Unlocking the potential of edge nodes: Range-extender for federated learning

With the rapid development of wireless networks and the widespread popularity of smart terminals, federated learning (FL) has attracted much attention as a distributed machine learning framework. This technique decentralizes the modeling process to mobile edge nodes, exploiting local data and edge arithmetic through collaboration. Although FL has many advantages, such as privacy protection, it still faces challenges in time management. Current FL frameworks suffer from inefficiencies in resource utilization (both synchronous and asynchronous), mainly due to idle arithmetic caused by communication gaps. In this case, collaboration time is usually wasted in long communication waits. To cope with this problem, we propose an edge node training range extender which can effectively utilize the communication window period for local training, thus compensating for edge node idling conditions and unleashing the potential of edge node training. This novel FL strategy revisits the FL process and provides two fused forms of additional training gradients. We critically analyze the convergence of additional FL and compare it with the mainstream FL frameworks at this stage. We demonstrate the potential benefits of this new strategy by performing a comprehensive analysis of the CIFAR10 and CIFAR100 datasets for a classification task.