A Trust and Energy-Aware Double Deep Reinforcement Learning Scheduling Strategy for Federated Learning on IoT Devices

Federated learning is a revolutionary machine learning approach whose main idea is to train the machine learning model in a distributed fashion over a large number of edge/end devices without having to share the raw data. We consider in this work a federated learning scenario wherein the local training is carried out on IoT devices and the global aggregation is done at the level of an edge server. One essential challenge in this emerging approach is scheduling, i.e., how to select the IoT devices to participate in the distributed training process. The existing approaches suggest to base the scheduling decision on the resource characteristics of the devices to guarantee that the selected devices would have enough resources to carry out the training. In this work, we argue that trust should be an integral part of the decision-making process and therefore design a trust establishment mechanism between the edge server and IoT devices. The trust mechanism aims to detect those IoT devices that over-utilize or under-utilize their resources during the local training. Thereafter, we design a Double Deep Q Learning (DDQN)-based scheduling algorithm that takes into account the trust scores and energy levels of the IoT devices to make appropriate scheduling decisions. Experiments conducted using a real-world dataset (https://www.cs.toronto.edu/similar to kriz/cifar.html) show that our DDQN solution always achieves better performance compared to the DQN and random scheduling algorithms.