Semantic watermarks keep ai images unchanged and detectable after edits

Distortion-Free Semantic Watermarks for Black-Box Generative Models

Machine LearningCryptography and SecurityComputer Vision and Pattern Recognition

Summary

Generating images with AI can introduce watermarks to prove origin, but these often lower quality or break when the image is resized or compressed. The authors describe a way to add hidden watermarks that don’t change the image’s usual look and stay detectable even after edits. Their method treats the AI as a black box, needing only sample access, and uses a custom image encoder to focus on meaningful content, not pixel details. This makes the watermark both reliable and robust, working on many AI image models without needing to see their internal workings.

What this means in practice

  • For ai image service providers: Add undetectable watermarks that survive image edits to prove images come from specific AI models.$Commercial implications: Enables commercial AI platforms to protect model-generated images and enforce usage policies by embedding robust provenance marks.
  • For digital content moderators: Detect and verify AI-generated images reliably even after compression, cropping, or resizing.

Authors

Aviv Yossef, Or Patashnik, Or Zamir

Abstract

We present a provably distortion-free semantic watermarking algorithm that requires only black-box sampling access to a generative model and black-box access to a semantic embedding. A watermark is called distortion-free if, averaged over its secret detection key, it leaves the generator's output distribution unchanged; this provides a formal guarantee that watermarking does not reduce output quality. Conventional such watermarks are often brittle to transformations. This is especially problematic for images, where benign operations such as compression, cropping, or resizing change the entire pixel representation. Semantic watermarks therefore require detection to be robust to content-preserving transformations. Previous methods combining these properties rely heavily on specific white-box properties of the underlying model: either to autoregressive text generation or to diffusion-specific access, such as control of the initial noise or diffusion inversion. In contrast, our construction is fully generic: It only interacts with the generative model and the semantic embedding, a function that determines the notion of "preserving content" under which detection should be robust, via sample and query access. We then instantiate this framework for image generation. We characterize the two competing properties required of the semantic embedding: it should remain stable under transformations that preserve an image's content, but should distinguish independent generations from the same prompt. We train an image encoder directly for this tradeoff. The resulting encoder improves separation while maintaining stability over off-the-shelf alternatives and enables strong aggregate detection while remaining robust to various transformations. We demonstrate the effectiveness of our watermark scheme on both open-source and closed models, which is possible only due to restriction to black-box access.