Identifying AI-generated images with SynthID

New tool helps watermark and identify composite images created by Imagen

AI-generated images are becoming more and more popular every day. But how can we better identify them, especially when they are so realistic in appearance?

Today, in partnership with Google Cloud, we’re releasing a beta version of SynthID, a tool for watermarking and identifying AI-generated images. This technology embeds a digital watermark directly into the pixels of an image, making it imperceptible to the human eye but detectable for identification.

SynthID will be released to a limited number of Vertex AI customers using Imagen, one of the latest text transformation models that uses input text to create photorealistic images.

Generative AI technology is rapidly evolving, making it difficult to distinguish computer-generated images, also known as “synthetic images,” from images that were not created by an AI system.

While generative AI unlocks huge creative potential, it also brings new risks, such as allowing creators to intentionally or unintentionally spread misinformation. Being able to identify AI-generated content is important to inform people as they interact with generated media and prevent the spread of misinformation.

We are committed to connecting people with quality information and maintaining the trust of creators and users across society. Part of this responsibility is to provide users with more sophisticated tools to identify AI-generated images, allowing images (and even some edited versions) to be identified later.

Google Cloud is the first cloud provider to provide the tools to responsibly create and confidently identify AI-generated images. This technology is based on our approach to responsible AI development and deployment, developed by Google DeepMind and refined in partnership with Google Research.

Although SynthID does not fully address extreme image manipulation, it does offer a promising technical approach to help people and organizations responsibly interact with AI-generated content. . The tool could also evolve with other AI models and modalities beyond images, such as audio, video, and text.

A new type of watermark for AI images

A watermark is a design that can be placed over an image to identify it. They have evolved throughout history, from physical traces on paper to the translucent text and symbols found in today’s digital photographs.

Traditional watermarks are not sufficient to identify images generated by AI, as they are often applied like a stamp to images and can be easily edited and removed. For example, a separate watermark in the corner of an image can be cropped using basic editing techniques.

Finding the right balance between imperceptibility and robustness to image manipulation is difficult. Highly visible watermarks are often added as a name or logo layer on top of an image, but they also pose aesthetic challenges for creative or commercial purposes. Similarly, some previously developed imperceptible watermarks may be lost through simple editing techniques such as resizing.

SynthID allows watermarks to remain detectable even after changes such as adding filters, changing colors, or saving with various lossy compression schemes (most commonly used in JPEG) without compromising image quality. Designed to be maintained.

SynthID uses two deep learning models trained together on different image sets for watermark insertion and identification. The combined model is optimized based on various objectives, such as accurately identifying watermarked content and improving imperceptibility by visually aligning the watermark to the original content.

Robust and scalable approach

SynthID enables Vertex AI customers to responsibly create and confidently identify AI-generated images. Although this technology is not perfect, our internal testing shows it to be accurate for many common image operations.

SynthID’s combined approach:

Watermarks: SynthID can add imperceptible watermarks to synthetic images generated by Imagen. Identification: By scanning an image for watermarks, SynthID can assess the likelihood that the image was created by Imagen.

This tool provides three confidence levels for interpreting watermark identification results. If a watermark is detected, part of the image may have been generated by Imagen.

SynthID contributes to a wide range of approaches to identifying digital content. One of the most widely used ways to identify content is through metadata, which provides information such as who created it and when. This information is saved with the image file. Adding a digital signature to metadata can indicate whether an image has been modified.

If the metadata information is intact, users can easily identify the image. However, metadata can be manually deleted or lost when editing files. Because SynthID watermarks are embedded in the pixels of images, they are compatible with other metadata-based image identification approaches and are detectable even when metadata is lost.

What’s next?

To build AI-generated content responsibly, we are working to develop safe, secure, and reliable approaches at every step, from image generation and identification to media literacy and information security. Masu.

These approaches need to be robust and adaptable as generative models advance and are extended to other mediums. We hope that our SynthID technology can work with a wide range of solutions for creators and users across society, and by gathering user feedback, enhancing features, and exploring new features, we hope that SynthID continues to evolve.

SynthID has the potential to be extended for use across other AI models, and in the near future SynthID will be integrated into more Google services and made available to third parties to help people and organizations take responsibility. We’re excited about the potential to work with AI-generated content.

Note: The models used to generate synthetic images in this blog may differ from the models used by Imagen and Vertex AI.