Search

Federated learning develops multiple clients with a universal framework platform to train deep learning models with excellent performance. During this process, the local data set is not leaked to the outside, making federated learning a powerful method to protect data security while solving the problem of isolated data islands. However, in some usage scenarios, different clients are holding local data sets which are non- Independent Identically Distribution, i.e., non-IID. There have been significant attempts made to overcome this difficulty, and one of these efforts is the model-contrastive federated learning framework, as known as MOON algorithm. The main concept is to make use of the similarities between the global and local models to adjust the local training. The final global model performs well with non-IID data. The paper aims to replicate and analyze this algorithm to provide data reference for subsequent use. The algorithm is tested on the MNIST data set to get its performance on the handwritten digit classification task. This paper processed the data set to make it non-IID, but the result achieved a high accuracy rate with an acceptable computational cost.