ACMFed: Fair Semi-Supervised Federated Learning With Additional Compromise Model

초록

One of the major drawbacks of federated learning (FL) is data imbalance and uneven reliability of clients, which adversely impacts model performance and generalization ability. Among these problems, data imbalance can be addressed through semi-supervised federated learning (SSFL), in which fully labeled, and fully unlabeled clients are jointly trained to derive a global model. Existing approaches work well when local clients have independent identically distributed (IID) data but fail to generalize under a more practical SSFL setting, namely fairness, non-IID. Since the non-IID problem stems from unfairness, to address this issue, we propose ACMFed, a fair and improved concurrent federated learning method that balances client connections, computing resources, data quality and quantity based on fairness. ACMFed reaches the compromise by performing fair multiple sampling of clients in models with largely deviating labeled and unlabeled clients. Each client distills the model through a multi-sub-sampling process without aggregating the local model to the central server, and builds a global model based on it. The additional compromise model applies a fair adjustment method of the new distance-based dynamic model to reduce data bias and optimize performance. ACMFed outperforms other federated learning methods and achieves faster convergence on six benchmark datasets. To code is available at https://github.com/kimdohyoung96/ACMFed.

키워드