FedFit: Server Aggregation Through Linear Regression in Federated Learning
Taiga Kashima, Ikki Kishida, Ayako Amma, Hideki Nakayama · IEEE Access · 2024
We present a conceptually novel framework for Federated Learning (FL) called FedFit for a flexible solver to address FL problems. FedFit framework consists of two components: model compression to upload a local model from a client to the server and reconstruction of the compressed local model in the server. Clients upload a compressed local model using a “key” shared with the server to formulate the server aggregation aslinear regression. Therefore, the global model’s parameters are updated through a linear regression solver in the server while naturally contributing to reducing upload costs from clients to the server. Thanks to our framework design, the server can flexibly utilize various established linear regression techniques to address some open problems of FL by considering server aggregation from a different perspective—linear regression. As an example of the broad applicability of our idea, we demonstrate the effectiveness of robust regression and LASSO regression implemented on FedFit, which can alleviate the vulnerability issues against attacks on the global model from the collapsed clients and introduce sparsity to the global model toward the reduction of model size, respectively.