Comparison of Federated Learning Algorithms for Image Classification

In the Traditional machine learning algorithms, users are required to transmit source data to the cloud server with huge computing power during private training. It leads to a lot of data out of control flow and sensitive data be leaked. Then federated learning (FL) is proposed. This a concept that leverages data spread across many devices, to learn classification tasks separately without recourse to data sharing. But federated learning also has many defects. To solve these problems, many improved algorithms based on traditional federated learning are proposed, they are designed to solve one aspect of the problems. This work introduces five algorithms of them, including the federated averaging algorithm (Fedavg), federated matched averaging algorithm (FedMA), FedProx, federated attentive message passing (FedAMP) algorithm, and model-contrastive learning (MOON) algorithm. The background of their birth and the problems they aspired to solve are introduced. In addition, their advantages and disadvantages are also mentioned for better analysis and comparison. Last, their performances are compared on Tiny-ImageNet, a dataset similar to the ImageNet project but is refined and smaller, to get more recognition of these algorithms.