The image and video generative artificial intelligence (AI) tools are very impressive. In the wrong hands, however, they can quickly turn out to be very dangerous as it is sometimes difficult to tell the difference with authentic content.
The ability to identify these AI-generated images and videos is therefore essential to avoid misinformation. Some tools already exist, like AI or Not, but they are far from infallible. To go further, the Google DeepMind teams want to set up a watermark that is imperceptible to the naked eye, called SynthID.
An imperceptible watermark resistant to image modifications
This method can recall nearly undetectable signatures on sensitive company photos or files, unique to each employee (hashing), aimed at identifying the source of a possible information leak, for example. However, Google’s technology is said to be resistant to image cropping and editing.
“We designed SynthID in such a way that it does not compromise image quality and allows the watermark to remain detectable, even after modifications such as adding filters, changing colors and saving with different lossy compression systems — most commonly used for JPEGs”explains Google. “Because SynthID’s watermark is embedded in the pixels of an image, it is compatible with other metadata-based image identification methods, and remains detectable even when metadata is lost.”
This tool is currently available in beta version to a small number of users of the Google Cloud Vertex AI platform, and more specifically of the Imagen image generation AI. SynthID allows both to add a watermark and to identify the images likely to have been marked on three levels: watermark detected, watermark not detected and watermark possibly detected.
“SynthID could be extended to other AI models and we are very excited to integrate it into other Google products and make it available to third parties in the near future, to enable people and organizations to work responsibly with AI-generated content”Google announces.