基于联邦学习的图像语义通信系统

An Image Semantic Communication System Based on Federated Learning

作为一种全新的通信范式,语义通信旨在提取并传递数据源中的语义信息而不是符号级的精确传输,极大地提高了通信效率,为6G移动通信提供了新思路。针对未来万物智联场景,为实现模型在多用户之间有效训练、动态更新、及时分发,同时在提供隐私保护的前提下充分利用广泛分布的用户数据,首先构建基于自编码器的图像语义通信系统,然后提出利用联邦学习对语义通信系统进行分布式的训练、更新和分发,为了进一步减少学习过程中模型频繁更新带来的巨大通信负载,根据模型更新前后参数差值动态范围较小(-0.1~0.1)的特点,提出针对模型差值更新的、包含放大,取整,边界限制,缩小共4步的量化方案。仿真结果表明,通过联邦学习训练得到的语义通信模型在图像恢复以及下游语义任务两方面的性能均接近集中训练的,且所提的量化方案在几乎不影响模型性能的情况下减少模型更新过程75%的通信负载。

As a new communication paradigm,semantic communication(SemCom)aims to extract and transmit the semantic information from data source rather than the symbol-level precise transmission,thereby greatly enhancing communication eficiency and providing new insights for 6G mobile communication.In the era of intelligence of everything(loE),the models are need to achieve effective training,dynamic updating and timely distribution among multiple users,and the widely distributed data should be fully used without leaking the privacy.Therefore,an image SemCom system is first established based on autoencoders and federated learning(FL)is used to perform the distributed training,updating,and distribution of the SemCom system.In order to further reduce the significant transmission overhead caused by frequent model updates during the learning process,a quantization scheme is proposed for the model difference updates taking advantage of the small dynamic range of model differences(-0.1-0.1),including four steps of scaling up,rounding,limit,and scaling down.The simulation results show that the performance of the proposed FL-based SemCom system is comparable to the centralized one in terms of both image recovery and downstream tasks.Additionally,the proposed quantization scheme greatly reduces the transmission overhead by 75%during the model updating process with a negligible effect on the performance of SemCom system.