The Effect of Hyper-parameters in Model-contrastive Federated Learning Algorithm
Chen Shen, Zekai Lin, Jing Ma · 2023
Federated learning enables all parties to perform data analysis and machine learning through the updated use of global models. The key difficulty in federation learning is data heterogeneity, and there are numerous studies that address this challenge but with poor results. The model-contrastive federated learning (MOON) algorithm described in this paper, an improvement on FedAvg, is a Model-level federation learning algorithm that is concise and efficient. The core idea is to train a local model by comparing between global models, which in turn updates the global model without the need for privacy data uploads while model-level training solves the non-IID problem. Widely tests are performed by the authors to confirm the MOON algorithm’s efficacy. Contrasting various image databases, lr, µ and other different parameters reveals that MOON outperforms other federal learning algorithms in all aspects such as learning accuracy, convergence speed and training time. It demonstrates its advanced and robustness.