SynthID-Image: Image watermarking at internet scale
paperAuthors
Credibility Rating
Good quality. Reputable source with community review or editorial standards, but less rigorous than peer-reviewed venues.
Rating inherited from publication venue: arXiv
A key industry paper on practical AI-generated content authentication; relevant to policy discussions around mandatory watermarking of synthetic media and technical approaches to mitigating AI-enabled misinformation.
Paper Details
Metadata
Abstract
We introduce SynthID-Image, a deep learning-based system for invisibly watermarking AI-generated imagery. This paper documents the technical desiderata, threat models, and practical challenges of deploying such a system at internet scale, addressing key requirements of effectiveness, fidelity, robustness, and security. SynthID-Image has been used to watermark over ten billion images and video frames across Google's services and its corresponding verification service is available to trusted testers. For completeness, we present an experimental evaluation of an external model variant, SynthID-O, which is available through partnerships. We benchmark SynthID-O against other post-hoc watermarking methods from the literature, demonstrating state-of-the-art performance in both visual quality and robustness to common image perturbations. While this work centers on visual media, the conclusions on deployment, constraints, and threat modeling generalize to other modalities, including audio. This paper provides a comprehensive documentation for the large-scale deployment of deep learning-based media provenance systems.
Summary
SynthID-Image is Google DeepMind's system for imperceptibly watermarking AI-generated images to enable detection of synthetic content at scale. The paper describes the technical approach to embedding and detecting watermarks that survive common image transformations, and reports on its deployment across Google's image generation products serving hundreds of millions of users.
Key Points
- •Presents a production-scale invisible watermarking system for AI-generated images deployed across Google products including Gemini and Imagen.
- •Watermarks are designed to be robust against common image manipulations like cropping, compression, and color adjustments while remaining imperceptible.
- •Addresses the challenge of provenance verification for synthetic media without requiring centralized databases or metadata that can be stripped.
- •Evaluates watermark robustness, false positive/negative rates, and image quality trade-offs at internet scale.
- •Represents a practical deployment of AI content authentication technology relevant to misinformation and deepfake detection efforts.
Cited by 1 page
| Page | Type | Quality |
|---|---|---|
| AI Content Authentication | Approach | 58.0 |
Cached Content Preview
[2510.09263] SynthID-Image: Image watermarking at internet scale
\correspondingauthor
sgowal@google.com, pushmeet@google.com \paperurl \reportnumber
SynthID-Image : Image watermarking at internet scale
Sven Gowal
Google DeepMind
Core contributor (randomized order)
Lead
Rudy Bunel
Google DeepMind
Core contributor (randomized order)
Florian Stimberg
Google DeepMind
Core contributor (randomized order)
David Stutz
Google DeepMind
Core contributor (randomized order)
Guillermo Ortiz-Jimenez
Google DeepMind
Core contributor (randomized order)
Christina Kouridi
Google DeepMind
Core contributor (randomized order)
Mel Vecerik
Google DeepMind
Core contributor (randomized order)
Jamie Hayes
Google DeepMind
Core contributor (randomized order)
Sylvestre-Alvise Rebuffi
Google DeepMind
Core contributor (randomized order)
Paul Bernard
Google DeepMind
Contributor
Chris Gamble
Google DeepMind
Contributor
Miklós Z. Horváth
Google DeepMind
Contributor
Fabian Kaczmarczyck
Google DeepMind
Contributor
Alex Kaskasoli
Google DeepMind
Contributor
Aleksandar Petrov
Google DeepMind
Contributor
Ilia Shumailov
Google DeepMind
Contributor
Meghana Thotakuri
Google DeepMind
Contributor
Olivia Wiles
Google DeepMind
Contributor
Jessica Yung
Google DeepMind
Contributor
Zahra Ahmed
Google DeepMind
Support
Victor Martin
Google DeepMind
Support
Simon Rosen
Google DeepMind
Support
Christopher Savčak
Google DeepMind
Support
Armin Senoner
Google DeepMind
Support
Nidhi Vyas
Google DeepMind
Support
Pushmeet Kohli
Google DeepMind
Lead
Sponsor
Abstract
We introduce SynthID-Image , a deep learning-based system for invisibly watermarking AI-generated imagery.
This paper documents the technical desiderata, threat models, and practical challenges of deploying such a system at internet scale, addressing key requirements of effectiveness, fidelity, robustness, and security.
SynthID-Image has been used to watermark over ten billion images and video frames across Google’s services and its corresponding verification service is available to trusted testers.
For completeness, we present an experimental evaluation of an external model variant, SynthID-O , which is available through partnerships.
We benchmark SynthID-O against other post-hoc watermarking methods from the literature, demonstrating state-of-the-art performance in both visual quality and robustness to common image perturbations.
While this work centers on visual media, the conclusions on deployment, constraints, and threat modeling generalize to other modalities, including audio.
This paper provides a comprehensive documentation for the large-scale deployment of deep learning-based media pro
... (truncated, 98 KB total)10607c87667b587e | Stable ID: sid_mmw7KgnrBT