Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Channel Coding for Unequal Error Protection in Digital Semantic Communication

Seonjung Kim, Yongjeong Oh, Yongjune Kim, Namyoon Lee, Yo-Seb Jeon

Abstract

Semantic communication is an emerging paradigm that prioritizes transmitting task-relevant information over accurately delivering raw data bits. In this paper, we address an unequal error protection (UEP) problem in digital semantic communication, where bits of higher semantic importance require stronger protection. To quantify bit-level importance, we leverage bit-flip probabilities of semantic bits as target error protection levels, which are jointly learned with semantic encoder and decoder. We propose two novel channel coding frameworks aimed at minimizing the total blocklength while satisfying UEP constraints. First, we develop a bit-level UEP framework based on repetition coding, in which the repetition number for each bit is optimized to precisely meet its target bit-flip probability. Second, we introduce a block-level UEP framework utilizing modern channel codes, where semantic bits with similar target bit-flip probabilities are grouped to exploit coding gains. Within this framework, we propose a bit-grouping algorithm guided by finite blocklength capacity analysis. Simulation results conducted on image transmission tasks confirm that the proposed frameworks significantly outperform conventional approaches, yielding substantial improvements in both task performance and transmission efficiency.

Channel Coding for Unequal Error Protection in Digital Semantic Communication

Abstract

Semantic communication is an emerging paradigm that prioritizes transmitting task-relevant information over accurately delivering raw data bits. In this paper, we address an unequal error protection (UEP) problem in digital semantic communication, where bits of higher semantic importance require stronger protection. To quantify bit-level importance, we leverage bit-flip probabilities of semantic bits as target error protection levels, which are jointly learned with semantic encoder and decoder. We propose two novel channel coding frameworks aimed at minimizing the total blocklength while satisfying UEP constraints. First, we develop a bit-level UEP framework based on repetition coding, in which the repetition number for each bit is optimized to precisely meet its target bit-flip probability. Second, we introduce a block-level UEP framework utilizing modern channel codes, where semantic bits with similar target bit-flip probabilities are grouped to exploit coding gains. Within this framework, we propose a bit-grouping algorithm guided by finite blocklength capacity analysis. Simulation results conducted on image transmission tasks confirm that the proposed frameworks significantly outperform conventional approaches, yielding substantial improvements in both task performance and transmission efficiency.

Paper Structure

This paper contains 16 sections, 3 theorems, 37 equations, 12 figures, 3 tables, 5 algorithms.

Table of Contents

  1. Introduction
  2. System Model and Preliminary
  3. Digital Semantic Communication Scenario
  4. Error Protection Level Extraction via End-to-End Training
  5. Channel Coding for Bit-Level UEP in Semantic Communication
  6. Problem Formulation for Bit-Level UEP
  7. Channel Coding for Bit-Level UEP
  8. Channel Coding for Block-Level UEP in Semantic Communication
  9. Problem Formulation for Block-Level UEP
  10. Bit Grouping Algorithm
  11. Channel Coding for Block-Level UEP
  12. Simulation Results
  13. Conclusion
  14. Proof of Proposition 1
  15. Proof of Proposition 2
  16. ...and 1 more sections

Key Result

Proposition 1

Suppose that $r^{\rm (ub)}$ is defined as in eq:initial_points, given a target bit-flip probability $\mu$ and a coded-bit-flip probability $\epsilon$. Then, the following inequality holds:

Figures (12)

  • Figure 1: An illustration of a digital semantic communication system with the proposed UEP framework.
  • Figure 2: An illustration of the training procedure of the end-to-end training method in oh2025digital along with the resulting bit-flip probabilities for the MNIST dataset when $\lambda=10^{-4}$.
  • Figure 3: Visualization of reconstructed CIFAR-10 images and corresponding PSNR and SSIM values when flipping the $i$-th segment of sorted semantic bits.
  • Figure 4: An illustration of the channel encoding process in the proposed block-wise UEP framework.
  • Figure 5: Comparison of PSNR versus total blocklength for various channel coding frameworks using the CIFAR-10 dataset with $\lambda=10^{-4}$ ($K=12288$).
  • ...and 7 more figures

Theorems & Definitions (3)

  • Proposition 1
  • Proposition 2
  • Theorem 1