DeepMind, Google Research and Google Cloud launch the beta version of a new tool for watermarking and identifying AI images.

Google is launching a tool to watermark and identify AI images. SynthID is the name of the new feature developed by DeepMind, Google Research and Google Cloud.

Your ID, please

With SynthID it is possible to integrate a watermark directly into the pixels of an image. For example, it is not visible to the human eye, but a system can identify the image.

Traditional watermarks are not enough for AI images, they are too easy to remove. Google firmly believes that SynthID does not degrade image quality and that a watermark is not affected by filters, color changes, or compressing an image by resizing it.

Part of the identification takes place via the metadata of an image. However, they are relatively easy to remove. However, the SynthID watermark should remain as it is built into the pixels.

Currently, SynthID is available to a limited number of Vertex AI customers who are also using Imagen. The latter is one of DeepMind’s newest AI models for converting text into images.

keeping up with the times

Generative AI is getting smarter and evolving rapidly. This also applies to image generators that work with artificial intelligence. AI images can look more and more realistic and sometimes it becomes difficult to distinguish them from real images. Therefore, such a watermark is a step in the right direction.

According to Google, SynthID is not foolproof when it comes to extreme image manipulation. However, the company emphasizes that it offers users and organizations a technical solution to work more responsibly with AI images. In addition, the tool can be further developed with other AI functions.

To achieve this, SynthID trains with two deep models for watermarking and identification. Both were trained together on a series of images. This merged model is optimized to make a watermark as unobtrusive as possible by fitting it as closely as possible to the image in question. This also increases the possibility of identification.

Possibilities

Imagen users get three levels of trust with SynthID:

• Digital watermark detected
• No digital watermark detected
• A digital watermark may have been detected

Google is already dreaming of future applications for SynthID. The tech giant has already hinted that the tool can be expanded to other AI models and integrated with other Google products. Providing SynthID to third parties can also be an option.

With DeepMind, Google has already created other eye-catching AI applications in recent months. Late last month saw RT-2 combining robotics and AI, and earlier this month it was announced that it was working on several dozen AI projects simultaneously.