TKAGFL: A Federated Communication Framework Under Data Heterogeneity

Federated learning still faces many problems from research to technology implementation and the most critical problem is that the communication efficiency is not high. Therefore, the main task of this work is to improve the communication efficiency, affected by communication devices, bandwidth, data distribution and so on. Our work focus on two main aspects of federated learning - updates strategy and data heterogeneity, and propose a new federated framework called TKAGFL, which is composed by three key factors. Firstly, in response to the problem of data heterogeneity that often occurs in actual federated learning, federated generative adversarial networks are proposed as a method of data preprocessing. Secondly, the improved homomorphic encryption method counterbalances the conflicts between data-sharing and privacy protection. Thirdly, in the communication process, classic AdaGrad optimization in deep learning and top-K algorithms are combined to compress communication parameters to improve communication efficiency. The results of experiments demonstrate our TKAGFL framework's accuracy is about 15% - 20% better than other algorithms or frameworks and it can converge at 150 communication rounds and is 50 rounds faster than others. Also, our TKAGFL algorithm reduces 10 times communication volume than other algorithms,which is beneficial for federated learning application in industry.

Automated summary

This research focuses on addressing the communication efficiency problem in federated learning and proposes a new framework called TKAGFL. The TKAGFL framework improves communication efficiency through the use of federated generative adversarial networks, improved homomorphic encryption methods, and optimized communication parameter compression algorithms. Experimental results demonstrate that the TKAGFL framework achieves higher accuracy and faster convergence compared to other algorithms or frameworks.