Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adversarial Shallow Watermarking

Guobiao Li, Lei Tan, Yuliang Xue, Gaozhi Liu, Zhenxing Qian, Sheng Li, Xinpeng Zhang

TL;DR

This paper tackles the robustness gap of learning-based deep watermarking (LDW) against unknown distortions by introducing Adversarial Shallow Watermarking (ASW). ASW fixes a randomly parameterized shallow decoder and performs adversarial embedding to produce watermarked images that reliably trigger the decoder, without training encoders or modeling distortions. Empirical results show ASW delivers competitive performance on known distortions and superior robustness to unknown distortions compared with state-of-the-art LDW methods, while remaining training-free and computationally efficient at extraction. The work provides a new perspective on watermark design by leveraging the distortion-insensitivity of shallow networks, offering practical robustness for real-world media protection.

Abstract

Recent advances in digital watermarking make use of deep neural networks for message embedding and extraction. They typically follow the ``encoder-noise layer-decoder''-based architecture. By deliberately establishing a differentiable noise layer to simulate the distortion of the watermarked signal, they jointly train the deep encoder and decoder to fit the noise layer to guarantee robustness. As a result, they are usually weak against unknown distortions that are not used in their training pipeline. In this paper, we propose a novel watermarking framework to resist unknown distortions, namely Adversarial Shallow Watermarking (ASW). ASW utilizes only a shallow decoder that is randomly parameterized and designed to be insensitive to distortions for watermarking extraction. During the watermark embedding, ASW freezes the shallow decoder and adversarially optimizes a host image until its updated version (i.e., the watermarked image) stably triggers the shallow decoder to output the watermark message. During the watermark extraction, it accurately recovers the message from the watermarked image by leveraging the insensitive nature of the shallow decoder against arbitrary distortions. Our ASW is training-free, encoder-free, and noise layer-free. Experiments indicate that the watermarked images created by ASW have strong robustness against various unknown distortions. Compared to the existing ``encoder-noise layer-decoder'' approaches, ASW achieves comparable results on known distortions and better robustness on unknown distortions.

Adversarial Shallow Watermarking

TL;DR

This paper tackles the robustness gap of learning-based deep watermarking (LDW) against unknown distortions by introducing Adversarial Shallow Watermarking (ASW). ASW fixes a randomly parameterized shallow decoder and performs adversarial embedding to produce watermarked images that reliably trigger the decoder, without training encoders or modeling distortions. Empirical results show ASW delivers competitive performance on known distortions and superior robustness to unknown distortions compared with state-of-the-art LDW methods, while remaining training-free and computationally efficient at extraction. The work provides a new perspective on watermark design by leveraging the distortion-insensitivity of shallow networks, offering practical robustness for real-world media protection.

Abstract

Recent advances in digital watermarking make use of deep neural networks for message embedding and extraction. They typically follow the ``encoder-noise layer-decoder''-based architecture. By deliberately establishing a differentiable noise layer to simulate the distortion of the watermarked signal, they jointly train the deep encoder and decoder to fit the noise layer to guarantee robustness. As a result, they are usually weak against unknown distortions that are not used in their training pipeline. In this paper, we propose a novel watermarking framework to resist unknown distortions, namely Adversarial Shallow Watermarking (ASW). ASW utilizes only a shallow decoder that is randomly parameterized and designed to be insensitive to distortions for watermarking extraction. During the watermark embedding, ASW freezes the shallow decoder and adversarially optimizes a host image until its updated version (i.e., the watermarked image) stably triggers the shallow decoder to output the watermark message. During the watermark extraction, it accurately recovers the message from the watermarked image by leveraging the insensitive nature of the shallow decoder against arbitrary distortions. Our ASW is training-free, encoder-free, and noise layer-free. Experiments indicate that the watermarked images created by ASW have strong robustness against various unknown distortions. Compared to the existing ``encoder-noise layer-decoder'' approaches, ASW achieves comparable results on known distortions and better robustness on unknown distortions.
Paper Structure (14 sections, 6 equations, 5 figures, 4 tables, 1 algorithm)

This paper contains 14 sections, 6 equations, 5 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related works
  3. Learning-based Deep Watermarking
  4. Adversarial Perturbation
  5. Analysis of the LDW Framework
  6. The Proposed Method
  7. The Shallow Decoder
  8. Adversarial Watermark Embedding
  9. Experiments
  10. Visual Quality and BER.
  11. Robustness
  12. Computational Efficiency
  13. Ablation Study
  14. Conclusion

Figures (5)

  • Figure 1: Illustration of the existing LDW framework and the proposed ASW framework, with the solid line representing the case where watermarked images undergo known distortions, and the dashed line representing the case where watermarked images undergo unknown distortions.
  • Figure 2: The BER (%) of the extracted watermark message for HiDDeN-decoders of different depths under $n^{+}$ or $n^{-}$ distortions.
  • Figure 3: Architecture of the shallow decoder, with $s$ and $t$ representing the stride and output length, respectively.
  • Figure 4: BER (%) of the extracted watermark message for the compared LDW methods and the proposed ASW under different distortions. The solid and dashed lines represent the results of these methods under known and unknown distortions, respectively. The distortion strength increases along the horizontal axis from left to right in all subfigures.
  • Figure 5: Visualization of the watermarked images generated using the compared LDW methods and the proposed ASW.