TRACEMARK: Editor-Aware Watermarking for User Attribution in Text-Guided Image Editing
Authors: Xiaodong Wu, Xiangman Li, Jianbing Ni, Rongxing Lu
Organizations: Department of Electrical and Computer Engineering, Queen’s University, Kingston, ON K7L 3N6, Canada. · School of Computing, Queen’s University, Kingston, ON K7L 3N6, Canada.
An image editing service records which user requests each edit, but the edited image itself does not contain this account information. Once the image is downloaded and shared, the service cannot identify the requesting user from the image alone. Embedding a user-specific watermark could enable such attribution, but semantic editing may erase the signal. We propose Tracemark, an editor-aware watermarking framework for identifying users from edited images. A shared encoder embeds a user-specific signal before editing, and a shared detector recovers the user by comparing features of the edited image with learned reference vectors. Both components are trained jointly through an editor with fixed parameters. User identification errors encourage the encoder to produce watermarks that survive editing, while fidelity losses limit changes to the input image and intended edit. Each image domain and editing instruction uses a separately trained model shared across registered users. Experiments on three image domains show that editor-aware training improves user recovery with both reference-vector matching and binary watermark decoding. Separately trained models reach 87.5% mean identification accuracy among 500 users. When tested with a different editor without retraining, the watermark modules still recover user information. These results demonstrate that learning watermarks through the editing process enables user attribution without modifying the editor.
Figures & tables
Fig. 1: System and watermark-integration paths for user attribution in text-guided image editing.
Fig. 2: Registered-user identification (a) and capacity margin from bit recovery (b) for three released 100 -bit watermarks.
Method
#Bits
IP2P
SDXL-Turbo
TrustMark [ 10 ]
100
0.495
0.009
RoSteALS [ 11 ]
100
0.975
0.249
VINE-R [ 3 ]
100
0.529
0.922
Robust-Wide [ 4 ]
64
0.917
0.648
TABLE I: Registered-user identification under InstructPix2Pix and Stable Diffusion XL (SDXL)-Turbo.
Fig. 3: The TraceMark training loop for user-conditioned watermark embedding and attribution through a fixed-weight instruction editor.
Editing channel
Domain
TraceMark
TrustMark
RoSteALS
VINE-R
Robust-Wide
InstructPix2Pix, training channel
CelebA-HQ
0.901
0.554
0.984
0.677
0.932
AFHQ-Dog
0.806
0.320
0.955
0.376
0.884
CUB-Bird
0.891
0.611
0.986
0.533
0.934
sd-turbo, s=0.2
CelebA-HQ
0.877
0.019
0.355
0.140
0.255
AFHQ-Dog
0.864
0.012
0.422
0.069
0.291
CUB-Bird
0.775
0.009
0.531
0.071
0.427
TABLE II: Cross-editor identification in three image domains using a common 256-pixel pipeline. Each domain contains 30 editing features.
Fig. 4: Four-channel mean identification accuracy with and without training through the editor, using two attribution methods in each domain.
Domain
Full
−Lcar
−Lcons
−Lout
CelebA-HQ
0.927
0.944
0.963
0.743
AFHQ-Dog
0.854
0.850
0.885
0.593
CUB-Bird
0.818
0.839
0.905
0.613
Mean
0.866
0.878
0.918
0.650
TABLE III: Post-edit identification accuracy after removing individual objective terms in three image domains.
Fig. 5: Identification accuracy as a function of carrier peak signal-to-noise ratio under InstructPix2Pix and sd-turbo editing.
Fig. 6: Mean identification accuracy of population-specific models on marked and unmarked edited images.
Training instruction
Marked mean
Marked range
Unmarked mean
CelebA-HQ
Add glasses
0.843
0.680–0.975
0.003
Make smiling
0.613
0.120–0.905
0.003
Make anime
0.858
0.735–0.920
0.023
AFHQ-Dog
Turn into cat
0.540
0.365–0.805
0.015
TABLE IV: Cross-instruction identification with fixed watermark models. Each source model is tested on three target instructions.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Method
PSNR ↑
SSIM ↑
LPIPS ↓
TraceMark
30.39
0.9566
0.0783
TrustMark
42.08
0.9883
0.0010
RoSteALS
29.07
0.8599
0.0348
VINE-R
36.53
0.9833
0.0089
Appendix
TABLE V: Carrier fidelity at each method’s native watermark strength over 90 semantic editing features.
β
0.35
0.50
0.70
1.00
1.40
Carrier PSNR (dB)
TraceMark
39.67
36.58
33.65
30.55
27.66
TrustMark
50.83
47.73
44.81
41.71
38.79
RoSteALS
36.90
33.81
30.88
27.79
24.89
VINE-R
44.47
41.37
38.45
35.35
32.79
Identification accuracy: InstructPix2Pix
Appendix
TABLE VI: Carrier PSNR and post-edit identification accuracy across watermark strengths.
Method
Pixels
Identity
IP2P
sd-turbo s=0.2
sd-turbo p2p-style s=0.3
ChordEdit
TraceMark
256
0.985
0.907
0.873
0.780
0.287
TrustMark
256
1.000
0.557
0.020
0.007
0.015
512
1.000
0.988
0.035
0.013
0.003
RoSteALS
256
1.000
0.995
0.420
0.138
0.083
512
1.000
0.990
0.425
0.133
0.083
VINE-R
256
1.000
0.605
0.100
0.062
0.025
Appendix
TABLE VII: Four-feature identification accuracy at 256- and 512-pixel carrier resolutions. IP2P denotes InstructPix2Pix.
Editing channel
256
512
Ratio
InstructPix2Pix
0.513
0.436
0.85
sd-turbo, s=0.2
0.474
0.370
0.78
sd-turbo, p2p-style s=0.3
0.506
0.400
0.79
ChordEdit
0.869
0.557
0.64
Appendix
TABLE VIII: Edit displacement at carrier resolutions of 256 and 512 pixels.
Fig. 7: TraceMark identification accuracy by image domain and editing channel.
Setting
Value
User embedding dim du
64
Encoder blocks / width
4 / 96
Carrier strength α
0.07
Detector backbone widths
64,96,128,192,256
Embedding dim d
128
Logit scale (init / cap)
16 / 100
Appendix
TABLE IX: Full TraceMark configuration for the paper-scale runs.
Prototype matching
100-bit decoding
Method effect
Feature
Exposed
Unexposed
Training
Exposed
Unexposed
Training
Exposed
Unexposed
Blond hair, CelebA-HQ
0.730
0.253
+0.477
0.665
0.445
+0.220
+0.065
−0.192
Angry, AFHQ-Dog
0.718
0.550
+0.167
0.520
0.320
+0.200
+0.198
+0.230
Crest, CUB-Bird
0.653
0.280
+0.373
0.628
0.390
+0.238
+0.025
−0.110
Mean
0.700
0.361
+0.339
0.604
0.385
+0.219
+0.096
−0.024
Appendix
TABLE X: Identification accuracy for the training-exposure and attribution-method control by image domain and editing channel.
Fig. 8: Identification accuracy of editor-exposed variants with prototype matching, 16-bit decoding, and 100-bit decoding across six channels.
Training instruction
Test instruction
Marked
Unmarked
CelebA-HQ
Add glasses
Make older
0.875
0.005
Make smiling
0.975
0.000
Make anime
0.680
0.005
Make smiling
Add glasses
0.905
0.000
Make older
0.815
0.005
Appendix
TABLE XI: Identification accuracy for all cross-instruction pairs, with 200 images per pair.
Registry size N
100
200
300
400
500
Training rounds
170
200
300
400
700
Workers
6
8 / 3
6–8
6
6
Glasses
0.967
0.891
0.916
0.782
0.870
Makeup
0.952
0.876
0.891
0.700
0.869
Red hair
0.939
0.754
0.851
0.718
0.900
Smiling
0.961
0.735
0.837
0.831
0.861
Appendix
TABLE XII: Identification accuracy and optimization budgets for four CelebA-HQ tasks with registries from 100 to 500 users.
This paper investigates a fundamental yet underexplored question: can watermarked images remain editable without compromising watermark integrity? We propose SafeMark, a framework for watermark-preserving text-guided image manipulation that explicitly integrates watermark integrity into the editing process. Specifically, SafeMark adds a thresholded watermark-decoding loss directly to the diffusion editor's training objective, fine-tuning the editor so that semantically valid edits also preserve the embedded watermark at the final output. This design admits a clean information-theoretic justification: maintaining high bit-accuracy on the edited image lower-bounds the mutual information that the editor channel preserves between watermark and edited output, the quantity that fundamentally controls watermark recoverability. SafeMark is compatible with differentiable diffusion-based editors, and requires no architectural modification. Extensive evaluations across multiple datasets, text-guided editing methods, and post-edit distortion settings demonstrate that SafeMark achieves high watermark bit accuracy across diverse editing settings while maintaining high-quality semantic edits, without sacrificing robustness to common post-edit distortions. These results demonstrate that semantic editability and watermark integrity are fundamentally compatible, enabling trustworthy image provenance in generative editing pipelines.
Text-guided diffusion editing raises disinformation concerns, making reliable image provenance essential. While watermarks are commonly used for this purpose, most methods carry a fixed ID that cannot explain what was changed and which prompt produced it. Furthermore, under open-source white-box access, attackers can easily locate and remove watermarks added as separate modules. Targeting this setting, we propose AngelFingerprint, a novel watermarking framework ensuring edit traceability, explainability, and white-box stealthiness. It integrates a LoRA into the diffusion model to embed the editing prompt's CLIP text embedding directly into the model's weights. An extractor then recovers this embedding from the image pixels alone. This semantic payload explains the edit, while the weight-integrated design makes it hard to detect and isolate even under full white-box access. Two techniques make this possible: a velocity-alignment anchor that preserves edit quality, and a specially designed frequency filter that keeps the watermark imperceptible yet recoverable and robust. On the MagicBrush dataset, our extractor achieves 86% top-1 accuracy in a 200-way prompt retrieval, versus 20% for prompt inversion.
Bo-Han Kung, Futa Waseda, Ching-Chun Chang +2
National Taiwan University, Taiwan · National Institute of Informatics, Japan
Generative image editing facilitates content creation but complicates source attribution and edit verification. Existing watermarking methods combine copyright protection with tamper localization, yet background reconstruction can weaken spatial evidence even when source identifiers remain recoverable. We propose ReliaGuard, a reconstruction-aware watermarking framework for source attribution and blind edit localization. The writer optimizes keyed feature responses in both the protected image and its variational autoencoder reconstruction under image-distortion and source-preservation constraints. Using only the received image and a secret key, the reader combines watermark responses with local image context through residual evidence refinement and predicts an edit mask through native-resolution-guided multiscale reading. It is trained on complete editing outputs, locally spliced counterparts, and processed versions of both. On an independent test set of 300 source images and 900 editing cases from three image domains, our method achieves edit-region intersection-over-union scores of 85.61% for complete outputs and 94.90% for locally spliced outputs at a threshold calibrated to a nominal 1% background false-positive rate on complete outputs. The complete-output result exceeds the strongest baseline by 45.02 percentage points, while decoded identifier-bit accuracy reaches 100% in both conditions. Across seven families of image processing, the corresponding localization scores are 81.01% and 86.49%, with identifier-bit accuracies of 92.76% and 92.79%, respectively. These results support source recovery and blind edit localization under background reconstruction and subsequent image processing.
Hailong Wang, Wuzheng Tan, Zixin Zhang +2
College of Cyber Security, Jinan University, Guangzhou 510632, China