We, the "Seallage" team¹, are proud to announce that our novel approach to watermarking generative AI outputs emerged victorious at the Meta x Entrepreneur First AI Safety Hackathon, a two-day competition hosted at Meta’s Paris office.

Following a stringent pre-selection process, the hackathon challenged 40+ students, researchers and industry professionals to create solutions addressing AI risks. Our project integrated key concepts from cutting-edge GenAI watermarking papers²³, enabling us to embed an information-rich, invisible signature in AI outputs.

Our method embeds multi-bit signatures in the latent vectors central to Stable Diffusion and similar Latent Diffusion Models (LDMs). We jointly train a watermark encoder and extractor model, maximizing recoverability of the embedded metadata while minimizing the impact of our watermark on image quality. As such, we are able to subtly encode metadata that is undetectable to humans yet recoverable algorithmically.

Our approach provides robustness to adversarial distortions, and works without fine-tuning target models, enabling zero-shot watermarking. We call it “Learned Latent Watermarking for Diffusion”, or LeLaWD for short.

We validated our early prototype on Stable Diffusion outputs, impressing an exceptional jury presided by France's National AI Coordinator, Dr. Guillaume Avrin, and comprising Dr. Alexandre Défossez (Research Scientist at Meta/FAIR Paris), Dr. Damien Henry (Senior VP of Product at Stability AI), Dr. Ania Kaci (Ambassador for Women in AI in France), and Coralie Chaufour (Partner & General Manager at Entrepreneur First).

Participating in the hackathon allowed us to receive invaluable guidance from leading experts including Meta AI/FAIR researchers, Entrepreneur First staff, and EffiSciences AI safety specialists. The competition was fierce, but our team stood out thanks to the potential of our model-agnostic latent watermarking approach to enable improved transparency and auditability of model outputs.

We hope to bring this work to publication in the near future.

---

Need high-impact technical consulting?

---

¹ Yohaï-Eliel Berreby, Sylvain Girard, Rania Ferchichi, Antoine Delplace, and Géraud Martin-Montchalin.

² Yuxin Wen, John Kirchenbauer, Jonas Geiping, Tom Goldstein: “Tree-Ring Watermarks: Fingerprints for Diffusion Images that are Invisible and Robust”, 2023; arXiv:2305.20030.

³ Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, Teddy Furon: “The Stable Signature: Rooting Watermarks in Latent Diffusion Models”, 2023; arXiv:2303.15435.