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

Krishna Panthi

School of Computing, Clemson University
kpanthi@clemson.edu

Images rendered using EDICT vs. without using EDICT with same prompts.

Using EDICTUsing EDICT Image 1Using EDICT Image 2Using EDICT Image 3Using EDICT Image 4Using EDICT Image 5
Not using EDICTNo EDICT Image 1No EDICT Image 2No EDICT Image 3No EDICT Image 3No EDICT Image 3
Figure 1: Close inspection shows that the images generated using EDICT are of higher quality.

Summary

This research implements Exact Diffusion Inversion via Coupled Transformation (EDICT) [1] with the Gaussian Shading [2] watermarking technique for stable diffusion models. The results show a slight improvement in the performance of Gaussian Shading. The implementation is tested on manipulated images after watermarking, and results in the table below shows better performance for most image manipulation methods, except for ColorJitter and Salt & Pepper Noise. For more details, please refer to the paper.

Results

Table 1. The following 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 & Pepper noise is added.
  • TPR: True Precision Rate with fixed false positive rate of 1e-6
  • mean_acc: Mean bit accuracy
  • std_acc: Standard deviation on bit accuracy

Image ManipulationsTPR_detection ↑TPR_traceability ↑mean_acc (higher is better) ↑std_acc (smaller is better) ↓
DefaultEDICTDefaultEDICTDefaultEDICTDefaultEDICT
ColorJitter (f = 6)0.9790.9590.9570.9340.9520.9390.0920.107
GauBlur (r=4)11110.9850.9880.0200.015
GauNoise0.9950.9980.9860.9950.9540.9710.0700.053
Identity1111110.00.0
Jpeg (QF 25)11110.9870.9870.0310.032
MedBlur (k=7)1111110.0050.002
RandomCrop (60%)11110.9750.9760.0170.013
RandomDrop (80%)110.99810.9660.9690.0290.013
Resize (25%)11110.99910.0100.003
S&PNoise (p=0.05)110.99910.9350.9340.0710.067