An Adaptive Weighting Approach to Enhance the Efficiency and Robustness of Federated Learning

Federated Learning (FL) is a distributed machine learning method that allows multiple devices to train models locally and aggregate updates at a central server to ensure data privacy. However, constructing a robust global model is challenging due to limited access to local data and heterogeneous client data. Furthermore, malicious clients and high communication and computation costs further complicate the design of FL algorithms. Existing methods for evaluating client contributions are computationally intensive, unsuitable for large-scale applications. To overcome these challenges, this paper proposes an adaptive weighting FL framework based on the Banzhaf index, providing a more efficient mechanism for evaluating and allocating client contributions. The Banzhaf index simplifies the computation process compared to traditional methods, allowing for better scalability. Additionally, we introduce a Neyman allocation stratified sampling strategy to optimize client selection, which effectively reduces variance and ensures balanced client representation, further enhancing both model performance and computational efficiency. Experimental results show that the proposed method significantly improves training efficiency, model accuracy, and robustness across various datasets, while reducing computation costs, making it well-suited for large-scale FL.