Multi-Scale Quantized Joint Source-Channel Coding for Energy-Efficient Federated Semantic Communication in Heterogeneous Edge Networks

This paper proposes Energy-Efficient Semantic Communication (EESC) for the Federated Learning (FL)-enabled Semantic Communication system, addressing the high energy consumption and data privacy issues in traditional communication systems. This system uses Multi-Scale Joint Source-Channel Coding (MSJSCC) for efficient extraction and transmission of image semantic information. Each device trains a Convolutional Neural Network (CNN) with low-bit weights to reduce energy consumption and only upload the quantized values of changes in training results to the edge server, further reducing the energy expenditure during communication. Experimental results on the CIFAR-10 dataset show that EESC outperforms typical Joint Source-Channel Coding (JSCC) methods, improving PSNR by over 4 dB and reducing energy consumption by 31.46% compared to full-precision methods. Additionally, EESC demonstrates improved robustness under varying Signal-to-Noise Ratios (SNRs).