Energy and Latency-aware Computation Load Distribution of Hybrid Split and Federated Learning on IoT Devices

Abstract

Split learning (SL) and Federated Learning (FL) are popular distributed learning frameworks used to increase data privacy and reduce computation loads of Internet of Things (IoT) devices. However, one of the major challenges of distributed learning on IoT devices is determining the portion of computation load to be assigned for the devices compared to the server side. The contributions of the existing works in the literature are either limited by consideration of homogeneous resources available at all IoT devices or by not distributing computation loads among the devices in an efficient way. In this paper, we propose an adaptive clustering-based computation load distribution method for IoT devices, with heterogeneous resource capacities, participating in the model training. The clustering makes the optimal determination of the split point of the learning model, which is scalable even for a large number of devices. The numerical evaluation of the proposed learning model implemented using Python 3.0 and the comparative performance results show that the proposed load distribution policy for the learning models reduces the time by 160 times on average compared to the usual brute force method.