The giant Google is adorning itself with a revolutionary new tool to identify virtual images generated by artificial intelligence.
Google DeepMind, a subsidiary devoted to artificial intelligence from the search engine firm, announced a few days ago the launch of the beta version of SynthID, in partnership with Google Cloud. This innovation can be described as a revolutionary AI-generated image watermarking and identification tool. The pioneering technology embeds a digital watermark directly into image pixels, making it invisible to the human eye while remaining detectable to identification systems.
SynthID, a response to the challenges of AI-generated images, including fake news
The growing popularity of AI-generated images raises a major challenge for authenticity and identification. It sometimes becomes difficult to disentangle the true from the false, and Google has already been working for several months to help its billions of users see things more clearly. With technology advancing rapidly, this AI-created content is becoming almost indistinguishable from human creations and shots, although Nikon recently proved that the real has no equal yet by teasing the Midjourney generator. To top it off, this advancement comes with new risks related to misinformation.
Here, Google Cloud is committed to maintaining the quality of information online and maintaining trust between creators and users. With this in mind, SynthID offers an advanced solution to identify AI-generated images, including their versions that may have been altered. On this last point, the technique is not yet perfected, but the first results are promising.
SynthID was developed by Google DeepMind and perfected in collaboration with Google Research. It represents the first tool of its kind to both responsibly create and identify AI-generated images, supporting Google Cloud Division’s commitment to more responsible artificial intelligence.
Currently, the SynthID tool is limited to a limited number of users of Vertex AI, a machine learning platform (formerly Google AI Platform). This makes it possible to develop specific models and generative AIs. In this case, the users we were talking about must use Imagen, an image generator from Google that works from a text query.
A digital watermark to distinguish the images, and a vision for the future?
The SynthID module can therefore spot invisible watermarks inserted into AI-generated visuals. These watermarks are patterns that can be overlaid on images to identify them. We could also compare them to the physical fingerprints found on printed photos. Because traditional watermarks can be easily removed, Google has added to SynthID the ability to digitally watermark images that Imagen produces that is imperceptible, but still detectable. The goal ? Preserve the quality of the image while leaving the possibility of detecting it, even in the event of modification, addition of filters, change of colors or others.
And while SynthID isn’t infallible against all image manipulation, it does represent a breakthrough with great potential. Google explains that the technology could be extended to media other than images, such as video, audio and text. It would thus become a first-rate all-round solution for identifying digital content.
SynthID marks the first step towards a more transparent content landscape, in which images generated by artificial intelligence can be verified with precision. By remaining open to evolution and extension to other AI models, SynthID could symbolize a step forward in the fight against the creation and dissemination of false information, and perhaps restore the confidence of users, who are in great need.
Source : Google DeepMind
1