FedCon: Scalable and Efficient Federated Learning via Contribution-Based Aggregation

With the increasing application of federated learning to medical and image data, the challenges of class distribution imbalances and Non-IID heterogeneity across clients have become critical factors affecting the generalization ability of global models. In the medical domain, the phenomenon of data silos is particularly pronounced, leading to significant differences in data distributions across hospitals, which in turn hinder the performance of global model training. To address these challenges, this paper proposes FedCon, a federated learning method capable of dynamically adjusting aggregation weights, while accurately evaluating client contributions. Specifically, FedCon initializes aggregation weights based on client data volume and class distribution and employs Monte Carlo sampling to effectively simplify the computation of Shapley values. Subsequently, it further optimizes the aggregation weights by comprehensively considering the historical contributions of clients and the similarity between clients and the global model. This approach significantly enhances the ability to generalize and update the stability of the global model. Experimental results demonstrate that, compared to existing methods, FedCon achieved a superior generalization performance on public datasets and significantly accelerated the convergence of the global model.