Federated Client Selection and Attention-Based Aggregation Algorithm on Non-IID Data

Federated learning is a decentralized machine learning approach that enables multiple devices to collaboratively train a global model while keeping their data localized, thus preserving user privacy. However, the presence of non-independent and identically distributed (Non-IID) data across clients presents significant challenges, including performance degradation and difficulties in model optimization. This paper proposes a novel Federated Client Selection and Attention Aggregation (FCSA) algorithm, designed to address these issues. FCSA improves client selection by considering factors such as training time, data diversity, and historical performance, thereby prioritizing clients that can contribute the most to the global model. On the server side, an attention-based aggregation mechanism is introduced to assign different weights to client updates based on their contributions, optimizing the model aggregation process. Experimental evaluations conducted on the MNIST dataset under Non-IID conditions demonstrate that FCSA outperforms traditional methods like FedAvg, improving model accuracy by approximately 17.5%. Future work will focus on dynamically adjusting weighting strategies to further enhance efficiency across various training stages.