Robust Federated Learning: The Case of Affine Distribution Shifts

摘要: Federated learning is a distributed paradigm that aims at training models using samples across multiple users in network while keeping the on users' devices with aim of efficiency and protecting privacy. In such settings, data often statistically heterogeneous manifests various distribution shifts users, which degrades performance learnt model. The primary goal this paper to develop robust federated algorithm achieves satisfactory against samples. To achieve goal, we first consider structured affine shift captures device-dependent heterogeneity settings. This perturbation model applicable problems as image classification where images undergo imperfections, e.g. different intensity, contrast, brightness. address propose Learning framework Robust Affine (FLRA) provably Wasserstein observed solve FLRA's minimax problem, fast efficient optimization method provide convergence guarantees via gradient Descent Ascent (GDA) method. We further prove generalization error bounds for classifier show proper from empirical true underlying distribution. perform several numerical experiments empirically support FLRA. an indeed suffices significantly decrease new test user, our proposed significant gain comparison standard adversarial methods.