/ground-report/media/post_banners/wp-content/uploads/2023/08/Source-Canva-23-2.jpg)
Google introduces SynthID tool, adding digital watermark to AI-generated images for easy identification. According to a DeepMind blog post, SynthID directly adds an identifier into the pixels of an image, making it imperceptible to the human eye. The watermark does not compromise image quality and allows for subsequent editing without losing its effectiveness. This implies that adding filters, cropping the image, or converting it to another format will retain the embedded markup within the file.
In recent months, the necessity of implementing a watermark on images created by artificial intelligence has emerged. Models such as Midjourney or Stable Diffusion have been making significant progress, producing realistic content that can deceive the average user. For instance, the viral photos of the arrest of Donald Trump or Pope Francis wearing a padded coat were so convincing that Midjourney had to discontinue their free trials.
Not only can you add a watermark using SynthID, but you can also scan an image to check if it is a product of generative AI. The tool uses two deep learning models: one for watermarks and another for identification. DeepMind engineers trained them with a diverse image set. The identifier visually aligns with the original content, determining its presence.
New type of watermark for AI images
SynthID uses three levels of confidence to inform the user when presenting its results. It is worth mentioning that the tool is not foolproof against extreme image manipulation. However, Google will continue to improve it and could apply it to other types of content, such as audio, video, or text.
A limited number of people will have access to SynthID by using Image, a Vertex AI model capable of creating photorealistic images from text input. Photoshop supports digital watermarking through metadata-based image identification methods.
"Traditional watermarks alone can't identify AI-generated images, as they're usually stamped on and easily editable. We designed SynthID to not compromise image quality and allow the watermark to remain discoverable, even after modifications."
Google confirmed that it intends to extend it to third-party generative AI models in the near future. The technology company also promised to maintain the development of SynthID to include new improvements in future versions. The testing phase will serve to collect user feedback and refine the models in the future.
In discussions on regulating this technology, a critical matter involves identifying AI-generated content. Politicians and organizations are both pushing for systems to detect synthetic images or texts potentially utilized in disinformation campaigns. Some governments are already utilizing this strategy to alter their inhabitants' perception with deepfakes, while others are contemplating employing AI in campaigns to destabilize their opponents.
Robust and scalable approach
Vertex AI customers can use SynthID to create AI-generated images responsibly and confidently identify them. Although this technology isn't flawless, our internal testing demonstrates its accuracy in detecting many common image manipulations.
SynthID's combined approach:
- Watermarking: Imagen can add an imperceptible watermark to synthetic images produced by SynthID.
- Identification: SynthID can assess the likelihood of Imagen creating an image by scanning its digital watermark.
/ground-report/media/post_attachments/b81cfc1473f31e3abbec0be2555bc1af8f52c3b71532f3a08868dd283305c208.png "publive-image")
This tool interprets the results of watermark identification by providing three confidence levels. If Imagen detects a digital watermark, it is likely that part of the image is generated by it.
SynthID is a contributor to the broad suite of approaches used for identifying digital content. One of the most widely used methods of identifying content is through metadata, which reveals information such as the creator and the creation date. The image file stores this information. Digital signatures, once added to metadata, can indicate whether an image has undergone any alterations.
When the metadata information is intact, users can easily identify an image. However, users can manually remove or even lose metadata when they edit files. SynthID implants watermark within image pixels, aligning with metadata-free identification approaches while maintaining detectability.
Keep reading
Follow Ground Report for Climate Change and Under-Reported issues in India. Connect with us on Facebook, Twitter, Koo App, Instagram, Whatsapp and YouTube. Write us on [email protected].
