Weighted Averaging Federated Learning Based on Example Forgetting Events in Label Imbalanced Non-IID

Federated learning, a data privacy-focused distributed learning method, trains a model by aggregating local knowledge from clients. Each client collects and utilizes its own local dataset to train a local model. Local models in the connected federated learning network are uploaded to the server. In the server, local models are aggregated into a global model. During the process, no local data is transmitted in or out of any client. This procedure may protect data privacy; however, federated learning has a worse case of example forgetting problem than centralized learning. The problem manifests in lower performance in testing. We propose federated weighted averaging (FedWAvg). FedWAvg identifies forgettable examples in each client and utilizes that information to rebalance local models via weighting. By weighting clients with more forgettable examples, such clients are better represented and global models can acquire more knowledge from normally neglected clients. FedWAvg diminishes the example forgetting problem and achieve better performance. Our experiments on SVHN and CIFAR-10 datasets demonstrate that our proposed method gets improved performance compared to existing federated learning algorithm in non-IID settings, and that our proposed method can palliate the example forgetting problem.