Towards Blind Watermarking: Combining Invertible and Non-invertible Mechanisms

Rui Ma, Mengxi Guo, Yi Hou, Fan Yang, Yuan Li, Huizhu Jia, Xiaodong Xie

Abstract

Blind watermarking provides powerful evidence for copyright protection, image authentication, and tampering identification. However, it remains a challenge to design a watermarking model with high imperceptibility and robustness against strong noise attacks. To resolve this issue, we present a framework Combining the Invertible and Non-invertible (CIN) mechanisms. The CIN is composed of the invertible part to achieve high imperceptibility and the non-invertible part to strengthen the robustness against strong noise attacks. For the invertible part, we develop a diffusion and extraction module (DEM) and a fusion and split module (FSM) to embed and extract watermarks symmetrically in an invertible way. For the non-invertible part, we introduce a non-invertible attention-based module (NIAM) and the noise-specific selection module (NSM) to solve the asymmetric extraction under a strong noise attack. Extensive experiments demonstrate that our framework outperforms the current state-of-the-art methods of imperceptibility and robustness significantly. Our framework can achieve an average of 99.99% accuracy and 67.66 dB PSNR under noise-free conditions, while 96.64% and 39.28 dB combined strong noise attacks.

Introduction

Digital watermarking is a classic cat-and-mouse game. You want to hide information (a watermark) inside an image so well that a human can't see it (imperceptibility), but you also need that information to survive if someone crops, compresses, or adds noise to the image (robustness). Traditionally, researchers have struggled with a "zero-sum" trade-off: the more robust you make the watermark, the more it degrades the image quality.

The rise of deep learning brought us Invertible Neural Networks (INNs). INNs are elegant because they are mathematically reversible; the same parameters used to "hide" the watermark are used to "extract" it. This leads to incredible imperceptibility because the network learns a nearly lossless mapping. However, INNs are extremely sensitive to non-differentiable noise, specifically lossy quantization like JPEG compression. Because JPEG removes information, the mathematical "reversibility" of the INN breaks down, and the watermark extraction fails. The authors set out to bridge this gap by combining the strengths of invertible and non-invertible architectures.

The CIN Framework

The proposed CIN (Combining Invertible and Non-invertible mechanisms) framework splits the watermarking task into two specialized paths:

1. The Invertible Path (The "Clean" Specialist)

This path focuses on high-fidelity embedding and extraction in noise-free scenarios.

  • Diffusion and Extraction Module (DEM): Spreads the watermark bits across the spatial dimensions of the image.
  • Fusion and Split Module (FSM): Uses a Haar Transform to embed the watermark into the frequency domain. This allows the model to hide data in high-frequency details where it's less visible to humans.

2. The Non-invertible Path (The "Noise" Specialist)

This path handles extraction when the image has been degraded by lossy processes.

  • Non-invertible Attention-based Module (NIAM): A parallel decoder based on SENet (Squeeze-and-Excitation Network). It is trained specifically to "guess" the watermark from images mangled by lossy compression.
  • Noise-specific Selection Module (NSM): A smart "switch" that routes the extraction to the NIAM if it detects JPEG-style artifacts; otherwise, it uses the standard invertible path.

Experimental Results

The researchers evaluated CIN on major benchmarks like COCO and DIV2K against 14 types of distortions. The results demonstrate a superior trade-off between quality and survival.

Comparison of CIN against state-of-the-art MBRS under noise-free and JPEG noise conditions. CIN achieves significantly higher PSNR (67.66 dB) and accuracy (99.99%) in clean conditions.

In noise-free conditions, the model achieved a PSNR of 67.66 dB. For context, anything above 40 dB is usually considered indistinguishable to the human eye; 67 dB is virtually perfect. Under a "combined noise" attack, the model maintained 96.64% extraction accuracy with a PSNR of 39.28 dB. CIN outperformed existing state-of-the-art models like MBRS and HiDDeN. For example, against JPEG compression (Q=50), CIN achieved 99.24% accuracy, while many previous models struggled to break 90%.

Conclusion

By combining the high-fidelity embedding of INNs with a specialized non-invertible attention module, the CIN framework achieves near-perfect invisibility while maintaining high extraction accuracy even under harsh JPEG compression and noise attacks. Future work may involve more nuanced "soft" selection mechanisms or a fusion of multiple decoders to handle even more complex, overlapping attacks.

Code Availability: https://github.com/RM1110/CIN

Summary updated

Abstract

Blind watermarking provides powerful evidence for copyright protection, image authentication, and tampering identification.However, it remains a challenge to design a watermarking model with high imperceptibility and robustness against strong noise attacks. To resolve this issue, we present a framework Combining the Invertible and Non-invertible (CIN) mechanisms. The CIN is composed of the invertible part to achieve high imperceptibility and the non-invertible part to strengthen the robustness against strong noise attacks. For the invertible part, we develop a diffusion and extraction module (DEM) and a fusion and split module (FSM) to embed and extract watermarks symmetrically in an invertible way. For the non-invertible part, we introduce a non-invertible attention-based module (NIAM) and the noise-specific selection module (NSM) to solve the asymmetric extraction under a strong noise attack. Extensive experiments demonstrate that our framework outperforms the current state-of-the-art methods of imperceptibility and robustness significantly. Our framework can achieve an average of 99.99% accuracy and 67.66 dB PSNR under noise-free conditions, while 96.64% and 39.28 dB combined strong noise attacks. The code will be available in https://github.com/RM1110/CIN.
BibTeX
@inproceedings{ma2022towards,
  author = {Rui Ma and Mengxi Guo and Yi Hou and Fan Yang and Yuan Li and Huizhu Jia and Xiaodong Xie},
  title = {{Towards Blind Watermarking: Combining Invertible and Non-invertible Mechanisms}},
  booktitle = {ACM Multimedia},
  year = {2022},
  url = {https://openreview.net/forum?id=loaSjo0kbrH},
}
Figure