Identification of images generated by AI with a synthesizer

The new tool helps with a watermark and identify synthetic images created by Imagen

Images generated by AI are becoming more and more popular. But how can we identify them better, especially when they look so realistic?

Today, in cooperation with Google CloudWe run the beta synthid version, a water marking tool and identification of images generated by AI. This technology embeds the digital watermark directly in the image pixels, which makes it imperceptible to the human eye, but detectable for identification.

Synthid is released to a limited number Vertex AI Customers use PictureOne of our latest text models for the image, which uses the input text to create photorealistic images.

AI generative technologies quickly evolve, and images generated by a computer, also known as “synthetic images”, become more difficult to distinguish from those that were not created by the AI ​​system.

Although generative artificial intelligence can unlock huge creative potential, it also presents a new risk, such as enabling creators to disseminate false information – both intentionally and unintentionally. The possibility of identifying the content generated by AI is of key importance to strengthening the position of people of knowledge about when they interact with generated media and to prevent the spread of disinformation.

We are involved in joining people with high -quality information and maintaining trust between creators and users throughout society. Part of this responsibility is to provide users with more advanced tools for identifying images generated by AI, so that their images-even some edited versions-could be identified at a later date.

Google Cloud is the first cloud supplier that offers a tool for responsible creating images generated by AI and identifying them certainly. This technology is based on our approach to developing and implementing responsible artificial intelligence and was developed by Google Deepmind and improved in cooperation with Google Research.

Synthid is not reliable against extreme image manipulations, but provides a promising technical approach to strengthen the position of people and organization to work with the content generated by AI. This tool can also evolve with other AI models and methods outside of images such as sound, video and text.

A new type of watermark for AI images

Water marks are projects that can be arranged in the paintings to identify them. From physical imprints on paper to a translucent text and symbols visible in digital photos, they have evolved throughout history.

Traditional water signs are not enough to identify images generated by AI, because they are often used as a stamp in the image and can be easily edited. For example, discreet water signs found in the corner of the image can be selected using basic editing techniques.

Finding the right balance between imperceptibility and resistance to image manipulation is difficult. Highly visible watermarks, often added as a layer called or logo at the top of the image, are also aesthetic challenges for creative or commercial purposes. Similarly, some previously developed unnoticeable watermarks can be lost using simple editing techniques such as size.

We designed Synthid so that it does not threaten the image quality and allows the water marker to remain detectable, even after modifications, such as adding filters, color change and saving with various lossy compression schemes – most commonly used for JPEG.

Synthid uses two deep learning models – for water marking and identification – which have been trained together on a variety of images. The combined model is optimized in terms of a number of targets, including a properly identifying water marker content and improving imperceptibility through visual adaptation of the watermark to the original content.

A solid and scalable approach

Synthid allows Vertex AI clients to create images generated by AI and identifying them with confidence. Although this technology is not perfect, our internal tests show that it is accurate compared to many typical image manipulation.

Connected Synthid approach:

• Perfectly aquatic: Synthid can add an imperceptible watermark to synthetic images produced by Imagen.
• Identification: By scanning the image for a digital watermark, Synthid can assess the likelihood of image creation by image.

This tool provides three trust levels to interpret the results of identification of the watermark. If a digital watermark has been detected, part of the image is probably generated by Imagen.

Synthid contributes to a wide set approach to the identification of digital content. One of the most commonly used content identification methods are metadata, which provide information such as the one who created it and when. This information is stored in the image file. Digital signatures added to metadata can then show whether the image has been changed.

When information about metadata is intact, users can easily identify the image. However, metadata can be manually deleted and even lost after editing the files. Because the Synthid watermark is built into the image pixels, it is compatible with other image identification approaches based on metadata and remains detectable, even when the metadata is lost.

What next?

To responsibly build content generated by AI, We are involved in developing safe, safe and trustworthy approaches At every stage – from generating image and identification to reading skills and media skills and information security.

These approaches must be solid and flexible, because generative models develop and develop to other media. We hope that our synthesizer technology can work with a wide range of solutions for creators and users throughout society, and we are still evolving Synthid, gathering feedback from users, increasing its capabilities and exploring new functions.

Synthid can be extended to use in other AI models and we are excited about the potential of integration with more Google products and sharing the third parties in the near future-by-ordering people and organizations for responsible work with content generated by AI.

Note: The model used for the production of synthetic images on this blog may be different from the model used for Imagen and Vertex AI.