Federated learning with ℓ1 regularization

Yong Shi, Yuanying Zhang, Peng Zhang, Yang Xiao, Lingfeng Niu

Abstract: Highlights•Present a federated learning method called Fed ℓ1 to address client drift.•Introduce ℓ1 regularizer to control clients and avoid unnecessary parameter updates.•Design a stochastic subgradient descent algorithm to train the model.•Compare Fed ℓ1 with baselines on several datasets and discuss the improvement.