Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Watermarking in Diffusion Model: Gaussian Shading with Exact Diffusion Inversion via Coupled Transformations (EDICT)

Krishna Panthi

TL;DR

This work addresses watermark fidelity in diffusion-based image generation by tackling the inexact inversion of Gaussian Shading with standard DDIM methods. It introduces Exact Diffusion Inversion via Coupled Transformations (EDICT) to derive exact inverse mappings using two coupled latents $x$ and $y$, which denoise each other in an alternating fashion during the forward pass and reintroduce noise in the reverse pass. The approach duplicates the watermark-infused latent, enabling reciprocal denoising and more precise watermark recovery, and claims to be the first integration of EDICT with Gaussian Shading. Experiments on Stable Diffusion 2.1 show a modest but statistically significant improvement in watermark recovery fidelity across several image-manipulation scenarios, without requiring model retraining; the method is roughly 2× slower and has some limitations under certain perturbations, suggesting avenues for speedups and broader applicability.

Abstract

This paper introduces a novel approach to enhance the performance of Gaussian Shading, a prevalent watermarking technique, by integrating the Exact Diffusion Inversion via Coupled Transformations (EDICT) framework. While Gaussian Shading traditionally embeds watermarks in a noise latent space, followed by iterative denoising for image generation and noise addition for watermark recovery, its inversion process is not exact, leading to potential watermark distortion. We propose to leverage EDICT's ability to derive exact inverse mappings to refine this process. Our method involves duplicating the watermark-infused noisy latent and employing a reciprocal, alternating denoising and noising scheme between the two latents, facilitated by EDICT. This allows for a more precise reconstruction of both the image and the embedded watermark. Empirical evaluation on standard datasets demonstrates that our integrated approach yields a slight, yet statistically significant improvement in watermark recovery fidelity. These results highlight the potential of EDICT to enhance existing diffusion-based watermarking techniques by providing a more accurate and robust inversion mechanism. To the best of our knowledge, this is the first work to explore the synergy between EDICT and Gaussian Shading for digital watermarking, opening new avenues for research in robust and high-fidelity watermark embedding and extraction.

Watermarking in Diffusion Model: Gaussian Shading with Exact Diffusion Inversion via Coupled Transformations (EDICT)

TL;DR

This work addresses watermark fidelity in diffusion-based image generation by tackling the inexact inversion of Gaussian Shading with standard DDIM methods. It introduces Exact Diffusion Inversion via Coupled Transformations (EDICT) to derive exact inverse mappings using two coupled latents and , which denoise each other in an alternating fashion during the forward pass and reintroduce noise in the reverse pass. The approach duplicates the watermark-infused latent, enabling reciprocal denoising and more precise watermark recovery, and claims to be the first integration of EDICT with Gaussian Shading. Experiments on Stable Diffusion 2.1 show a modest but statistically significant improvement in watermark recovery fidelity across several image-manipulation scenarios, without requiring model retraining; the method is roughly 2× slower and has some limitations under certain perturbations, suggesting avenues for speedups and broader applicability.

Abstract

This paper introduces a novel approach to enhance the performance of Gaussian Shading, a prevalent watermarking technique, by integrating the Exact Diffusion Inversion via Coupled Transformations (EDICT) framework. While Gaussian Shading traditionally embeds watermarks in a noise latent space, followed by iterative denoising for image generation and noise addition for watermark recovery, its inversion process is not exact, leading to potential watermark distortion. We propose to leverage EDICT's ability to derive exact inverse mappings to refine this process. Our method involves duplicating the watermark-infused noisy latent and employing a reciprocal, alternating denoising and noising scheme between the two latents, facilitated by EDICT. This allows for a more precise reconstruction of both the image and the embedded watermark. Empirical evaluation on standard datasets demonstrates that our integrated approach yields a slight, yet statistically significant improvement in watermark recovery fidelity. These results highlight the potential of EDICT to enhance existing diffusion-based watermarking techniques by providing a more accurate and robust inversion mechanism. To the best of our knowledge, this is the first work to explore the synergy between EDICT and Gaussian Shading for digital watermarking, opening new avenues for research in robust and high-fidelity watermark embedding and extraction.
Paper Structure (6 sections, 2 equations, 2 figures)

This paper contains 6 sections, 2 equations, 2 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Approach
  4. Experiments and Results
  5. Performance
  6. Conclusions

Figures (2)

  • Figure 1: Example of images manipulated. (a) Brightness, factor = 6 (Color Jitter), (b) Gaussian Blur, r=4 (GauBlur), (c) Gaussian Noise, $\mu = 0, \sigma = 0.05$ (GauNoise), (d) Identity, (e) JPEG, QF=25, (f) Median Filter, k=7 (MedBlur), (g) 60% area random crop, (h) 80% area random drop, (i) 25% Resize and restore (Resize), (j) Salt and Pepper Noise, p = 0.05 (S&PNoise)
  • Figure 2: The table shows the results obtained by testing our method against the baseline. It demonstrates that when EDICT is used, performance improves or remains consistent across all image manipulation methods, except when brightness is increased (ColorJitter) and when Salt and Pepper noise is added.