A General Federated Learning Scheme with Blockchain on Non-IID Data

The security of machine learning has received a lot of attention from the community. Federated learning enables more secure training processes of models in machine learning via local training and parameter interactions of participants. However, participants' data usually shows significant differences, i.e., the characteristics of non-IID, affecting the convergence speed and accuracy of models to a large extent. In this paper, we propose a general federated learning scheme with blockchain to cover the shortage of federated learning caused by non-IID data. Specifically, each participant trains a GAN via local data first and then shares the generator corresponding to the GAN with the assistance of the blockchain. Based on the generator parameters on the blockchain, each participant augments the local data and trains the local model, alleviating a series of problems caused by the non-IID data. The scheme achieves effective training of models while ensuring security. Experimental results show that the proposed scheme can speed up model convergence and improve the model's accuracy simultaneously. In the non-IID scenario, compared with the federated learning benchmark scheme, the accuracy in our scheme can be improved by up to 17%.