Unlocking the Future of Content Authentication: IMATAG's Breakthrough in AI-Generated Image Watermarking

‍

Digital Watermarking: How It Works

To understand IMATAG's innovation, it's essential to grasp how digital watermarking functions. The process consists of two crucial steps: watermarking and identification.

During watermarking, an invisible digital watermark is embedded into the image. Identification involves scanning the image or video to detect the watermark's presence.

‍

‍

Traditionally, watermarking occurred after content creation. However, with generative AI, a more suitable approach is to introduce watermarking during the generation process, as explained later in this article.

‍

IMATAG's Watermark for AI-Generated Images

IMATAG's "Watermarked SDXL-Turbo Demo" on Hugging Face

On November 28, 2023, Stable Diffusion released SDXL Turbo, an advanced AI image generation model. Just two days later, IMATAG experts made available an invisible watermarking solution on Hugging Face to identify images generated by this AI.

In this demo, users can generate an image by providing a prompt describing the desired image. An imperceptible watermark is inserted during the image generation process, certifying it as synthetic. The demo also allows for watermark detection (in the form of a probability) after subjecting the generated image to various attacks, such as compression or recolorization.

This technique is based on the refinement of a process developed by scientists at INRIA and Meta, known as StableSignature (see references). It involves modifying the weights of the generative AI model (in this case, SDXL Turbo) to make it naturally generate watermarked images. This approach is ideal for open-source generative AI, unlike the "generate-then-watermark" techniques. For enhanced robustness, IMATAG combined this method with its in-house zero-bit decoder (indicating the presence or absence of a watermark, without payload), capable of decoding even highly altered watermarks.

‍

Try It Out

To explore the potential of this technique, you can test it on Hugging Face, adjusting parameters to control watermark visibility and robustness. 

Visibility Impact: The training process balances watermark robustness and image quality. The perceptibility goal aligns with the original objective of Stable Diffusion, trained on real images.

Robustness: The demo allows you to adjust the watermark strength in the generated model, testing various trade-offs between perceptibility and robustness.

Demo Limitations

It's important to note that the detector used in this demo is a "degraded" version of our in-house detector. It has been trained to resist cropping, JPEG compression, and screen capture by smartphones. Challenges like image flips, extreme perspectives, rotations, and text addition were not part of the training. This exercise focuses on a single watermark (no traitor tracing), without a hidden message. The detector will only identify this specific watermark and no other. 

For additional features, such as security keys, multiple watermarks, concealed messages, or increased robustness, please don't hesitate to contact us.

‍

What Makes This Demo Remarkable?

IMATAG's solution stands out as the first independent watermarking technology for AI models. While some AI developers, like DeepMind with SynthID, already apply watermarking algorithms, they typically limit these algorithms to their own models.

The primary goal of this demo is scientific advancement. In the case of Stable Diffusion, one can simply reload the non-watermarked model to generate watermark-free images, since the original weights are publicly available. IMATAG aims to encourage model producers, such as StabilityAI, to watermark their models before any public weight release.

This initiative underscores Imatag's message to the open-source AI community: Seek digital watermarking experts for protection and GenAI watermarking compliance. The speedy and effective implementation of a tailored solution, as shown in the demo released just two days after the model's launch, is achievable, avoiding the typical quality loss associated with post-generation watermarking.

‍

Conclusion and Future Outlook

In an age of increasingly convincing AI-generated images, the importance of authentication is paramount, notably in the context of deepfakes and highly realistic AI creations. European and U.S. authorities now insist on providing users with tools to distinguish genuine content from synthetic creations.

While recent industry announcements, such as DeepMind's SynthID, offer reassurance, they often remain in beta for watermarking, with limited immediate impact. Moving forward, the journey involves refining tools, adapting regulations, raising awareness, and fostering collaboration among tech leaders, regulators, and users. In this AI-dominated era, trust, security, and authenticity remain pivotal concerns, and the digital watermarking journey, initiated in the 1950s, continues to evolve.

‍

-----

About StableSignature: Stable Signature’s approach is to train the generative AI model to generate natively watermarked images. This blog post reflects the work of Matthijs Douze and Pierre Fernandez, with the contributions of Guillaume Couairon, Teddy Furon, and Hervé Jégou to this research. Other resources available on their GitHub Repository.

More technical details from Imatag-lab about the algorithm here.