Skip to main content
Back to timeline
arXivSource publication:

Semantic watermarks for black-box generators need only sampling: ten-image batches reach 0.998 AUROC

Synopsis

A semantic watermarking scheme that touches a generative model only through sampling and a semantic embedding only through queries leaves the generator's output distribution unchanged when averaged over the secret key, so it does not degrade output quality. It samples several candidate outputs per prompt, maps each to a semantic code, and returns one chosen by a keyed score; detection aggregates several outputs from the same source. On Stable Diffusion 2.1, single-image AUROC is 0.802 and rises to 0.998 for batches of ten, while OpenAI's closed GPT Image API reaches 0.978 for batches of five.

Source-provided article image: Distortion-Free Semantic Watermarks for Black-Box Generative Models
arXiv

Interpretation

The watermark no longer needs the diffusion model's initial noise or latent space: the generator is treated as a black box returning independent samples and the semantic embedding as a black box mapping a finished image to a vector, and the output distribution is preserved after averaging over the key. Earlier semantic watermarks relied either on autoregressive token-by-token generation or on diffusion-specific access such as control of the initial noise and diffusion inversion, which closed image APIs do not expose. On Stable Diffusion 2.1 with held-out COCO-2017 prompts, single-image AUROC is 0.802 and ten-image batch AUROC is 0.998; on the sampling-only GPT Image API, five-image batches reach 0.978.

The method selects among complete candidate images: each candidate is embedded, a random-hyperplane hash assigns its embedding to a semantic region, a secret key gives each region a pseudorandom score, and the highest-scoring candidate is returned. Because images in the same semantic region share a score, an edit that keeps the region unchanged keeps the detection score; a boundary-aware detector also checks neighboring codes near hyperplane boundaries. On 1,000 held-out COCO-2017 prompts with Stable Diffusion 2.1, matched-population FID under fixed-key reuse ranges from 4.75 to 5.22, below Tree-Ring at 5.61, Gaussian Shading at 5.42 and SEAL at 6.88, with same-prompt diversity staying near clean generation.

Perspective

The demonstrated scope is image generation with Stable Diffusion 2.1 on held-out COCO-2017 prompts, plus OpenAI's GPT Image API accessed only through sampling; detection is mainly reported for batches of outputs from one source, with single-image detection reported separately. Text-to-audio is demonstrated once with AudioLDM 2 and LAION-CLAP.

Distortion-freeness holds for a single output averaged over the secret key, not for every fixed key; the detection guarantee is likewise averaged over the key and random hyperplanes. Single-image detection has a ceiling inherent to black-box sampling, with true-positive rate bounded by a multiple of the false-positive rate. Robustness experiments cover non-adaptive appearance, compression, geometric and generative edits, and do not evaluate adaptive attacks on the encoder or detector.