IOFL: Intelligent-Optimization-Based Federated Learning for Non-IID Data
Xinyan Li, Huimin Zhao, Wu Deng
Abstract
Federated learning (FL) algorithm has been widely studied in recent years due to its ability for sharing data while protecting privacy. However, FL has risks such as model inversion attack, and is less effective when data is non-independent and identically distributed (non-IID). In response to these challenges, an intelligent optimization-based federated learning (IOFL) framework is developed to improve the privacy protection performance and global model performance in this paper. In the IOFL, the server searches model parameters by using intelligent optimization algorithm and distributes it to the clients. The clients use local data to validate the issued model by the server and return the validation results to the server. The server calculates the fitness function based on the weighted average of the received validation results, which guide the intelligent optimization algorithm to search for new model parameters. The experimental results on MNIST and Fashion-MNIST dataset show that the accuracy of the IOFL can reach over 0.8 and 0.68 under different non-IID settings with 200 round communications, whose performance is not affected by non-IID data distribution at clients.