Scholarlink

<p>Digital image watermarking technology has been widely used in the fields of image copyright protection and content authentication. A well-performed watermarking algorithm has a certain robustness while ensuring image quality. Attack-based robustness testing is an effective evaluation method for digital watermarking algorithms. Therefore, inspired by self-supervised denoising and residual learning, we propose a digital image watermarking attack scheme based on residual and machine learning self-supervised training. It trains the model without carrier images and affects the embedded watermark information. We will evaluate the effectiveness of this attack scheme using DCT-based and DWT-based watermarking algorithms on standard images and compare it with other watermarking attack schemes. The results of the comparison experiments show that, for both DCT-based and DWT-based watermarking algorithms, the proposed scheme improves the PSNR on BSD dataset by 16.48% and 17.26%, respectively. In particular, the UNet-based attack scheme achieves very high PSNR and SSIM. In addition, we also compute the residual maps of the attacked image with the original watermarked image, and the results show that the proposed attack method has less impact on the details of the image, which indicates that it is difficult to detect the signs of the attacked from the attacked image.</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

Attacking the Digital Watermark of Watermarked Images Through Residual Learning and Self-supervised Training

  • 期刊论文

  • 发表日期：

    2025年7月28日

查看全文
  • Zigang Chen,
  • Deyang Xie,
  • Zhenghao Liu,
  • Tao Leng,
  • Yuhong Liu,
  • Haihua Zhu

摘要

Digital image watermarking technology has been widely used in the fields of image copyright protection and content authentication. A well-performed watermarking algorithm has a certain robustness while ensuring image quality. Attack-based robustness testing is an effective evaluation method for digital watermarking algorithms. Therefore, inspired by self-supervised denoising and residual learning, we propose a digital image watermarking attack scheme based on residual and machine learning self-supervised training. It trains the model without carrier images and affects the embedded watermark information. We will evaluate the effectiveness of this attack scheme using DCT-based and DWT-based watermarking algorithms on standard images and compare it with other watermarking attack schemes. The results of the comparison experiments show that, for both DCT-based and DWT-based watermarking algorithms, the proposed scheme improves the PSNR on BSD dataset by 16.48% and 17.26%, respectively. In particular, the UNet-based attack scheme achieves very high PSNR and SSIM. In addition, we also compute the residual maps of the attacked image with the original watermarked image, and the results show that the proposed attack method has less impact on the details of the image, which indicates that it is difficult to detect the signs of the attacked from the attacked image.