Federated Domain Generalization by Intra-Client Stable Feature Learning and Inter-Client Domain Adversarial Alignment

Federated Learning (FL), which serves as a distributed learning model focused on privacy protection, has received widespread attention due to its ability to collaboratively train models without exchanging clients’ raw data. However, since each client independently collects its local dataset, a single client may contain data from multiple domains, which leads to the existence of an agnostic distribution bias both intra-client and inter-client biases, and it is still a challenge to implement cross-domain generalization of the model in such an FL setting. In this paper, we propose a new stable federated adversarial domain generalization (Stable-FedADG) algorithm, which considers both intra-client and inter-client multi-domain distributions and through intra-client stable feature learning and inter-client domain adversarial alignment to learn domain-invariant features in federated scenarios. Experiments verify that the Stable-FedADG algorithm enhances the overall model performance across various federated environments.