Identifying AI-generated images with SynthID

The new tool helps a watermark and identify the artificial images created by Imagen

The images created from artificial intelligence have become more popular every day. But how can we get to know it better, especially when it looks very realistic?

Today, in partnership with Google Cloud, we launch an experimental version of Synthid, a watermark tool and define images created by artificial intelligence. This technology destroys a digital watermark directly in the pixel units of the image, making it imperceptible to the human eye, but it can be detected to identify identity.

Synthid is released to a limited number of Airex AI customers using Imagen, which is one of the latest text models to the image that uses the entry text to create realistic images.

The techniques of the artificial intelligence develops rapidly, and the images created as a computer, also known as “artificial images”, have become more difficult to distinguish between those that were not created by the artificial intelligence system.

While the obstetric intelligence can open huge creative capabilities, it also represents new risks, such as enabling creators to publish wrong information – intentionally or unintentionally. The ability to identify the content created by artificial intelligence is very important to enable people to know when interacting with the created media, and to help prevent the spread of wrong information.

We are committed to linking people with high -quality information, supporting confidence between creators and users across society. Part of this responsibility gives users more advanced tools to determine the AI’s images so that their photos can be identified-even some versions that have been edited-later.

Google Cloud is the first cloud provider to present a tool to create images created by artificial intelligence with responsibility and identify them with confidence. This technology is based on our approach to the development and dissemination of responsible artificial intelligence, developed by Google DeepMind and its refining in partnership with Google Research.

Synthid is not guaranteed against extremist image manipulation, but it provides a promising technical approach to enabling people and institutions to work with the content created by artificial intelligence responsibly. This tool can also develop alongside other Amnesty International models and methods outside images such as sound, video and text.

A new type of watermark for artificial intelligence images

Water signs are designs that can be applied to the images to select. From material fingerprints on paper to the transparent text and symbols that were seen on digital images today, have evolved throughout history.

Traditional watermarks are not enough to determine the images created by artificial intelligence because they are often applied like a seal on an image and can be easily edited. For example, separate watermarks in the image angle with basic editing techniques can be held.

It is difficult to find the right balance between ability and durability to manipulate pictures. Very visible aquatic signs, often added as a layer bearing a name or logo across the top of the image, and also provides aesthetic challenges for creative or commercial purposes. Likewise, some improper watermarks can be lost by simple liberation techniques such as changing their size.

We designed Synthid so that it does not weaken the image quality, and allows the water mark to remain discovered, even after adjustments such as adding filters, changing colors, and providing different pressure schemes – the most common in JPEGS.

Synthid uses two deep learning models – for water marks and definition – that are training together on a variety of images. The built -in model is improved on a set of goals, including properly defining water content and improving the ability to visually align the water mark with the original content.

A strong and developmental approach

Synthid allows AIREX AI to create images created by artificial intelligence responsibly and identify them with confidence. Although this technology is not perfect, our internal test indicates that it is accurate for many common pictures.

Common Synthid approach:

• watermarkSYNTHID can add an imperceptible watermark to the IMAGEN artificial images.
• identification: By wiping an image of digital watermark, Synthid can evaluate the possibility of creating an image with Imagen.

This tool provides three levels of confidence to explain the results of the watermark identification. If a digital watermark is discovered, a portion of the image is likely to be created by imagen.

Synthid contributes to a wide range of methods to determine digital content. One of the most widely used ways to identify content is through descriptive data, which provides information like those who created it and when. This information is stored with the image file. Digital signatures added to the descriptive data can appear if a picture has been changed.

When descriptive data information is sound, users can easily determine an image. However, descriptive data can be manually removed or lost when the files are edited. Since the water mark in Synthid is built into the pixel units, it is compatible with other images that are based on descriptive data, and are still discovered even when descriptive data is lost.

What next?

To build content created by artificial intelligence responsibly, we are committed to developing safe, secure and trustworthy methods in every step on the road-from generating images and identifying identity to literacy and information security.

These methods should be strong and adaptable as obstetric models advance and expand to other means. We hope that our SYNTHID technology can work with a wide range of solutions for creators and users across society, and we continue to develop Synthid by collecting comments from users, enhancing its capabilities, and exploring new features.

Synthid can be expanded for use through other models of artificial intelligence, we are excited about the possibility of combining them in more Google products and making them available to third parties in the near future-empowering people and institutions to work responsibly with the content created from artificial intelligence.

Note: The model used to produce artificial images in this blog may differ from the model used in Imagen and Vertex Ai.