Federated Codebook for Multi - User Deep Source Coding

This paper investigates Deep Learning based Source Coding (DeepS C) with multiple users. While most of the existing works focus on a single pair of DeepSC transceivers, we consider multiple pairs of transceivers, each of which is modeled as a Vector Quantized Variational Autoencoder (VQ-VAE) architecture. Furthermore, in contrast to existing DeepSC works exploiting the trainability of encoders and decoders, in this work we focus on the trainability of codebooks. Inspired from this and Federated Learning (FL), we propose a novel DeepSC framework with federated codebook (FC- DeepSC) wherein each transceiver iteratively exchanges their codebooks during training, so as to construct an averaged global codebook that is downloaded by each transceiver. Simulation results corroborate that FC- DeepSC achieves faster convergence than DeepSC.