Communication efficiency optimization in federated learning based on multi-objective evolutionary algorithm

• Published in: Evolutionary intelligence Vol. 16; no. 3; pp. 1033 - 1044
• Main Authors: Chai, Zheng-yi, Yang, Chuan-dong, Li, Ya-lun
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.06.2023; Springer Nature B.V
• Subjects: Accuracy; Applications of Mathematics; Artificial Intelligence; Artificial neural networks; Bandwidths; Bioinformatics; Communication; Control; Costs; Data compression; Data integrity; Deep learning; Efficiency; Energy consumption; Engineering; Evolutionary algorithms; Experiments; Federated learning; Genetic algorithms; Machine learning; Mathematical and Computational Engineering; Mechatronics; Multilayers; Multiple objective analysis; Neural networks; Optimization; Pareto optimization; Privacy; Research Paper; Robotics; Sorting algorithms; Sparsity; Statistical Physics and Dynamical Systems; Federated learning; Multi-objective evolutionary optimization; Deep neural networks; Network connectivity; Communication cost
• ISSN: 1864-5909; 1864-5917
• DOI: 10.1007/s12065-022-00718-x

Federated learning is an emerging technology that can effectively safeguard personal information. As opposed to traditional centralized learning, federated learning can avoid data sharing while maintaining global model training. However, in the process of updating the global model, it will consume huge client communication resources, which hinder the wide application of this technology. To reduce the communication overhead without seriously reducing the accuracy of the global model, under the federated learning framework, we use decomposition based multi-objective optimization algorithm (MOEA/D) to optimize the structure of the global model. For the structure of the global model, a highly scalable coding method is used for coding, which improves the efficiency of the evolutionary neural network. As a comparison, we use the non-dominated sorting genetic algorithm II(NSGA II) to optimize the problem under the same conditions and verify the effectiveness of both algorithms according to the obtained Pareto solution. We verify that MOEA/D has better convergence when using multilayer and convolutional neural networks as the global model. Overall, MOEA/D can further strengthen the structure optimization of the federated learning model and reduce communication costs.