Energy-Efficient and Data-Optimized Federated Learning for Distributed On-Device Intelligence

The advent of Artificial Intelligence (AI), requiring large data from diverse devices, necessitates novel architectures that incorporate edge devices in the pipeline of AI operations. Federated Learning (FL) is becoming the state-of-practice for distributed, collaborative training of AI models at the network edge, respecting data privacy and anonymity concerns. However, implementing FL on battery-powered edge devices with limited computing capabilities necessitates innovative approaches to balance training accuracy and energy efficiency. In this work, we propose a framework for energy-efficient and data-optimized FL to facilitate on-device intelligence. The proposed framework jointly optimizes the (a) selection of the most important data samples from the devices’ local training datasets to maintain FL model accuracy, and their (b) computing frequency and (c) uplink transmission power to minimize the energy consumption during local computation and communication phases. A multi-processor model is considered for each device, extending data selection to allocate data samples to each device’s processors. The initially formulated non-convex and combinatorial optimization problem is decomposed into sub-problems, each equivalently transformed into a convex form. An Alternating Optimization (AO) approach is then applied to the sub-problems, yielding a close-to-optimal solution to the original problem. The proposed algorithm is distinguished in terms of scalability, achieving at least 30% energy reduction on the MNIST and CIFAR10 datasets compared to the second-best benchmarking schemes, while maintaining FL model accuracy at nearly the same level. Overall, the proposed framework proves scalable and suitable for deployment in realistic resource-constrained edge environments.