Channel-Aware Vector Quantization for Robust Semantic Communication on Discrete Channels

TL;DR

The paper tackles the challenge of deploying semantic communication on discrete digital infrastructures by addressing the disconnect between discrete semantic representations and digital channel constraints. It introduces VQJSCC, a fully digital joint source-channel coding framework based on VQ-VAE, and enhances it with Channel-Aware Vector Quantization (CAVQ) that leverages channel transition probabilities to guide codebook updates, plus a multi-codebook/subchannel architecture to decouple codebook size from modulation order. A channel-aware loss and end-to-end training strategy, including CVQ with EMA anchors and a hierarchical index-depth parameter l for rate control, enable robust performance across diverse modulation schemes on a discrete memoryless channel. Experiments on CIFAR-10 demonstrate constellation-aligned codebooks, higher codeword entropy, and superior PSNR compared to analog and state-of-the-art digital baselines, with clear advantages when codebook and modulation orders are aligned and flexible rate adaptation via index depth is possible. The work thus provides a practical, scalable path toward robust, digital-friendly semantic communication that integrates naturally with existing digital infrastructures and modulation schemes.

Abstract

Deep learning-based semantic communication has largely relied on analog or semi-digital transmission, which limits compatibility with modern digital communication infrastructures. Recent studies have employed vector quantization (VQ) to enable discrete semantic transmission, yet existing methods neglect channel state information during codebook optimization, leading to suboptimal robustness. To bridge this gap, we propose a channel-aware vector quantization (CAVQ) algorithm within a joint source-channel coding (JSCC) framework, termed VQJSCC, established on a discrete memoryless channel. In this framework, semantic features are discretized and directly mapped to modulation constellation symbols, while CAVQ integrates channel transition probabilities into the quantization process, aligning easily confused symbols with semantically similar codewords. A multi-codebook alignment mechanism is further introduced to handle mismatches between codebook order and modulation order by decomposing the transmission stream into multiple independently optimized subchannels. Experimental results demonstrate that VQJSCC effectively mitigates the digital cliff effect, achieves superior reconstruction quality across various modulation schemes, and outperforms state-of-the-art digital semantic communication baselines in both robustness and efficiency.

Figures (8)

• Figure 1: The proposed VQJSCC framework. The encoder $f_{\phi}$ maps image $\mathbf{x}$ to $\mathbf{Z}$; vector quantization with a shared codebook $\mathbf{M}$ yields $\mathbf{Z}_q$ and indices $\mathbf{y}$. The indices are bit-serialized and mapped to constellation indices $\mathbf{c}$ ($M=2^{m_c}$), transmitted over the wireless channel and demodulated to $\hat{\mathbf{c}}$, then pass through a discrete memoryless channel $P(\hat{c}\mid c)$. The decoder $g_{\theta}$ reconstructs $\hat{\mathbf{x}}$ from $\hat{\mathbf{y}}$.
• Figure 2: Subsequence decomposition and multi-codebook mapping address mismatched codebook and modulation orders. The index stream is divided based on the repeating bit-alignment period into separate subsequences. Each subsequence extracts a consistent bit slice pattern from sequential constellation symbols, creating an i.i.d. discrete memoryless subchannel via the digital channel.
• Figure 3: Heatmaps of pairwise distance matrices for modulation constellations (left), codebooks learned by the proposed CAVQ (middle), and codebooks from vanilla VQ-VAE (right) under 16-QAM, 64-QAM, and 256-QAM. Warmer colors indicate larger distances. CAVQ produces structured, constellation-like patterns aligned with modulation geometry, while vanilla VQ-VAE exhibits irregular distributions without channel awareness.
• Figure 4: Heatmaps of codeword distance matrices under mismatched codebook and modulation orders, generated by the multi-codebook algorithm with subsequent alignment via CAVQ. Each sub-codebook maintains structured patterns reflecting the estimated sub-channel transition probabilities.
• Figure 5: Codeword activation heatmaps under different codebook orders ($m_b=4,6,8$) corresponding to 16-QAM, 64-QAM, and 256-QAM. Each cell denotes the empirical usage frequency of a codeword, with darker colors indicating higher activation probability. Codeword positions are arranged according to the square QAM constellation.

Theorems & Definitions (2)

• Theorem 1
• Proposition 1