Se-HiLo: Noise-Resilient Semantic Communication with High-and-Low Frequency Decomposition

2025-03-10

AI Summary

To address unreliable reconstruction caused by unpredictable semantic noise—such as ambiguity and distortion—in semantic communication, this paper proposes a high-low frequency decoupled noise-robust semantic transmission framework. Methodologically, it introduces the first frequency-domain disentanglement of image semantic representations into structure-dominant low-frequency and detail-dominant high-frequency components, each mapped independently into finite scalar quantization (FSQ) spaces. Crucially, FSQ constraints replace conventional adversarial training, yielding more efficient and stable noise robustness. Built upon a Transformer architecture, the framework enables end-to-end joint optimization of semantic encoding and decoding. Experiments demonstrate that, across diverse noisy channels, it achieves significantly higher PSNR and SSIM than state-of-the-art methods, with superior semantic fidelity, strong generalization, and reduced training overhead.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Machine Learning: Adversarial Learning & Robustness
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs

Abstract

Semantic communication has emerged as a transformative paradigm in next-generation communication systems, leveraging advanced artificial intelligence (AI) models to extract and transmit semantic representations for efficient information exchange. Nevertheless, the presence of unpredictable semantic noise, such as ambiguity and distortions in transmitted representations, often undermines the reliability of received information. Conventional approaches primarily adopt adversarial training with noise injection to mitigate the adverse effects of noise. However, such methods exhibit limited adaptability to varying noise levels and impose additional computational overhead during model training. To address these challenges, this paper proposes Noise-Resilient extbf{Se}mantic Communication with extbf{Hi}gh-and- extbf{Lo}w Frequency Decomposition (Se-HiLo) for image transmission. The proposed Se-HiLo incorporates a Finite Scalar Quantization (FSQ) based noise-resilient module, which bypasses adversarial training by enforcing encoded representations within predefined spaces to enhance noise resilience. While FSQ improves robustness, it compromise representational diversity. To alleviate this trade-off, we adopt a transformer-based high-and-low frequency decomposition module that decouples image representations into high-and-low frequency components, mapping them into separate FSQ representation spaces to preserve representational diversity. Extensive experiments demonstrate that Se-HiLo achieves superior noise resilience and ensures accurate semantic communication across diverse noise environments.

Problem

• Addresses unpredictability of semantic noise in communication.
• Proposes Se-HiLo for noise-resilient image transmission.
• Enhances noise resilience and preserves representational diversity.

Innovation

• Finite Scalar Quantization for noise resilience
• High-and-low frequency decomposition module
• Transformer-based image representation mapping