Blind Deep-Learning-Based Image Watermarking Robust Against Geometric Transformations

TL;DR

The paper addresses protecting images with watermarks that remain detectable after geometric transformations while remaining imperceptible to viewers. It extends the HiDDeN framework by introducing differentiable geometric noise layers and a differentiable JPEG approximation (JPEGdiff), coupled with an adversarial discriminator to improve imperceptibility. The proposed method demonstrates superior robustness to geometric attacks compared to the state of the art and maintains high invisibility, enabling practical deployment on consumer devices. This approach advances secure, blind watermarking by explicitly modeling geometric distortions within a differentiable end-to-end framework.

Abstract

Digital watermarking enables protection against copyright infringement of images. Although existing methods embed watermarks imperceptibly and demonstrate robustness against attacks, they typically lack resilience against geometric transformations. Therefore, this paper proposes a new watermarking method that is robust against geometric attacks. The proposed method is based on the existing HiDDeN architecture that uses deep learning for watermark encoding and decoding. We add new noise layers to this architecture, namely for a differentiable JPEG estimation, rotation, rescaling, translation, shearing and mirroring. We demonstrate that our method outperforms the state of the art when it comes to geometric robustness. In conclusion, the proposed method can be used to protect images when viewed on consumers' devices.

Figures (2)

• Figure 1: Architecture of HiDDeN hidden, consisting of an encoder and decoder, separated by noise layers. We include noise layers that simulate geometric transformations to provide robustness against geometric attacks.
• Figure 2: Robustness results.