Adaptation Diffusion Strategy Over Wireless Fading Channels
Amjad Ali · 2022
Federated learning has introduced the solution to train machine learning model with data distributed over network nodes, where the training data remains stored on the nodes, which may be phones or network sensors [1]. With federated learning the nodes don't exchange the data which may be of large volumes and may be considered too sensitive but exchange just the parameters of the models, thereby ensuring a basic level of privacy [2]. In this work, we introduce new strategy for federated learning over nodes of wireless network. Since FEDAVG algorithm is centralized (uses the structure client-server), we distribute the learning over the graph nodes using the concept of diffusion strategy, where we suppose that the graph is adapted in time. We show how this strategy overcomes heterogeneity and converges in fewer rounds of communication. The main goal is to employ diffusion strategy in a decentralized manner to improve the convergence time without increasing the communication rounds. We compare the performance of the new approach with the consensus algorithm for the client's IID data (IID-Independent and Identically Distributed) and non-IID client data.