As diffusion models (DMs) democratize high-fidelity image generation, they simultaneously open a Pandora’s box of security risks, ranging from copyright infringement to the creation of sophisticated deepfakes. Digital watermarking is the standard defense, but it faces a fundamental "Goldilocks" problem: a watermark robust enough to survive JPEG compression or rotation (copyright protection) is often too blunt to detect subtle, localized edits (tamper localization). Conversely, a highly sensitive "fragile" watermark is easily destroyed by benign image processing.
In their paper ROBIN++, Huang et al. propose a paradigm shift to break this stalemate. Instead of entangling these conflicting requirements in a single domain, they introduce a Dual-Domain Synergistic design that decouples robustness and fragility during injection and re-couples them for mutual reinforcement during verification.
TL;DR
ROBIN++ resolves the inherent trade-off between copyright robustness and tamper sensitivity by embedding watermarks into two distinct domains: frequency for robustness and spatial for fragility. By using a synergistic detector that allows these signals to cross-reference each other during verification, the system achieves state-of-the-art performance in both provenance and integrity verification.
What Problem Does This Paper Address?
Existing unified watermarking frameworks typically inject multiple signals into the same domain—usually the pixel space. This creates an optimization conflict. Strengthening a watermark’s robustness tends to smooth out the fine-grained variations needed for precise localization.
As shown in the conceptual comparison below, traditional "entangled" injection leads to independent verification where the two tasks don't help each other. ROBIN++ argues that these tasks are inherently coupled: you only care about tampering if the image is yours, and heavy tampering shouldn't necessarily break your ability to prove ownership.

Methods
The core of ROBIN++ is its Dual-Domain Injection and Synergistic Extraction pipeline.
1. Dual-Domain Injection
The framework embeds two distinct signals at different stages of the diffusion process:
- Frequency-Domain (Robustness): Using an Adversarial Optimization Algorithm, a copyright watermark ($w_{id}$) is injected into the latent trajectory at an intermediate timestep. This ensures the watermark is "absorbed" by the model's generation process, maximizing strength while a learnable guidance signal minimizes visual artifacts.
- Spatial-Domain (Fragility): A lightweight Content-Aware Perturbation (CAP) generator injects a fragile watermark ($w_{loc}$) into the fully converged latent space. This signal is designed to be spectrally disjoint from the robust watermark to prevent interference.
2. Bidirectional Synergistic Extraction
The "magic" happens during verification. Instead of decoding the signals in isolation, ROBIN++ uses a dual-branch detector:
- Frequency-to-Spatial (F2S) Synergy: The robust spectral signal acts as a global anchor. It helps the spatial detector distinguish between benign global distortions (like JPEG noise) and malicious local tampering.
- Spatial-to-Frequency (S2F) Refinement: If tampering is detected, the predicted spatial mask is used to "quarantine" corrupted regions. The system then extracts the copyright watermark only from the intact areas, significantly improving recovery rates for heavily edited images.
Datasets
The authors evaluated ROBIN++ across several benchmarks:
- Training: Self-supervised training on the MS-COCO dataset.
- Tamper Localization: Evaluated on 5,000 image-mask pairs from the StableGuard dataset, featuring inpainting from five different models (SD, SDXL, Kandinsky, ControlNet, and LaMa).
- Non-Generative Edits: Tested on the CASIA 2.0 dataset for traditional copy-move and splicing attacks.
- Models: Experiments were conducted on Stable Diffusion v2, SDXL, and PixArt-α.
Key Insights and Ideas
- Explicit Decoupling: By assigning robustness to the frequency domain and fragility to the spatial domain, the researchers allow each signal to be optimized within its "natural habitat" without suppressing the other.
- In-Process vs. Post-Hoc: Unlike methods that add watermarks to finished images, ROBIN++ operates during generation. This allows the watermark to become a structural part of the image, making it much harder to remove without destroying the image itself.
- Mutual Promotion: The S2F refinement is a standout idea—it acknowledges that a tampered image is "partially broken" and uses the localization mask to fix the copyright extraction process.
Key Results
ROBIN++ effectively eliminates the performance trade-off seen in previous SOTA methods like StableGuard and TAG-WM.
- Robustness: It achieved an average AUC of 0.984 across various degradations. Most impressively, it maintained a 0.935 AUC under 75-degree rotation, a scenario where most baselines dropped to near-random (0.50) performance.
- Localization: It reached a state-of-the-art average precision of 0.985 for tamper localization, outperforming both active watermarking and passive forensic methods.
- Black-Box Forgery: In tests against "latent forgery" attacks (where an attacker tries to imitate the watermark trace), ROBIN++ showed the lowest Forgery Success Rate (0.411) compared to baselines.
Limitations and Future Work
The primary limitation is that ROBIN++ is an in-generation method. It requires access to the diffusion sampling process, meaning it cannot be applied to images after they have already been generated by an un-watermarked model.
The authors suggest that the "decoupled injection + synergistic verification" principle could be extended beyond 2D images to other generative modalities, such as video, 3D assets, and multi-view content, where temporal and geometric consistency could provide even more avenues for robust watermarking.
Summary updated