Machine Heart report
Heart of the Machine Editorial Office
Can you tell if the image is AI-generated or real?
With the rapid development of generative AI, there is increasing concern about the potential risks posed by AI-generated content. For example, people may knowingly or unintentionally spread AI-generated false information. Especially in the field of image generation, it is especially important to distinguish between AI-generated images and real images.
Some studies have tried to "watermark" AI-generated content. ICML 2023 Outstanding Paper "A Watermark for Large Language Models" proposes to add watermarking to the output of the model for text generation.
Now, Google DeepMind has announced the launch of SynthID, a watermarking tool for image generation to identify AI-generated images, and the beta version of SynthID is currently live. Google CEO Sundar Pichai and Google DeepMind CEO both retweeted about the SynthID tool.
SynthID can currently add watermarks to images generated by Google's Vensen graph model, Imagen, and are visually imperceptible. As shown in the image below, images with or without watermarks look indistinguishable:
By scanning the digital watermark in the image, SynthID can assess the likelihood that the image was generated by Imagen.
SynthID does not affect image quality, and watermarks can still be detected even after modifications such as adding filters, changing colors, and saving with various lossy compression schemes such as JPEG.
SynthID uses two deep learning models (for watermarking and recognition) that have been trained together on multiple image sets and optimized for a range of targets.
Currently, some users can use the SynthID tool (beta) on Imagen on Google's Vertex AI platform.
Video loading...
It's worth noting that SynthID doesn't guarantee that watermarks will be foolproof against extreme image processing, but it offers a promising technological approach that enables people to responsibly use AI-generated content. The research team says similar tools may be rolled out later for audio, video, and text modalities.
Reference link: https://www.deepmind.com/blog/identifying-ai-generated-images-with-synthid