Adaptive Model Aggregation in Federated Learning Based on Model Accuracy
Rebekah Wang, Yingying Jennifer Chen · IEEE Wireless Communications · 2024
Federated learning is useful when predicting user preferences due to its ability to keep user data private. As such, certain data samples may be more useful than others. For instance, with a product recommendation system, recent purchase history is more useful and relevant than old purchase history. Models trained with more useful data samples would make better predictions. Additionally, weighted model aggregation should be implemented. A client model trained with more useful data should be given a heavier aggregation weight (the relative proportion that model updates are multiplied by to obtain a global model during model aggregation). In this article, accuracy-based federated learning with adaptive model aggregation (A-FLAMA) is introduced. With this algorithm, client models that achieve higher accuracy with the server's test data are assigned heavier aggregation weights. This eliminates the need for extra information from clients. To improve accuracy and reduce computation in A-FLAMA, variations were explored. To assess these approaches, they were compared to each other, FLAMA, and benchmark federated learning (i.e., FedAvg) via experiments using the modified National Institute of Standards and Technology dataset. After each training round, global model accuracies were calculated to graph a global model convergence chart. The results showed A-FLAMA and its variations achieved similar to, or higher global accuracies than, the original FLAMA, establishing them as valid alternatives and improvements to FLAMA.