Federated Learning (FL) enables collaborative training without cen- tralizing sensitive medical data, but conventional weight exchange in- curs high communication costs and privacy risks. This work compares parameter-based and knowledge-based strategies for federated fall de- tection using three public datasets. While unweighted FedAvg achieves the strongest performance (0.9692 global PR-AUC), no-weight alterna- tives maintain competitive local results. Notably, soft-label communi- cation reduces the communication payload from 49,665 parameters to only two values per round, offering the most efficient trade-off. These findings demonstrate the potential of lightweight knowledge-sharing for resource-constrained wearable healthcare systems.

This work was partially supported by Fundação para a Ciência e a Tec- nologia (FCT), I.P., through the ISTAR-Iscte Multiannual Funding Pro- gramme under Project UID/04466/2025 (https://doi.org/10. 54499/UID/04466/2025).

Federated Learning,Communication,Security,Privacy