18.7% of the baseline model on multiple home health datasets, improved the cross-domain recognition accuracy by 12.4 percentage points, and reduced communication overhead by 63.2%, providing a feasible technical path for distributed intelligent health monitoring in resource-constrained environments. The distributed design is consistent with wireless home-health monitoring where physiological and activity data remain locally protected.