[KongchangAI]
· 1 min read· 845 words

Distortion-Free Semantic Watermarking for Black-Box Generative Models

Distortion-Free Semantic Watermarking for Black-Box Generative Models

New framework achieves provably distortion-free and robust semantic watermarking with only black-box access.

The paper proposes a fully general semantic watermarking framework requiring only black-box sampling access to the generative model and semantic embeddings, achieving provable distortion-freeness and robustness to content-preserving transformations. To address the fragility of traditional distortion-free watermarks to compression and cropping and their reliance on white-box access, the authors base detection in semantic embedding space and train an image encoder to balance transformation stability and discriminability between independent generations. Experiments on open- and closed-source models show improved separation, strong aggregate detection, and no sacrifice in output quality. This approach reduces dependence on internal model access, enables third parties to watermark closed-source models, and offers a new path for copyright protection of AI-generated content.

A New Watermarking Approach for Black-Box Generative Models

Watermarking for generative models has always faced a dilemma: it must ensure the watermark does not affect generation quality (distortion-free), while also ensuring the watermark can still be detected after benign transformations such as compression and cropping (robustness). Traditional methods often rely on white-box access to the model—for example, requiring control over the initial noise of diffusion models or exploiting the characteristics of autoregressive text generation. A recent paper proposes a fully general semantic watermarking framework that requires only black-box sampling access to the generative model and a semantic embedding, achieving provable distortion-freeness and robustness to content-preserving transformations.

Paper title screenshot

The Tension Between Distortion-Freeness and Robustness

A distortion-free watermark is defined as one that, when averaged over the secret detection key, does not change the generator's output distribution. This formally guarantees that the watermark does not degrade output quality. However, traditional distortion-free watermarks are highly fragile to transformations—especially in the image domain, where compression, cropping, or scaling alters the entire pixel representation. Semantic watermarking requires detection to be robust to content-preserving transformations, but previous methods combining the two rely heavily on white-box properties of the model, such as initial noise control in diffusion models or diffusion inversion.

Distortion-free watermarking is formally defined based on indistinguishability from cryptography: if the detection key is uniformly random over the key space, the output distribution after watermark embedding is identical to the original generation distribution (total variation distance of 0). This means any attacker who only observes the generated content cannot distinguish whether a watermark is present, so distortion-freeness inherently provides resistance to watermark removal attacks. But this property also implies a corollary: the watermark signal must be "hidden" in the randomness of the generation process rather than explicitly modifying the output. Traditional distortion-free image watermarks are often implemented via constrained pixel-level perturbations, but such perturbations are destroyed after JPEG compression or cropping, causing detection to fail.

Semantic embedding is a function that maps an image to a low-dimensional semantic space, whose distance metric defines the boundary of "content preservation." A good semantic embedding should satisfy: after applying content-preserving transformations such as compression and scaling to the same image, the embedding vectors are close; while for images with different semantic content, the distance is large. Watermark detection is performed in this space, so robustness does not depend on pixel-level exact matching.

Building a General Black-Box Framework

The paper's construction is fully general: it interacts with the generative model and the semantic embedding only through samples and queries. The semantic embedding defines the notion of "content preservation," under which detection should remain robust. The authors instantiate the framework for image generation and characterize two competing properties of the semantic embedding: it should be stable under transformations that preserve image content, yet discriminative between independent generations from the same prompt. They directly train an image encoder to balance these two points.

Black-box sampling access means that one can only obtain output samples by querying the generative model, without access to model parameters, internal activations, or random seeds. This contrasts with white-box access: for example, white-box watermarks for diffusion models often control the initial noise distribution, while watermarks for autoregressive models modify token sampling probabilities. In the black-box setting, the watermark must be embedded in the output samples themselves or implemented via post-processing, while not changing the output distribution. The framework decouples the watermark signal from content via semantic embedding; detection only requires comparing statistical properties of semantic embeddings, making it applicable to any samplable model.

Aggregate detection refers to jointly statistically testing the watermark signal across multiple generated samples, rather than making per-sample decisions. Since the per-sample signal of a distortion-free watermark is extremely weak, aggregate detection can significantly improve detection power. A common approach is to compute the deviation of the mean semantic embedding of samples from a reference distribution and use a hypothesis test (e.g., chi-squared test) to determine the presence of the watermark. The robustness of aggregate detection depends on the stability of the semantic embedding under transformations and the discriminability between independently generated samples.

Experimental Results and Advantages

The authors validated the watermarking scheme on both open-source and closed-source models, thanks to the restriction of using only black-box access. The trained encoder improves separation compared to off-the-shelf alternatives while maintaining stability, achieving strong aggregate detection and robustness to various transformations. This method provides provable guarantees for watermarking black-box generative models without sacrificing output quality.

Practical Implications

For copyright protection and provenance of AI-generated content, this scheme reduces reliance on internal model access, enabling third parties to add watermarks to closed-source models. At the same time, the distortion-free property ensures the watermark does not affect the visual quality of generated content, which is crucial for real-world deployment.

Conclusion

Through a clever combination of black-box access and semantic embedding, this research solves the long-standing problem in watermarking where distortion-freeness and robustness are difficult to achieve simultaneously. Its general framework is not limited to images and may be extended to other modalities in the future.

#semantic watermarking#distortion-free watermark#black-box generative models#AI content provenance#image encoder
Share:

Related articles