Google DeepMind has developed a new tool called SynthID that aims to bring transparency to AI-generated images. SynthID is a watermarking and identification tool for generative art that embeds a digital watermark onto an image’s pixels, invisible to the human eye. This technology is being rolled out to a limited number of customers currently using Imagen, Google’s art generator available on its suite of cloud-based AI tools.
Generative art has faced criticism due to ethical concerns surrounding the use of artists’ work and the potential for creating deepfakes. The emergence of deepfakes was previewed by an AI image of the Pope wearing hip-hop attire, which went viral on social media. This raised concerns about the potential misuse of generative tools, such as AI-generated political ads that could cause significant damage. As a result, one of the voluntary commitments made by seven AI companies after a meeting at the White House was to watermark audio and visual content to indicate that it is AI-generated. Google is the first company to launch a system that addresses this commitment.
Although Google doesn’t provide much technical detail about SynthID’s implementation, it states that the watermark cannot be easily removed through simple editing techniques. The company emphasizes the difficulty of striking a balance between imperceptibility and robustness to image manipulations. SynthID is designed to maintain image quality while allowing the watermark to remain detectable even after modifications like adding filters, changing colors, and saving with lossy compression schemes commonly used for JPEG images.
The identification aspect of SynthID rates images according to three digital watermark confidence levels: detected, not detected, and possibly detected. By embedding the tool into the image’s pixels, Google’s system can work alongside metadata-based approaches like the one Adobe uses with its Photoshop generative features, currently available in an open beta.
SynthID comprises two deep learning models: one for watermarking and the other for identification. These models were trained on diverse images and combined to create a machine learning model optimized for objectives such as correctly identifying watermarked content and improving imperceptibility by visually aligning the watermark to the original content.
Despite its imperfections, Google describes SynthID as a promising technical approach for enabling responsible use of AI-generated content. The company believes the tool could expand to other AI models, including those generating text, video, and audio.
However, it’s important to consider the potential challenges and limitations that may accompany the adoption of digital watermarking. Hackers may engage in an arms race, attempting to overcome the watermarking technology, leading to a need for continuous updates to services that use SynthID. Additionally, the open-source nature of Stable Diffusion, a prominent generative tool, presents further obstacles to the industry-wide adoption of SynthID or similar solutions. Nevertheless, Google plans to make SynthID available to third parties in the near future to improve AI transparency across the industry.
By addressing the issue of transparency in AI-generated images, Google’s SynthID represents a significant step forward. Although it may not provide a foolproof solution against extreme image manipulations, the watermarking tool has the potential to empower individuals and organizations to work responsibly with AI-generated content. As the technology evolves, it may also extend to other types of AI-generated content, further enhancing transparency and accountability in the field.
