Towards Fair Federated Learning with Zero-Shot Data Augmentation

Abstract

Federated learning has emerged as an important distributed learning paradigm, where a server aggregates a global model from many client-trained models, while having no access to the client data. Although it is recognized that statistical heterogeneity of the client local data yields slower global model convergence, it is less commonly recognized that it also yields a biased federated global model with a high variance of accuracy across clients. In this work, we aim to provide federated learning schemes with improved fairness. To tackle this challenge, we propose a novel federated learning system that employs zero-shot data augmentation on under-represented data to mitigate statistical heterogeneity, and encourage more uniform accuracy performance across clients in federated networks. We study two variants of this scheme, Fed-ZDAC (federated learning with zero-shot data augmentation at the clients) and Fed-ZDAS (federated learning with zero-shot data augmentation at the server). Empirical results on a suite of datasets demonstrate the effectiveness of our methods on simultaneously improving the test accuracy and fairness.

No Fear of Heterogeneity: Classifier Calibration for Federated Learning with Non-IID Data

“…The key motivation behind data sharing is that a client cannot acquire samples from other clients during local training, thus the learned local model under-represents certain patterns or samples from the absent classes. The common practices are to share a public dataset [6], synthesized data [16,17] or a condensed version of the training samples [18] to supplement training on the clients or on the server. This line of works may violate the privacy rule of federated learning since they all consider sharing raw input data of the model, either real data or artificial data.…”

“…There have been a plethora of works exploring promising solutions to federated learning on non-IID data. They can be roughly divided into four categories: 1) client drift mitigation [5,8,9,10], which modifies the local objectives of the clients, so that the local model is consistent with the global model to a certain degree; 2) aggregation scheme [11,12,13,14,15], which improves the model fusion mechanism at the server; 3) data sharing [6,16,17,18], which introduces public datasets or synthesized data to help construct a more balanced data distribution on the client or on the server; 4) personalized federated learning [19,20,21,22], which aims to train personalized models for individual clients rather than a shared global model. However, as suggested by [7], existing algorithms are still unable to achieve good performance on image datasets with deep learning models, and could be no better than vanilla FedAvg [2].…”