Anchor-Based FedAvg

An Distributed Invariant Feature Learning Algorithm

  • Proposed a novel federated learning scheme making use of the data distribution of public and private data and can handle the out of distribution data on the clients;

  • Designed the training framework for the proposed scheme; Designed a problem setting on CelebA and constructed the dataset for federated learning that requires domain generalization; Outperformed the FedAvg optimizer on both i.d. and o.o.d accuracy.