DeepMind launches SynthID: an AI image recognition tool exclusively for Google

2023-08-30

Google's DeepMind has announced that they have solved a problem that has plagued image verifiers, which is determining whether an image is real or generated by artificial intelligence. The researchers claim that their new watermark format, SynthID, can accurately identify deepfake images generated by AI without compromising the original quality of the image. However, the program is currently only applicable to Google's local image generation system. In a blog post, DeepMind stated that the beta version of the SynthID system embeds watermarks into the pixels of AI-generated images, making them imperceptible to the human eye but detectable by detection tools. SynthID is designed for companies using Google's Vertex AI platform, which means it is only available to large enterprise partners using Google's Imagen model. DeepMind, the department recently merged with Google's Brain AI team, did not provide detailed information on how this watermark technology works. The researchers claim that SynthID retains the watermark even after resizing, compressing, or applying color filters to the image. Additionally, this secretive technology will remain confidential to prevent external parties from bypassing it. DeepMind CEO Demis Hassabis told The Verge that if the system proves effective, the company will eventually share the technology with "partners." However, this means that the watermark exists within a closed loop of Google's ecosystem. The system can also scan images to evaluate whether they were created by Imagen, but it is difficult to determine the usefulness of this feature since the number of AI image generators is unknown. Watermarks cannot prevent the spread of misinformation The Biden administration has emphasized the need for companies to find a way to identify AI-generated images. However, watermarks alone cannot solve the problem of deepfake and misinformation generated by AI. Researchers point out that most AI detection software is inadequate, especially when many companies are trying to promote their proprietary AI platforms. Hany Farid, a computer science professor at the University of California, Berkeley, told Scientific American this month that watermarks are merely a "mitigation strategy" against the harm caused by AI deepfakes. Gizmodo magazine has discussed methods to counter AI-manipulated images with other engineers, but critics argue that there is no foolproof method. Whenever a creator comes up with a method to protect images from distortion or mark AI-generated images, others may find ways to bypass it. Furthermore, watermarks pose a particular problem. As AI researcher Claire Leibowicz recently pointed out in an article in MIT Technology Review, watermarks can only help uncover potentially manipulated images before AI-generated content, such as false images of the Pentagon exploding, spreads on the internet. Watermarks can be removed, some more easily than others, but ultimately, internet sleuths will find ways to bypass them. Even DeepMind quickly acknowledged that this method is "not 100% reliable in extreme image processing" but rather a way for organizations to "responsibly use AI-generated content." Google announces more AI additions to Workspace Google has announced a plethora of new AI features at its Cloud Next conference. Firstly, Google has integrated its Duet AI into chats. Duet is similar to Microsoft's Copilot used in its 365 applications. The company states that Duet can gather information from users' Drive and Gmail accounts, as well as other applications like Slides and Sheets. Users can use AI to summarize documents or conversations in shared meeting rooms. These features are similar to those planned by Zoom and Slack. Google's Workspace Duet AI is now widely available for companies interested in trying it out. In addition to its own services, Google will also start offering its cloud platform to other major AI companies. Meta's Llama 2 language model and Anthropic's Claude 2 chatbot will be available through Google Cloud software. Llama 2 is already available on Microsoft's Azure cloud platform, so Google is making efforts to keep up with the latest commercial applications of AI in many aspects.

RECENT AI NEWS

RECENT AI TOOLS