OLG-FL: A Federated Learning Framework for Optimizing Local Training and Global Aggregation

Federated learning(FL) allows collaborative training of machine learning models across decentralized clients without compromising data privacy, demonstrating substantial success across various applications such as medical imaging, biometric recognition, and object detection. However, real-world federated learning scenarios typically involve non-independent and identically distributed (non-IID) data, creating local biases that significantly degrade global model performance. Existing approaches primarily address the non-IID issue through improved model aggregation, constrained local training, or simulated data sharing, but often overlook intrinsic client knowledge and their actual contributions to global model improvement. To effectively mitigate these limitations, this paper introduces a novel federated learning framework termed OLG-FL (Optimize Local training and Global aggregation Federated Learning). OLG-FL incorporates a contrastive learning mechanism at the client side, aligning local and global class feature spaces to reduce local biases caused by non-IID data. Simultaneously, client contributions are quantified based on the similarity between their local updates and the global update direction, guiding more effective global model aggregation. Extensive experiments under diverse non-IID scenarios demonstrate that OLG-FL significantly outperforms state-of-the-art methods, achieving higher accuracy and robustness with acceptable computational and communication overhead.