Deep Joint Source-Channel Coding Over Multi-User Access Channels With Random Activities

Abstract

Random access (RA) poses great challenges for modern wireless multi-user communication on how to handle the unknown and stochastic user access. Many existing RA schemes typically design codebooks for transmission with independent and statistically identical user access, but lack the adaptability to correlated user access behaviors and the resulting complex pattern dynamics. In this paper, we develop a deep joint source and channel coding (DeepJSCC) architecture accommodating non-stationary access patterns, enabling robust and efficient image transmission over Gaussian RA channels (RACs). Specifically, our method features these innovations: A DeepJSCC framework employing user-specific encoders and state-dependent decoders, allowing flexible non-orthogonal multiple access (NOMA) codebook construction for varying user access correlations; An intelligent discriminator that identifies real-time access conditions directly from received signals, enabling adaptive decoding and facilitating codebook adaptation to access pattern dynamics. An adversarial training mechanism incorporating the discrimination accuracy-aware loss and employing the balanced sampling strategy to generate high-quality NOMA codebooks. Experimental results show that our scheme outperforms existing deep learning-based MA approaches under unstable RAC scenarios.