Paper Info
Title
Privacy is not Free: Energy-Aware Federated Learning for Mobile and Edge Intelligence
Abstract
In mobile and edge intelligence systems, federated learning (FL) enables local data training and learning model sharing without obtaining actual data from mobile and edge users, which are data owners. Data training processes are performed at the user side with only trained gradients passed to an aggregator, i.e., learning server. The learning server continually trains and updates corresponding learning models by collecting gradients. The updated learning models are delivered back to the mobile and edge users for improved data training results. Despite the advantages of federated learning in preserving privacy, the local data training process will consume an adequate amount of energy from the perspective of mobile and edge users. In mobile and edge intelligence systems, mobile users may not always prefer to apply federated learning to reduce energy consumption. There is a trade-off between applying federated learning to reserve data privacy and updating actual data for learning servers to train. In this work, a Markov decision process (MDP) based system model is proposed for mobile and edge users to make federated learning decisions to optimize long-term performance in terms of utility function consists of data training reward and data processing delay. A deep reinforcement learning approach is proposed to solve the MDP problem in highly dynamic systems with a large state space.
Year
DOI
Venue
2020
10.1109/WCSP49889.2020.9299703
2020 International Conference on Wireless Communications and Signal Processing (WCSP)
Keywords
DocType
ISSN
Federated learning,edge intelligence,Markov decision process,reinforcement learning
Conference
2325-3746
ISBN
Citations 
PageRank 
978-1-7281-7237-8
0
0.34
References 
Authors
7
6
Name
Order
Citations
PageRank
Wenqi Yang100.34
Yang Zhang231319.51
Wei Yang Lim319313.37
Zehui Xiong458654.94
Yutao Jiao500.34
Jiangming Jin6151.18