Google AI researchers have announced a tool that will help identify images generated by AI. At a time when AI images are becoming more realistic as the technology continues to evolve rapidly, there’s a need to develop mechanisms to identify such images. This will help fight some abuse and exploitation of said images and aid in preventing the spread of misinformation. In line with this, Google is launching a new invisible and permanent watermark in AI-generated images.
Google Deepmind, an AI research lab, explained that the tool called SynthID will embed a digital watermark directly into the pixels of an image. The digital watermark is imperceptible to the human eye but detectable making it easy to identify images generated by AI. According to the announcement, will help users identify they are interacting with an image created by an AI model and thus help prevent misinformation.
While there already exists some safeguards to identify AI generate images, some can be easily altered or edited out. This includes watermarks “printed” on such images such as Bing’s image watermark found on the edge of images which you can easily crop out with basic editing tools. Google Deepmind however warns that SynthID is not foolproof against “extreme image manipulations.”
However, the research blog explains that the watermark remains intact even after image modifications such as adding filters, changing colors, brightness and even saving with various lossy compression schemes.
Using metadata to identify images generated by AI
Another way used to detect digital content includes the use of metadata. This is the information stored in the image file that tells who created the image, and when and how the image was created.
However, metadata can be lost or removed manually making it often unreliable to identify AI-generated images. As the SynthID watermark is directly embedded in the image pixels, it remains detectable even if the metadata is lost or corrupted.
The tool is currently available in Beta for select Vertex AI customers using Google’s text-to-image generative model, Imagen.