Delay and Energy Efficient Federated Learning over Heterogeneous Mobile Devices

The unprecedented growth in mobile data has created an opportunity for artificial intelligence (AI) to leverage this wealth of information for learning models, which in turn make remarkable improvements on personalized and efficient mobile services, ranging from predictive typing to location-based recommendations. The privacy-sensitive nature of mobile data precludes its central gathering to form the necessary large datasets for the development of high-performance AI models. To overcome this challenge, federated learning (FL) has emerged as a promising distributed machine learning paradigm that enables mobile devices to collaboratively learn a model without sharing raw data. However, practical deployment of FL over mobile devices is very challenging because (i) conventional FL training is power-hungry and time consuming for mobile devices due to interleaved local on device training and communications of model updates, (ii) there are heterogeneous training data across mobile devices, and (iii) mobile devices have hardware heterogeneity in terms of computing and communication capabilities.

This dissertation focuses on delay- and energy-efficient FL over mobile devices. In this dissertation, two delay-efficient FL schemes over mobile devices, i.e., SDEFL and FedCR, are proposed to minimize the training latency of mobile devices from the aspects of algorithm design and system design, respectively. Additionally, inspired by our observation of the huge amount of energy saved by high-speed wireless communications during the FL training, we propose a energy-aware local stochastic gradient descent policy to minimize the energy of FL over mobile devices. Furthermore, a full prototype of FL system is implemented to validate the effectiveness of the proposed designs under various FL tasks, learning models, datasets and wireless transmission rates.