Server Aggregation as Linear Regression: Reformulation for Federated Learning

Taiga Kashima, Ikki Kishida, Ayako Amma, Hideki Nakayama

Keywords: Federated Learning, Machine Learning

Abstract: We propose a conceptually novel framework for Federated Learning (FL) called FedFit to mitigate issues of FL. FedFit is a reformulation of the server aggregation in FL, where the global model is updated by linear regression. This reformulation naturally enables us to utilize the established linear regression techniques for several FL issues. For example, we apply robust regression to alleviate the vulnerability issue against attacks on the global model from collapsed clients, and we apply LASSO regression to introduce sparsity into the model to reduce the communication cost in FL. Moreover, FedFit enables clients to upload compressed model parameters to the server, significantly reducing the data traffic. In experiments, we demonstrate that FedFit successfully improves robustness against attacks on a global model by robust regression and reduces the global model size by LASSO regression.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: General Machine Learning (ie none of the above)