Auto-weighted Robust Federated Learning with Corrupted Data Sources

Federated learning provides a communication-efficient and privacy-preserving training process by enabling statistical models with massive participants while keeping their data in local clients. However, standard federated techniques that naively minimize an average loss function are vulnerable to corruptions from outliers, systematic mislabeling, or even adversaries. In addition, it is often prohibited for service providers verify the quality of samples due increasing concern user privacy. this paper, we address challenge proposing Auto-weighted Robust Learning (arfl), novel approach jointly learns global model weights updates provide robustness against corrupted sources. We prove bound on expected risk respect predictor clients, which guides definition objective robust learning. The allocated comparing empirical client best p clients (p-average), thus can downweight significantly high losses, thereby lower contributions model. show achieves when distributed differently benign ones. To optimize function, propose algorithm based blockwise minimization paradigm. conduct experiments multiple benchmark datasets, including CIFAR-10, FEMNIST Shakespeare, considering different deep neural network models. results our solution scenarios label shuffling, flipping noisy features, outperforms state-of-the-art methods most scenarios.