Iterative Optimization for Edge Federated Learning

Federated learning emphasizes the protection of data privacy in distributed machine learning. It is similar to the data-centered distributed machine learning that trains a model for making predictions or decisions without being explicitly programmed. The computing workers in federated learning provide training for the model at the edges of network where the data are stored. Thus, they can control the data and decide whether and when participating in the learning is needed. This paper analyzes the difference between centralized machine learning and federated learning, and the affects of communication frequency between the server and clients on the learning accuracy. It proposes two variations of federated learning optimizations by studying its process. The experimental results demonstrate that the centralized machine learning often receives a far better training result under the same number of training samples compared with federated learning. Furthermore, increasing the communication frequency between the server and clients can improve the learning result.