DIF-WAN: watermark attack network based on diffusion model architecture
Journal of Supercomputing, cilt.82, sa.15, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 82 Sayı: 15
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s11227-026-08931-6
- Dergi Adı: Journal of Supercomputing
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Applied Science & Technology Source, Compendex, INSPEC, zbMATH, Academic Search Ultimate (EBSCO), Biomedical Reference Collection: Corporate Edition (EBSCO), Engineering Source (EBSCO), Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest)
- Anahtar Kelimeler: Bit error rate, Imperceptibility, Robust watermarking algorithm, Watermarking attack model
- Ankara Üniversitesi Adresli: Evet
Özet
Although traditional digital watermark attacks can disrupt watermark information, they often introduce significant image distortion during the attack process, making it difficult to balance watermark disruption capability and visual quality. To address this issue, a universal and imperceptible, DIF-WAN, is proposed in this paper. The diffusion process is constructed based on a mean-reverting stochastic differential equation, and the effective extraction of watermark information is reduced through progressive perturbation and reverse reconstruction of watermarked images. Since the multistep diffusion process requires repeated U-Net inference, high computational and memory demands are introduced. Therefore, LinearAttention is introduced in high-resolution feature stages to reduce the computational cost of global feature interaction, while multihead attention is adopted in the bottleneck layer to enhance global structure and contextual information modeling. Meanwhile, a Mask module is introduced to provide spatial guidance for watermark-sensitive regions, and a joint loss function is utilized to constrain the reconstruction quality of attacked images. Training data are constructed based on the DIV2K2017 dataset and the QPHFMs watermark algorithm, and the proposed method is validated on the DIV2K2017 and MSCOCO datasets as well as under different watermark mechanisms, including StableSignature, StegaStamp, LSB, and DCT. Experimental results show that, on the DIV2K2017 dataset, DIF-WAN achieves BER, PSNR, and SSIM values of 0.4441, 29.8771, and 0.7971, respectively. Computational efficiency experiments further demonstrate that, under the same Stage 3 configuration, LinearAttention reduces the diffusion inference time per image from 2.196698 s to 1.841926 s, corresponding to a reduction of approximately 16.15%. The results indicate that DIF-WAN maintains stable watermark disruption capability and image quality while reducing part of the computational overhead introduced by multistep diffusion through efficient attention design, providing a computational foundation for large-scale image watermark security analysis based on GPU/HPC.