Communication-Efficient Federated Learning for Large-Scale Multiagent Systems in ISAC: Data Augmentation With Reinforcement Learning

Abstract

Integrated sensing and communication (ISAC) has attracted great attention with the gains of spectrum efficiency and deployment costs through the coexistence of sensing and communication functions. Meanwhile, federated learning (FL) has great potential to apply to large-scale multiagent systems (LSMAS) in ISAC due to the attractive privacy protection mechanism. Nonindependent identically distribution (non-IID) is a fundamental challenge in FL and seriously affects the convergence performance. To deal with the non-IID issue in FL, a data augmentation optimization algorithm (DAOA) is proposed based on reinforcement learning (RL), where an augmented dataset is generated based on a generative adversarial network (GAN) and the local model parameters are inputted into a deep Q-network (DQN) to learn the optimal number of augmented data. Different from the existing works that only optimize the training performance, the number of augmented data is also considered to improve the sample efficiency in the article. In addition, to alleviate the high-dimensional input challenge in DQN and reduce the communication overhead in FL, a lightweight model is applied to the client based on deep separable convolution (DSC). Simulation results indicate that our proposed DAOA algorithm acquires considerable performance with significantly fewer augmented data, and the communication overhead is reduced greatly compared with benchmark algorithms.