Scholarlink

Modern AI-based image compression methods significantly outperform traditional techniques. However, they remain vulnerable to adversarial attacks designed either to degrade the quality of reconstruction by injecting imperceptible noise into the original image. In this work, we further investigate this issue and propose using the Tanh-Arctanh function to generate perturbed images, which allows noise to be applied in a non-linear manner relative to pixel brightness. We demonstrate the effectiveness of this approach against those targeting maximal distortion on the Kodak dataset showing significant improvements in PSNR and SSIM compared to traditional methods.

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

Robustness of AI Image Compression: Non-linear Perturbation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Arofenitra Rarivonjy,
  • Anton Bibin,
  • Razan Dibo,
  • Aleksandr Kolomeitsev,
  • Anh-Huy Phan,
  • Ivan Oseledets

摘要

Modern AI-based image compression methods significantly outperform traditional techniques. However, they remain vulnerable to adversarial attacks designed either to degrade the quality of reconstruction by injecting imperceptible noise into the original image. In this work, we further investigate this issue and propose using the Tanh-Arctanh function to generate perturbed images, which allows noise to be applied in a non-linear manner relative to pixel brightness. We demonstrate the effectiveness of this approach against those targeting maximal distortion on the Kodak dataset showing significant improvements in PSNR and SSIM compared to traditional methods.