RSF: Reinforcement learning based hybrid split and federated learning for edge computing environments

The increasing usage of the Internet of Things (IoT) in various domains, particularly within the realm of artificial intelligence has raised concerns about data privacy. Federated learning has been established to address this issue. However, fed-erated learning faces challenges such as constrained computing resources on IoT devices, fluctuating network bandwidth, and low scalability due to network congestion during the transition of local weights in a large scale. All of these results in a low convergence time, or even failure in the training procedure. This paper presents a reinforcement learning-based dynamic neural network splitting method called RSF to reduce overall training time with a faster partitioning decision time. The reinforcement learning agent dynamically allocates computational tasks across IoT devices, Edge servers, and the cloud, accelerating the training process in edge computing environments. The neural network partitioning adapts iteratively. The agent decides on the neural network partitioning for each IoT device at the start of each training round. Simulation results demonstrate the effectiveness of our proposed method, achieving substantial speed improve-ments up to 27%, and higher scalability.