ROBIN++: Unified Copyright Protection and Tamper Localization for Diffusion Models via Dual-Domain Synergistic Watermarking
Authors: Huayang Huang, Siqi Zeng, Qian Wang, Bo Du, Yu Wu
Organizations: School of Computer Science, National Engineering Research Center for Multimedia Software, and Hubei Key Laboratory of Multimedia and Network Communication Engineering, Institute of Artificial Intelligence, Wuhan University, Wuhan, China · School of Cyber Science and Engineering, Wuhan University, Wuhan, China · School of Artificial Intelligence, National Engineering Research Center for Multimedia Software, and Hubei Key Laboratory of Multimedia and Network Communication Engineering, Institute of Artificial Intelligence, Wuhan University, Wuhan, China
Digital watermarking provides a promising solution for verifying the provenance and integrity of generated content. However, existing methods often suffer from an inherent trade-off between copyright robustness and localization sensitivity due to the entanglement of conflicting signals within a single domain. In this paper, we propose ROBIN++, a paradigm that shifts from single-domain entanglement to a Dual-Domain Synergistic design. Our key insight is to explicitly decouple robustness and fragility during injection by assigning them to distinct domains, while re-coupling them for mutual reinforcement during verification. Specifically, we introduce an Adversarial Optimization Algorithm to inject a frequency-domain copyright watermark during generation, maximizing extraction robustness without visual artifacts. Complementarily, a Content-Aware Perturbation generator is integrated into the latent space to inject the spatial-domain localization watermark, ensuring sensitivity to spatial tampering with minimal interference to the copyright watermark. During verification, we propose a dual-branch synergistic detector that facilitates Frequency-to-Spatial Synergy, where robust spectral priors act as global anchors for local anomaly detection. Conversely, a Spatial-to-Frequency Refinement strategy is employed to rectify corrupted latent features using predicted tamper masks. Extensive experiments show that ROBIN++ significantly outperforms existing approaches in both copyright recovery and tamper localization accuracy, effectively mitigating the long-standing trade-off between the two tasks. Code is available at https://github.com/Hannah1102/ROBIN.
Figures & tables
Fig. 1: Overview of the unified verification framework. Dual-functional watermarking framework embeds complementary signals for copyright ( wid ) and tamper localization ( wloc ). Upon receiving a potentially degraded or tampered image, the system performs dual extraction to classify the image into one of three states: Authentic & Intact, Authentic & Tampered, or Unverified.
Fig. 2: Robustness versus tamper localization precision of ROBIN++ and existing methods [ 10 , 11 , 13 , 12 , 14 ] . The results are evaluated based on the average performance across various common global degradations and local tampering operations. While existing methods often face a trade-off between robustness and localization, ROBIN++ achieves superior performance in both dimensions.
Fig. 3: Conceptual comparison between existing methods and our framework. (Left) Single-domain injection entangles robustness and localization, resulting in an inherent trade-off. (Right) Our proposal decouples the two objectives by dual-domain injection and enables mutual reinforcement during verification.
Fig. 4: The overall framework of ROBIN++. Our approach decouples robustness and fragility through a dual-domain synergistic strategy. (Left) Dual-Domain Injection: During diffusion generation, the robust copyright watermark ( wid∗ ) is embedded into the frequency domain at the detail formation stage ( tinj∈[200,300] ) via Adversarial Optimization, where a learnable prompt guidance ( wp∗ ) conceals visual artifacts. Complementarily, the fragile localization signal ( wloc ) is injected into the latent ( z0∗ ) at the late stage by a Content-Aware Perturbation Generator ( Gloc ). (Bottom) Self-Supervised Training: To ensure generalizability without annotated datasets, we employ a self-supervised splicing protocol. Gloc and the Tamper Detector ( Dloc ) are jointly optimized end-to-end by synthesizing “hard negative” spliced samples ( xspliced ) on the fly. (Right) Bidirectional Synergistic Extraction: The watermark extraction features two interacting processes. Copyright is authenticated via diffusion inversion to retrieve frequency signals, while tampering is localized by the dual-branch detector. Crucially, the robust spectral prior aids the detector in distinguishing benign distortions from malicious tampering, while the predicted spatial mask, in turn, filters out corrupted regions to refine spectral verification, thereby achieving synergistic performance.
TABLE I: Comparison of robustness against image degradations (Bit Acc/AUC), Inpainting Localization (average of F1, AUC, and IoU), and image quality (PickScore) on generative-model-based inpainting datasets [ 50 , 54 , 55 , 56 , 57 ] . The best and second-best results are marked in bold and underlined , respectively.
Method
Imperceptible ↑
Frequency-domain Attacks ↑
Generative-AI-based Removal ↑
Perturbation-removal ↑
Black-box Forgery ↓
Adv. Noise [ 59 ]
Freq. Noise
Low Pass
Freq. Smooth
Latent Remv. [ 60 ]
Diffusion Rec. [ 61 ]
VAE Rec. [ 61 ]
Rinse2x [ 61 ]
Pert. Defeat [ 62 ]
Latent Forg. [ 60 ]
EditGuard [ 10 ]
0.998 / 1.000
0.999 / 1.000
0.501 / 0.483
0.997 / 1.000
0.505 / 0.501
0.488 / 0.442
0.501 / 0.500
0.495 / 0.463
1.000 / 1.000
0.458
OmniGuard [ 11 ]
0.999 / 1.000
1.000 / 1.000
1.000 / 1.000
0.918 / 1.000
0.507 / 0.542
0.871 / 1.000
0.998 / 1.000
0.750 / 1.000
1.000 / 1.000
0.991
WAM [ 13 ]
0.998 / 1.000
0.999 / 1.000
0.999 / 1.000
0.993 / 1.000
0.496 / 0.500
0.716 / 0.857
0.705 / 0.922
0.639 / 0.780
0.999 / 1.000
0.500
TAG-WM [ 12 ]
0.973 / 1.000
0.999 / 1.000
0.969 / 1.000
0.978 / 1.000
0.820 / 0.981
0.973 / 1.000
0.953 / 1.000
0.936 / 1.000
0.999 / 1.000
0.614
StableGuard [ 14 ]
1.000 / 1.000
1.000 / 1.000
0.501 / 0.494
0.999 / 1.000
0.503 / 0.508
0.502 / 0.500
0.498 / 0.488
0.501 / 0.504
1.000 / 1.000
0.504
TABLE II: Comprehensive evaluation of ROBIN++ and state-of-the-art baselines under diverse image degradations and security attacks. We report Copyright Robustness (Bit Acc. / AUC, higher is better) for Imperceptible, Frequency-domain, Advanced GenAI Removal, and Perturbation-removal attacks, and Forgery Success Rate (FSR, lower is better) for Black-box Forgery attacks. The best and second-best results are marked in bold and underlined , respectively.
Method
Reference
SD Inpaint [ 50 ]
SDXL [ 54 ]
Kandinsky [ 55 ]
ControlNet [ 56 ]
LaMa [ 57 ]
F1 ↑
AUC ↑
IoU ↑
F1 ↑
AUC ↑
IoU ↑
F1 ↑
AUC ↑
IoU ↑
F1 ↑
AUC ↑
IoU ↑
F1 ↑
AUC ↑
IoU ↑
MVSS-Net [ 35 ]
TPAMI’22
0.862
0.934
0.791
0.848
0.929
0.775
0.848
0.928
0.775
0.856
0.930
0.782
0.860
0.934
0.789
HDF-Net [ 36 ]
TPAMI’24
0.556
0.762
0.468
0.763
0.470
0.560
0.544
0.759
0.457
0.551
0.764
0.463
0.565
0.767
0.476
EditGuard [ 10 ]
CVPR’24
0.806
0.985
0.709
0.766
0.981
0.662
0.772
0.983
0.669
0.787
0.983
0.687
0.767
0.978
0.666
OmniGuard [ 11 ]
CVPR’25
0.853
0.964
0.810
0.867
0.973
0.824
0.868
0.966
0.830
0.858
0.965
0.815
0.864
0.969
0.823
WAM [ 13 ]
ICLR’25
0.924
0.977
0.868
0.918
0.976
0.862
0.921
0.976
0.865
0.917
0.977
0.860
0.922
0.967
0.864
TABLE III: Localization precision comparison on the generative-model-based tampering dataset. The best and second-best results are highlighted in bold and underlined , respectively.
Fig. 5: Qualitative comparison of tamper localization on the generative-model-based tampering dataset [ 50 , 54 , 55 , 56 , 57 ] . Our method generates precise localization masks that align closely with the ground truth (GT), effectively handling complex boundaries where baselines fail.
Fig. 6: Performance in complex scenarios with coexisting degradations and tampering. ROBIN++ achieves the best trade-off between robustness and tamper localization accuracy compared to baselines, effectively distinguishing benign distortions from malicious attacks.
Freq
Adv
CAP
F2S
S2F
Localization
Robustness
Avg
✓
0.387
0.973
0.680
✓
✓
0.543
0.979
0.761
✓
✓
✓
0.947
0.887
0.917
✓
✓
✓
✓
0.971
0.970
0.971
✓
✓
✓
✓
✓
0.971
0.973
0.972
TABLE IV: Ablation Study Results. Note that “Localization” is measured by average IoU under five tampering types, while “Robustness” reports the average accuracy under clean, rotation, JPEG, and blur settings.
Param.
Value
PickScore
Robustness
Local.
( ↑ )
(Acc. / AUC) ↑
(Avg) ↑
ξ
0.25
18.923
0.951 / 0.974
0.985
0.50 (Def.)
19.124
0.973 / 0.984
0.985
0.75
18.825
0.951 / 0.970
0.975
Freq. Cov.
20%
19.034
0.950 / 0.968
0.976
50%
18.777
0.961 / 0.978
0.984
TABLE V: Ablation study on key hyperparameters of the framework: CAP perturbation constraint ( ξ ) and frequency domain coverage ratio.
Fig. 7: Spectral separation of dual watermarks. The robust copyright watermark (Freq) and fragile localization watermark (CAP) occupy distinct frequency bands, ensuring they function without mutual interference.
Fig. 8: Effectiveness of the frequency-to-spatial (F2S) synergy. By leveraging frequency features, ROBIN++ achieves robust immunity to JPEG noise (top) and superior precision in inpainting localization (bottom).
Fig. 9: Effectiveness of the spatial-to-frequency (S2F) refinement. By utilizing the predicted tamper mask to suppress interference from the spliced region (red box), the extraction error is significantly reduced.
Method
SD v2.1 → SD v1.4
SD v2.1 → SD v2.0
F1 ↑
AUC ↑
IoU ↑
F1 ↑
AUC ↑
IoU ↑
StableGuard [ 14 ]
0.007
0.492
0.004
0.971
0.999
0.953
ROBIN++
0.765
0.927
0.707
0.974
0.999
0.956
TABLE VI: Quantitative comparison of zero-shot transferability trained on Stable Diffusion (SD) v2.1.
Figure 16Figure 17Figure 18Figure 19Figure 20
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Copy-move
Splicing
F1 ↑
AUC ↑
IoU ↑
F1 ↑
AUC ↑
IoU ↑
EditGuard [ 10 ]
0.847
0.990
0.757
0.871
0.989
0.791
OmniGuard [ 11 ]
0.786
0.967
0.722
0.672
0.949
0.595
WAM [ 13 ]
0.944
0.999
0.903
0.955
0.999
0.919
TAG-WM [ 12 ]
0.740
0.849
0.636
0.791
0.875
0.689
StableGuard [ 14 ]
0.973
0.999
0.955
0.978
0.999
0.962
Appendix
TABLE VII: Localization performance comparison on the balanced CASIA subset. The best and second-best results are highlighted in bold and underlined , respectively.
Method
Small ( <5% )
Medium ( 5%−15% )
Large ( >15% )
F1 ↑
IoU ↑
F1 ↑
IoU ↑
F1 ↑
IoU ↑
EditGuard [ 10 ]
0.701
0.587
0.789
0.688
0.829
0.741
OmniGuard [ 11 ]
0.871
0.831
0.866
0.825
0.845
0.800
WAM [ 13 ]
0.872
0.794
0.927
0.871
0.949
0.910
TAG-WM [ 12 ]
0.528
0.402
0.778
0.666
0.873
0.784
StableGuard [ 14 ]
0.972
0.946
0.982
0.965
0.983
0.967
Appendix
TABLE VIII: Quantitative breakdown of tamper localization performance across different manipulated region sizes. The best and second-best results are highlighted in bold and underlined , respectively.
Fig. 10: Degradation profile of ROBIN++ under progressive Gaussian noise with varying intensities ( σ∈{0.05,0.10,0.20,0.30,0.50,0.75,1.00} ).
Fig. 11: Error maps for tamper localization under challenging degradation conditions. False positives are marked in red, and false negatives are marked in blue.
Fig. 12: Amplified residual maps between original and watermarked images in challenging cases.
Injection Scheduling
Peak
Robustness
Localization
PickScore ↑
tinj
(Bit Acc. / AUC) ↑
((F1+AUC+IoU)/3) ↑
Fixed (50 samples, cfg=7.5)
300
0.973 / 0.984
0.985
19.124
Adaptive (50 samples, cfg=7.5)
280
0.966 / 0.985
0.987
19.142
Adaptive (200 samples, cfg=7.5)
220
0.957 / 0.975
0.987
19.154
Adaptive (50 samples, cfg=5.5)
240
0.961 / 0.978
0.987
19.151
Adaptive (other 50 samples, cfg=7.5)
280
0.966 / 0.985
0.987
19.142
Appendix
TABLE IX: Comparison between fixed injection scheduling and the proposed Adaptive Injection Scheduling. Peak tinj denotes the timestep with the highest final sampling probability for the adaptive scheduler.
Notation
Description
zt , z0∗ , zw , z~0
Noisy, copyright-embedded, dual-watermarked, and rectified latent
wid , w^id
Embedded robust copyright watermark and its extracted estimate
wloc , wp
Fragile localization watermark, and learnable prompt guidance
I(⋅) , T(⋅)
Dual-domain injection function, and compound degradation process
Gloc(⋅) , Dloc(⋅)
Content-aware perturbation generator, and synergistic detector
Dec(⋅)
VAE decoder mapping latent representations to pixel space
Appendix
TABLE X: Summary of Key Notations
Pipeline Stage
Specific Component
Time Cost / Runtime
Memory Overhead (VRAM)
Frequency of Cost
Offline Preparation (Model-specific training)
Adversarial Optimization
∼ 0.20 Hours
12.7 GB
One-time
CAP Generator Training
∼ 50.00 Hours
41.7 GB
One-time
Online Injection (Generation pipeline)
Standard Diffusion (Baseline)
∼ 1.64 s / img
3.9 GB
Per generation
ROBIN++ Synergistic Injection
∼ 1.78 s / img
4.9 GB
Per generation
Relative Overhead Δ
+0.14 s / img
+1.0 GB
-
Practical Verification (Dual-branch detection)
Copyright Extraction (Freq.)
∼ 0.30 s / img
3.8 GB
Per verification
Appendix
TABLE XI: Comprehensive evaluation of the computational overhead and practical efficiency of the ROBIN++ framework. All measurements were conducted on a single NVIDIA RTX 4090 (48GB) GPU using a batch size of 1.
Method
Ref.
Robustness (Bit Acc ↑ / AUC ↑ )
Inpainting Localization ((F1+AUC+IoU)/3) ↑
Quality
Clean
Rotation
JPEG
Blur
Avg
SDXL
Kandinsky
LaMa
Avg
PickScore
EditGuard
CVPR’24
0.834 / 0.983
0.505 / 0.553
0.498 / 0.512
0.455 / 0.286
0.573 / 0.584
0.837
0.830
0.841
0.836
21.212
OmniGuard
CVPR’25
1.000 / 1.000
0.493 / 0.478
1.000 / 1.000
1.000 / 1.000
0.873 / 0.870
0.925
0.942
0.683
0.850
21.246
WAM
ICLR’25
1.000 / 1.000
0.479 / 0.405
0.990 / 1.000
1.000 / 1.000
0.867 / 0.851
0.950
0.946
0.953
0.950
21.028
StableGuard
NeurIPS’25
1.000 / 1.000
0.498 / 0.478
0.499 / 0.480
0.500 / 0.490
0.624 / 0.612
0.966
0.965
0.967
0.966
20.972
ROBIN++
-
1.000 / 1.000
0.673 / 0.650
1.000 / 1.000
1.000 / 1.000
0.918 / 0.913
0.970
0.968
0.972
0.970
21.009
Appendix
TABLE XII: Evaluation of ROBIN++ Generalization to Stable Diffusion XL (SDXL). We include the baselines with reliable implementations on SDXL. The best and second-best results are highlighted in bold and underlined , respectively.
Method
Ref.
Robustness (Bit Acc ↑ / AUC ↑ )
Inpainting Localization ((F1+AUC+IoU)/3) ↑
Quality
Clean
Rotation
JPEG
Blur
Avg
SDXL
Kandinsky
LaMa
Avg
PickScore
EditGuard
CVPR’24
1.000 / 1.000
0.487 / 0.456
0.701 / 0.964
0.472 / 0.414
0.665 / 0.709
0.666
0.606
0.662
0.645
20.588
OmniGuard
CVPR’25
1.000 / 1.000
0.502 / 0.497
0.994 / 1.000
0.997 / 1.000
0.873 / 0.874
0.975
0.963
0.885
0.941
20.558
WAM
ICLR’25
1.000 / 1.000
0.496 / 0.510
0.944 / 0.998
1.000 / 1.000
0.860 / 0.877
0.955
0.956
0.955
0.955
20.455
StableGuard
NeurIPS’25
1.000 / 1.000
0.498 / 0.486
0.495 / 0.491
0.490 / 0.458
0.621 / 0.609
0.970
0.961
0.958
0.963
20.431
ROBIN++
-
1.000 / 1.000
0.908 / 0.944
0.980 / 0.986
0.960 / 0.990
0.962 / 0.980
0.977
0.961
0.959
0.966
20.472
Appendix
TABLE XIII: Evaluation of ROBIN++ Generalization to PixArt- α . We include the baselines with reliable implementations on PixArt- α . The best and second-best results are highlighted in bold and underlined , respectively.
Diffusion-based generative models have transformed visual content synthesis, yet they remain vulnerable to unauthorized usage and lack reliable attribution methods. Existing watermarking techniques often treat latent tensors as static spatial feature maps or depend on pixel-domain modification, and most do not explicitly leverage the internal frequency structure of the latent space for dual-band redundant embedding, leaving them susceptible to the stochastic nature of diffusion and regeneration attacks. We introduce BiSLW, a trainable bi-spectral latent watermarking framework that jointly embeds aligned identity signals across complementary spectral bands of the decoded diffusion latent using learned encoders and decoders, going beyond fixed-pattern frequency approaches. We leverage the inherent frequency structure of diffusion latents to design a dual-band watermarking framework. Low-frequency components encode global semantics, while high-frequency components capture fine texture. We exploit this structure to embed watermarks across complementary spectral bands. The watermark is independently injected into both bands via learned encoders and recombined before decoding, ensuring it becomes intrinsic to the generative trajectory. Dual spectral decoders recover the watermark from each band, while a cross-band consistency constraint enforces alignment between semantic and textural embeddings. Experiments show that BiSLW achieves a strong balance between perceptual fidelity and robustness, improving PSNR by over 3 dB compared to prior latent diffusion watermarking methods while preserving near-perfect bit accuracy under aggressive regeneration and common distortions, all with negligible computational overhead.
Aryan Pandit
PDPM Indian Institute of Information Technology, Jabalpur, India
With the proliferation of AI-generated images, digital watermarking has become an essential safeguard for protecting intellectual property and mitigating malicious exploitation. Recent works on semantic watermarking have enabled efficient copyright protection for diffusion models. However, the dependence of semantic watermarking on diffusion inversion for watermark detection creates a critical vulnerability. Imprint removal and forgery attacks exploit this weakness to produce deceptive results. Our analysis reveals that these attacks succeed by displacing watermarked latents into the unwatermarked region, while guiding unwatermarked latents into the watermarked region. Based on that, we propose Progressive Guided Inversion and Denoising (PGID), the first plug-and-play, training-free noise extraction framework designed to defend against both attack strategies. PGID effectively defends by projecting perturbed latents back to the region where they originally belong. The projection is achieved by eliminating intermediate latent deflections and mitigating adversarial perturbations through progressive inversion-denoising cycles. Comprehensive evaluations across multiple schemes demonstrate that PGID successfully restores detection reliability by recovering removed watermarks and identifying forged instances.
As open-weights generative AI rapidly proliferates, the ability to synthesize hyper-realistic media has introduced profound challenges to digital trust. Automated disinformation and AI-generated imagery have made robust digital provenance a critical cybersecurity imperative. Currently, state-of-the-art invisible watermarks operate within one of two primary mathematical manifolds: the spatial domain (post-generation pixel embedding) or the latent domain (pre-generation frequency embedding). While existing literature frequently evaluates these models against isolated, classical distortions, there is a critical lack of rigorous, comparative benchmarking against modern generative AI editing tools. In this study, we empirically evaluate two leading representative paradigms, RivaGAN (Spatial) and Tree-Ring (Latent), utilizing an automated Attack Simulation Engine across 30 intensity intervals of geometric and generative perturbations. We formalize an "Adversarial Evasion Region" (AER) framework to measure cryptographic degradation against semantic visual retention (OpenCLIP > 75.0). Our statistical analysis (n=100 per interval, MOE=±3.92%) reveals that these domains possess mutually exclusive, mathematically orthogonal vulnerabilities. Spatial watermarks experience severe cryptographic degradation under algorithmic pixel-rewriting (exhibiting a 67.47% AER evasion rate under Img2Img translation), whereas latent watermarks exhibit profound fragility against geometric misalignment (yielding a 43.20% AER evasion rate under static cropping). By proving that single-domain watermarking is fundamentally insufficient against modern adversarial toolsets, this research exposes a systemic vulnerability in current digital provenance standards and establishes the foundational exigence for future multi-domain cryptographic architectures.
Jesse Yu, Nicholas Wei
Millburn High School · Williamsville East High School