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.
Figures & tables
Fig. 1: Overall framework of ReliaGuard. (a) Source-message encoding and REW produce the protected image; source decoding and localization recover its identifier and estimate the target edit region. (b) REW optimizes keyed responses in protected and VAE-reconstructed views under source-preservation and image-distortion constraints. (c) EBR combines shared dual-scale features with keyed responses and is trained on paired unpropagated and propagated observations. (d) BER refines the base prediction using background context, and NSR fuses native-, half-, and quarter-resolution features. Solid arrows denote forward computation; dashed arrows denote optimization or training supervision. Blue and red blocks indicate fixed and optimized components; reader colors refer to propagation training. Error-correcting coding (ECC) is used for the source message.
Method
COCO ( n=100 )
Generated ( n=100 )
ImageNet ( n=100 )
IoU ↑
F1 ↑
BAID↑
IoU ↑
F1 ↑
BAID↑
IoU ↑
F1 ↑
BAID↑
Full-output regeneration (FR)
Passive localization
TruFor
4.35
7.20
–
4.89
8.11
–
5.23
8.53
–
IML-ViT
1.39
2.45
–
1.43
2.56
–
2.04
3.54
–
Active protection
TABLE I: Localization and source-identifier recovery of active and passive methods across three image domains at fixed thresholds.
Fig. 2: FR localization examples for active methods across three image domains and four target areas. Each row uses the same source and target mask across methods; each method receives its own protected and edited image. The first column shows our received image. IoUs are evaluated at the fixed thresholds in Table I .
Fig. 3: SP localization examples for active methods across three image domains and four target areas. These cases differ from those in Fig. 2 ; IoUs use the same fixed-threshold convention.
Method
FR
SP
IoU ↑
F1 ↑
FPR ↓
IoU ↑
F1 ↑
FPR ↓
Fixed decision thresholds
WAM
23.54
35.18
60.76
77.60
83.85
0.97
StableGuard-C
16.73
26.69
94.61
98.16
99.00
0.23
OmniGuard
0.47
0.91
0.31
85.97
90.32
0.36
EditGuard
16.00
25.75
100.00
81.78
88.63
4.93
TABLE II: Localization at fixed thresholds and at thresholds calibrated to a nominal 1% FR background FPR.
Method
FR
SP
IoU ↑
F1 ↑
FPR ↓
BAID↑
ER ↑
IoU ↑
F1 ↑
FPR ↓
BAID↑
ER ↑
WAM
2.04
3.59
1.44
90.28
85.80
2.41
4.31
0.00
98.32
99.98
StableGuard-C
1.88
3.25
5.15
50.26
100.00
37.78
42.07
13.52
70.56
100.00
OmniGuard
1.64
2.99
1.46
N/A
N/A
28.17
31.39
2.98
N/A
N/A
EditGuard
0.63
1.20
0.63
53.79
100.00
0.40
0.75
2.91
75.11
100.00
GenPTW
0.53
1.02
0.90
91.78
100.00
0.00
0.00
0.00
92.98
100.00
TABLE III: Edit localization and source-identifier recovery under ordinary propagation.
Fig. 4: FR and SP localization of active methods under 18 image-processing settings, using fixed FR-calibrated thresholds. Heatmap values are IoUs rounded to whole percentages.
Method
COCO ( n=100 )
Generated ( n=100 )
ImageNet ( n=100 )
PSNR (dB) ↑
SSIM ↑
LPIPS ↓
PSNR (dB) ↑
SSIM ↑
LPIPS ↓
PSNR (dB) ↑
SSIM ↑
LPIPS ↓
WAM
38.32
0.9768
0.0348
37.80
0.9796
0.0238
39.32
0.9760
0.0444
StableGuard-C
25.87
0.7328
0.0733
31.46
0.9429
0.0256
31.01
0.8580
0.0517
OmniGuard
46.27
0.9935
0.0005
47.52
0.9948
0.0003
49.53
0.9951
0.0005
EditGuard
31.91
0.8370
0.0474
31.36
0.8554
0.0414
31.99
0.8141
0.0794
GenPTW
25.72
0.7174
0.0803
30.21
0.9147
0.0354
30.64
0.8406
0.0567
TABLE IV: Protected-image fidelity across three image domains.
Fig. 5: Protected images, local crops, and embedding residuals from the three domains. Examples are selected from the lowest-LPIPS quintile in each domain. Original and protected images use the same crop; absolute differences are amplified fourfold. Table IV reports quality over all 300 sources.
Fig. 6: Effects of target geometry. (a,b) FR IoU across target areas and shapes. (c,d) Inner-band FNR and outer-band FPR before and after propagation training, evaluated on unpropagated images. Solid and dashed curves denote FR and SP, respectively. Boundary bands are eight pixels wide.
Condition
Source recovery
Joint success
BAcode↑
CA ↑
@0.50 ↑
@0.75 ↑
FR
Unpropagated
90.58
99.67
96.33
90.78
Propagated
84.57
85.43
81.71
74.22
SP
Unpropagated
100.00
100.00
99.67
95.67
TABLE V: Source attribution and joint localization performance.
Method
FR
SP
Edit IoU ↑
Edit F1 ↑
Background IoU ↑
Edit IoU ↑
Edit F1 ↑
Background IoU ↑
APT
67.27
78.49
89.47
75.39
84.63
94.63
APT ∗
67.42
77.86
91.97
78.45
86.48
94.88
ReliaGuard
81.73
89.04
96.22
95.38
97.36
99.08
Before propagation training
83.11
90.15
96.59
95.44
97.30
99.12
TABLE VI: Localization under the APT SD1.5 protocol on 100 sources, using a fixed threshold of 0.5.
Fig. 7: Effect of spatial evidence and key correspondence on 60 sources. Reader weights and the decision threshold are fixed. Wrong-key results are averaged over five alternative keys.
Variant
FR
SP
IoU ↑
F1 ↑
FPR ↓
Inner FNR ↓
Outer FPR ↓
IoU ↑
F1 ↑
FPR ↓
Full architecture
83.17
90.16
1.15
31.93
22.26
93.33
96.30
0.28
Without BER
78.91
87.30
1.11
41.84
20.82
81.90
89.16
0.37
Without background contrast
83.08
90.08
1.15
31.97
22.42
93.32
96.29
0.29
Without NSR
82.31
89.62
1.18
33.57
22.84
91.50
95.21
0.37
Bilinear reassembly
83.18
90.16
1.15
31.88
22.28
93.38
96.33
0.28
TABLE VII: Structural ablations trained from the same localization initialization.
Fig. 8: Paired IoU differences between the full architecture and structural ablations. Positive values favor the full architecture. Error bars are 95% confidence intervals from 5,000 source-group bootstrap resamples on the 144-source ablation set, with the trained models fixed.
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.
Xiaodong Wu, Xiangman Li, Jianbing Ni +1
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.
Ordinary prompt-based editing can cause latent watermark detection to fail without explicitly targeting the watermark. We benchmark eight watermark methods against five editors across four generative backbones, four editing strengths, and five semantic categories, with edit-validity and threshold checks. Separating editing from seven subsequent distortions reveals that editing alone primarily distinguishes Tree-Ring, while added distortions expose a broader spectrum of detection survival. Sequential edits reveal a second hidden difference: score separation can decline while detection rates remain near their ceiling. Across methods, standardized clean score separation (d′) organizes composite-survival tiers, whereas spatial overlap adds little to predicting edit-only survival beyond clean detectability. Embedding-strength interventions in two methods link higher clean separation to higher post-edit separation. In HSTR, the margin contrast is positive, while the angular layout contrast at matched clean separation remains unresolved. Together, outcome decomposition and continuous separation expose differences hidden by aggregate TPR. Method tiers are stable under threshold recalibration at the main operating points and alternative composite weights. Clean d′ is thus a useful empirical diagnostic within this benchmark, with mixed transfer to unseen methods. Code and supporting artifacts are planned for a separate release.
Sung Ju Lee, Nam Ik Cho
Department of ECE & INMC Seoul National University Seoul, South Korea
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.