Optimization of Moon Algorithm Based on Absolute Value Pruning and Differential Privacy
TL;DR
This study proposes an optimization scheme of the MOON algo-rithm based on absolute value pruning and differential privacy and verifies the effectiveness of the two on the improvement of communication efficiency and privacy protection ability of MOON algorithm.
Abstract
: With the continuous expansion of AI application scenarios, data privacy and security issues are becoming more and more prominent, prompting federated learning as a distributed machine learning framework to be widely noticed. However, it has an obvious data heterogeneity problem in non-independent, same-distributed data scenarios. Model-Contrastive Federated Learning (MOON) effectively mitigates this problem through the contrastive learning mechanism, but there is still much room for optimization in terms of communication efficiency and privacy protection. Therefore, this study proposes an optimization scheme of the MOON algo-rithm based on absolute value pruning and differential privacy. Experiments show that under reasonable parameter settings, the optimized MOON algorithm significantly reduces the communication duration and enhances the privacy protection ability within the acceptable accuracy degradation range. This study verifies the effectiveness of the two on the improvement of communication efficiency and privacy protection ability of MOON algorithm, which is of significance for the application of federated learning in scenarios requiring both efficient communication and high data privacy.