Scholarlink

<p>JPEG steganography leverages discrete cosine transform (DCT) coefficients to embed secret messages, offering a practical medium for covert communication. However, the sparsity and block-based structure of DCT coefficients, combined with advances in steganalysis, pose significant challenges to feature extraction and anti-detection performance. To address these limitations, we propose AJSEAD, an adaptive JPEG steganographic framework that automatically learns embedding probabilities via a generative adversarial network (GAN). AJSEAD optimizes adversarial training for the unique characteristics of the DCT domain, thus providing a practical solution for secure and covert message embedding. Specifically, AJSEAD incorporates a U-shaped generator enhanced with dilated and large-kernel convolutions to improve feature extraction and employs upsampling layers to optimize decoding. For adversarial training, a dual-input steganalyzer enhances detection accuracy by jointly analyzing the spatial and DCT domains. Moreover, a refined loss function enhances AJSEAD’s resistance to detection. Extensive experiments verify the enhanced anti-detection capability and practicality of our proposed framework, demonstrating that it outperforms existing steganographic methods with various embedding strategies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AJSEAD: Adaptive JPEG steganography with enhanced anti-detection via generative adversarial network

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Yuxiang Peng,
  • Chong Fu,
  • Yu Zheng,
  • Yunjia Tian,
  • Guixing Cao,
  • Junxin Chen

摘要

JPEG steganography leverages discrete cosine transform (DCT) coefficients to embed secret messages, offering a practical medium for covert communication. However, the sparsity and block-based structure of DCT coefficients, combined with advances in steganalysis, pose significant challenges to feature extraction and anti-detection performance. To address these limitations, we propose AJSEAD, an adaptive JPEG steganographic framework that automatically learns embedding probabilities via a generative adversarial network (GAN). AJSEAD optimizes adversarial training for the unique characteristics of the DCT domain, thus providing a practical solution for secure and covert message embedding. Specifically, AJSEAD incorporates a U-shaped generator enhanced with dilated and large-kernel convolutions to improve feature extraction and employs upsampling layers to optimize decoding. For adversarial training, a dual-input steganalyzer enhances detection accuracy by jointly analyzing the spatial and DCT domains. Moreover, a refined loss function enhances AJSEAD’s resistance to detection. Extensive experiments verify the enhanced anti-detection capability and practicality of our proposed framework, demonstrating that it outperforms existing steganographic methods with various embedding strategies.