AdaptPSOFL: Adaptive Particle Swarm Optimization-Based Layer Offloading Framework for Federated Learning

Abstract

Following the rising privacy requirement enforced by regulations such as the European General Data Protection Regulation and the HIPAA, the centralized training of machine learning models has become difficult. Federated learning has come under extensive adoption because it allows clients to train a local model on a local dataset and then send the model to a server for aggregation. This aspect allows clients to preserve the privacy of their local data, while also allowing the training of a global model. A new area of study that is emerging is the possibility of offloading the training of the local model between a client and a server to influence the training time of the model. Studies such as FedAdapt use a reinforcement learning agent to decide the optimal split between a client and a server. In our study, we determine the optimal split between a client and server by formulating an optimization problem that simultaneously minimizes the CPU utilization and the round execution time while considering the number of layers placed on the client as decision variables. We apply particle swarm optimization to approach the optimization problem. Based on our experiments, we observed that our scheme outperforms FedAdapt and classic federated learning on the combined objectives by 36.60% and 36.73% on average respectively.