An Improved MOON Federated Learning Algorithm Based on Negative Sample Quality Awareness

Federated learning achieves privacy protection through multi-device collaborative training, but model drift caused by the heterogeneity of Non-IID data severely impacts performance. The MOON algorithm uses the previous-round model as negative samples and the global model as positive samples to introduce a contrastive loss to mitigate drift, but indiscriminately rejecting the previous-round model results in the loss of valuable knowledge. This paper proposes an adaptive contrastive federated learning method based on negative sample quality awareness: the rejection weight is dynamically adjusted by calculating the cosine similarity between the previous-round model and the global model—reducing rejection when quality is high to preserve valuable knowledge, and maintaining normal rejection when quality is low to correct bias. In the CIFAR-10 Non-IID scenario, the test accuracy reaches 74.78% (a 0.58% improvement over MOON), the average loss is reduced by approximately 10.09%, and convergence is better, validating the effectiveness of this adaptive mechanism.