Google and DeepMind, the AI research division of Google, have announced the launch of a new tool that can watermark and identify images that are generated by AI models. The tool, called SynthID, is designed to help users and creators of AI-generated content to work with it responsibly and prevent the spread of misinformation.
What is SynthID and how does it work?
SynthID is a tool that embeds a digital watermark directly into the pixels of an image, making it invisible to the human eye but detectable by an algorithm. The watermark is created by modifying some pixels of the original image in a subtle way, creating an embedded pattern that can be recognized by a dedicated AI detection tool.
SynthID uses two neural networks, one for watermarking and one for identifying, that were trained together on a diverse set of images. The tool can detect the watermark even after the image has been edited, cropped, resized, or compressed, as long as the modifications are not too extreme.
SynthID is available in beta for select users of Vertex AI, Google’s platform for building AI apps and models. The tool only supports Imagen, Google’s text-to-image model, which is exclusively available in Vertex AI.
Why is SynthID important?
SynthID is one of the first watermarking tools for AI-generated images that has been publicly launched by a major tech company. Google and DeepMind say that the tool is part of their commitment to developing responsible and ethical AI technologies.
AI-generated images have many potential applications, such as creating art, enhancing photos, generating captions, and more. However, they also pose some risks, such as enabling creators to spread false or misleading information, infringe on intellectual property rights, or harm people’s privacy or reputation.
By using SynthID, users and creators of AI-generated images can signal that their content is not real and provide a way for others to verify its origin and authenticity. This could help empower people with knowledge of when they are interacting with generated media and prevent the misuse or abuse of AI technologies.
What are some other watermarking techniques for AI-generated images?
Watermarking techniques for AI-generated images are not new. Several researchers and startups have been working on developing similar methods to protect or identify generative art.
For example, Imatag, a French startup launched in 2020, offers a watermarking tool that it claims is not affected by resizing, cropping, editing or compressing images. Another firm, Steg.AI, uses an AI model to apply watermarks that survive resizing and other edits.
Watermarking techniques are also being explored for other types of AI-generated content, such as audio, video, and text. For instance, Microsoft has developed a tool called Project Origin that can embed digital signatures into synthetic media files to verify their source and integrity.
What are some challenges and limitations of watermarking AI-generated images?
Watermarking AI-generated images is not a perfect solution. There are some challenges and limitations that need to be addressed before it can become widely adopted and effective.
One challenge is the lack of standardization and interoperability among different watermarking tools and models. There is no universal format or protocol for embedding or detecting watermarks in AI-generated images. This means that different tools may not be compatible with each other or may produce conflicting results.
Another challenge is the possibility of adversarial attacks or countermeasures that could remove or alter the watermarks in AI-generated images. For example, some researchers have shown that it is possible to use another AI model to erase or fool the watermark detection algorithm. This could undermine the reliability and security of the watermarking system.
A third challenge is the ethical and legal implications of watermarking AI-generated images. There are some questions that need to be answered, such as who owns the rights to the watermarked images, who is responsible for their use or misuse, and how to balance the trade-offs between transparency and creativity.
