Uniting farms: Federated learning for sensor-based animal activity recognition

Automated animal activity recognition (AAR) has achieved great success due to the development of deep learning methods trained on large-scale datasets, providing rich insights into animal health and welfare and alleviating the workloads of animal caretakers and veterinarians. However, constructing centralised data across diverse sources (e.g., farms) faces two challenges: 1) data ownership and privacy issues when accessing farm data, and 2) limited storage and computational capabilities in a single central repository. Federated learning (FL), which allows data owners to collectively train a model while keeping their data stored locally, provides a privacy-preserving decentralised solution. This study introduced the FL-based framework for the first time to AAR fields and explored its feasibility and effectiveness in improving model performance by uniting sensor data from different farms. Three state-of-the-art FL strategies (i.e., FedAvg, FedProx, and FedBN) were compared against SingleSet (i.e., training an individual model within each client) based on two public datasets. These two datasets consist of 87,621 and 42,943 2-s motion data (tri-axial acceleration and tri-axial angular velocity) acquired from horses and goats, respectively. The results demonstrated that FedAvg, FedProx, and FedBN could accurately classify activities of horses and goats, outperforming the SingleSet with different increments in average accuracy (horses: 12.07%, 12.05%, 11.89%; goats: 4.05%, 4.07%, 4.16%). This proved the promising capability of FL to enhance AAR's performance without privacy leakage. In addition, empirical analyses were conducted to assess FL's performance from two aspects, including data sizes and clients numbers, providing rich insights into FL's appropriate applications in the future.