.png){kind=icon}
.png){kind=icon}
Return
An attack-resilient Unet watermarking framework for copyright protection via adaptive weighting and resolution recovery
.png){kind=icon}
PRE
AI
Z
N
J
L
DOI：10.1016/j.sigpro.2026.110609.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In the era of information security, multimedia images propagate through networks and often face diverse attacks and malicious tampering. Ensuring the trustworthiness and integrity of such images is critical. However, many deep watermarking methods, including Unet-based designs, rely on fixed skip connection fusion strategies, which can cause redundant feature propagation and detail loss. To address this gap, an attack-resilient weighted Unet-based watermarking framework is proposed. It replaces the concatenated-based skip connection with a weighted-based mode to alleviate redundant propagation and cross-layer fusion imbalance. In this framework, an adaptive weighting module is proposed to learn skip-connection weights automatically, avoiding hand-crafted weights that cannot adapt to different image contents. Furthermore, a residual interaction spatial feature transform module predicts affine modulation parameters from down-sampled host priors to modulate the up-sampling features, restoring detailed information lost during feature fusion. Experiments on COCO show that the proposed scheme achieves strong imperceptibility with PSNR of 35.71 dB, SSIM of 0.9835, and LPIPS of 0.004. Moreover, it maintains high extraction robustness under common distortions, reaching 99.67% bit accuracy under JPEG compression with a quality factor of 40, outperforming several state-of-the-art deep watermarking methods.
Keywords:
Information security
Watermarking method
Skip connection
Weighted Unet
Adaptive weighting factor
Detail restoration
Journal
IF:
3.6
Papers:
9.8K
Citations:
1.7W
Organization
Cited Papers