Google launches SynthID tool to add watermark function to AI-generated images

With the rapid development of generative artificial intelligence, people are increasingly paying attention to the potential risks brought by AI-generated content. For example, people may intentionally or unintentionally spread disinformation generated by AI. Especially in the field of image generation, it has become particularly important to distinguish AI-generated images from real images

There are some studies trying to distinguish AI-generated content by adding watermarks. ICML 2023's outstanding paper "Watermarking for Large Language Models" proposes a method to add watermarks to the output of the model to address the challenge of text generation

Google DeepMind announces the launch of a new tool SynthID, used to add watermarks to generated images to identify AI-generated images. Currently, the Beta version of SynthID is online. Google CEO Sundar Pichai and Google DeepMind CEO both forwarded the introduction of the SynthID tool on Twitter

Currently, SynthID can be used for Google’s Wensheng Diagram The image generated by the model Imagen adds a watermark, and the presence of the watermark is almost invisible to the eye. As shown in the figure below, there is no difference in appearance whether the image is watermarked or without watermark:

SynthID can be scanned Digital watermark in the image to evaluate whether the image was generated by Imagen

SynthID does not affect the image quality and even when adding filters, changing colors and using various lossy compression schemes such as JPEG) after saving and other modifications, the watermark can still be detected

Two deep learning models are used in SynthID, respectively for watermarks and identification. The two models have been jointly trained on multiple image sets and optimized for a range of objectives

Currently, some users can use SynthID on Google’s Vertex AI platform Tool (Beta version) to operate on Imagen

It is worth noting that SynthID does not guarantee that the watermark is completely resistant to extreme image processing, but it provides a promising technical method to make People can use AI-generated content responsibly. The research team said that similar tools may be launched in the future for other modalities such as audio, video and text

The above is the detailed content of Google launches SynthID tool to add watermark function to AI-generated images. For more information, please follow other related articles on the PHP Chinese website!