Maximizing Global Model Appeal in Federated Learning

Yae Jee Cho, Divyansh Jhunjhunwala, Tian Li, Virginia Smith, Gauri Joshi

Abstract: Federated learning (FL) aims to collaboratively train a global model using local data from a network of clients. To warrant collaborative training, each federated client may expect the resulting global model to satisfy some individual requirement, such as achieving a certain loss threshold on their local data. However, in real FL scenarios, the global model may not satisfy the requirements of all clients in the network due to the data heterogeneity across clients. In this work, we explore the problem of global model appeal in FL, which we define as the total number of clients that find that the global model satisfies their individual requirements. We discover that global models trained using traditional FL approaches can result in a significant number of clients unsatisfied with the model based on their local requirements. As a consequence, we show that global model appeal can directly impact how clients participate in training and how the model performs on new clients at inference time. Our work proposes MaxFL, which maximizes the number of clients that find the global model appealing. MaxFL achieves a $22$-$40\%$ and $18$-$50\%$ improvement in the test accuracy of training clients and (unseen) test clients respectively, compared to a wide range of FL approaches that tackle data heterogeneity, aim to incentivize clients, and learn personalized/fair models.

Submission Length: Regular submission (no more than 12 pages of main content)

Code: https://openreview.net/attachment?id=8GI1SXqJBk&name=supplementary_material

Assigned Action Editor: ~Aurélien_Bellet1

Submission Number: 1790