Tests
The modern tool helps watermark and identify synthetic images created by Imagen
Images generated by artificial intelligence are becoming more and more popular every day. But how can we better identify them, especially when they look so realistic?
Today in cooperation with Google Cloud, we are launching the beta version of SynthID, a tool for watermarking and identifying images generated by artificial intelligence. This technology embeds a digital watermark directly into the image pixels, making it concealed to the human eye but detectable for identification purposes.
SynthID is made available to a narrow number Apex AI customers benefit Pictureone of our newest text-to-image models that uses input text to create photorealistic images.
Generative AI technologies are developing rapidly, and computer-generated images, also known as “synthetic images,” are becoming increasingly challenging to distinguish from those not created by an AI system.
While generative AI can unlock enormous original potential, it also poses modern risks, such as allowing creators to spread false information – both intentional and unintentional. The ability to identify AI-generated content is crucial to ensuring people know when they are interacting with generated media, and helps prevent the spread of misinformation.
We are committed to providing people with high-quality information and maintaining trust between creators and users across society. Part of that responsibility is to provide users with more advanced tools to identify AI-generated images, so that their images – and even some edited versions – can be identified at a later date.
SynthID generates an imperceptible digital watermark for AI-generated images.
Google Cloud is the first cloud service provider to offer a tool for responsible creation of AI-generated images and their reliable identification. This technology is based on our approach to developing and deploying responsible AI and was developed by Google DeepMind and refined in collaboration with Google Research.
SynthID is not foolproof against extreme image manipulation, but it provides a promising technical approach that enables people and organizations to work responsibly with AI-generated content. The tool could also evolve with other AI models and modalities beyond images, such as audio, video and text.
Novel watermark type for AI images
Watermarks are designs that can be applied to images to identify them. From physical prints on paper to the translucent text and symbols now evident in digital photos, they have evolved throughout history.
Established watermarks are not sufficient to identify AI-generated images because they are often applied to the image like a stamp and can be easily edited. For example, circumspect watermarks located in the corner of an image can be cut out using basic editing techniques.
Finding the right balance between imperceptibility and resistance to image manipulation is challenging. Highly evident watermarks, often added as a layer with a name or logo at the top of an image, also pose an aesthetic challenge for original or commercial purposes. Similarly, some previously created unnoticeable watermarks may be lost due to straightforward editing techniques such as resizing.
The watermark is detectable even after modifications such as adding filters, changing colors and brightness.
We designed SynthID to not degrade image quality and allow the watermark to remain detectable even after modifications such as adding filters, changing colors, and saving using various lossy compression schemes – most commonly used for JPEG files.
SynthID uses two deep learning models – watermarking and identification – that have been jointly trained on a diverse set of images. The combined model is optimized for a number of goals, including correctly identifying watermarked content and improving imperceptibility by visually matching the watermark to the original content.
A tough and scalable approach
SynthID enables Vertex AI customers to responsibly create AI-generated images and confidently identify them. While this technology is not perfect, our internal testing has shown it to be true for many common image manipulations.
SynthID combined approach:
- The watermark: SynthID can add an unnoticeable watermark to synthetic images created by Imagen.
- Identification: By scanning an image for a digital watermark, SynthID can assess the likelihood of Imagen creating the image.
SynthID can facilitate assess the likelihood that an image was created by Imagen.
This tool provides three confidence levels for interpreting watermark identification results. If a digital watermark is detected, part of the image will likely be generated by Imagen.
SynthID contributes to a broad set of approaches to identifying digital content. One of the most common methods of identifying content is metadata, which provides information about who created it and when. This information is stored with the image file. Digital signatures added to the metadata can then show whether the image has been altered.
When the metadata information is intact, users can easily identify the image. However, metadata can be manually removed or even lost when editing files. Because the SynthID watermark is embedded in image pixels, it is compatible with other metadata-based image identification methods and remains detectable even in the event of metadata loss.
What’s next?
To responsibly create AI-generated content, we are committed to developing safe and trustworthy approaches at every stage – from image creation and identity to media literacy and information security.
These approaches need to be tough and adaptable as generative models advance and expand to other media. We hope that our SynthID technology will be able to work with a wide range of solutions for creators and users across society, which is why we are constantly developing SynthID, collecting feedback from users, enhancing its capabilities and discovering modern features.
SynthID can be extended for exploit in other AI models. We’re excited about the potential to integrate it into more Google products and make it available to third parties in the near future, enabling people and organizations to work responsibly with AI-generated content.
Note: The model used to create synthetic images in this blog may be different from the model used in Imagen and Vertex AI.
Watermarked image of a metallic butterfly with prismatic patterns on its wings