Source: https://www.imatag.com/blog/6-examples-of-uses-of-digital-watermarking
Thanks to high-quality AI tools, people are now creating more and more deepfakes. They can come in the form of images, voice cloning, fake video content, and much more. This trend can have severe consequences for the financial system, democracy, the judicial system, and reputation. Passive defenses against synthetic media involve passive digital forensics and traditional machine learning methods that detect forged content by analyzing statistical irregularities in digital media. In practice, passive checks are often not enough due to machine learning models constantly improving and eliminating artifacts. The most promising direction in the research of active authentication of digital media is the cryptographic standard of provenance called C2PA (Coalition for Content Provenance and Authenticity), which uses standardized metadata manifests, and the method of invisible digital watermarks known as SynthID from Google. Proactive content provenance tools offer the best protection against AI fraud by cryptographically signing content, collecting information about subsequent changes to digital media, and storing this information in a special manifest. Watermarks, on the other hand, are embedded directly into the pixels of images, sounds, or video frames during content creation, thereby providing an additional level of security and privacy. Both watermarking methods and those based on manifests have advantages and weaknesses. The C2PA standard is an open standard that allows you to embed information about the origin of digital media in files in the form of JUMBF manifests. This information can be used, for example, to cryptographically verify the source of tweets or determine which editing tools were used after the original was created. One of the problems with such methods is that manifests can be removed during normal processing of files. When a user uploads an image to a social network that does not support the C2PA standard, or when a file is converted from one format to another, or when an image is captured on a screen, manifests may be lost. One way to protect against such attacks is to use not only manifest files but also to hide digital watermarks in images. Using a similar principle, invisible watermarks can be hidden in audio files. Watermarking tools typically work with raw pixels, audio samples, or video frames, changing some of their characteristics to values that are not visible to the eye or ear. Unlike manifest files, watermarks are hidden in the content itself, so they are much more difficult to remove when converting files from one format to another or reducing resolution. These two technologies fight different types of attacks, and their combined use provides more reliable protection. Both watermarking and manifest file technologies are not fully protected against attacks. Attackers may use various methods to circumvent the protection provided by such tools. One of the ways to protect against AI fraud is to exclude such files from being used in the creation of deepfakes in the first place by removing all signs of manipulation. However, this requires additional steps from the user when publishing photos or videos on websites that do not support such standards. One of the simplest ways to protect against such attacks is to use content provenance tools that use both manifest files and invisible digital watermarks. In the case of copy-paste attacks, content verification tools will show that the manifest file does not match the actual image. However, a malicious actor can copy real content credentials from a legitimate image and superimpose them on a fake image. To protect against such attacks, one must always check the integrity of manifests and ensure that they were not forged. The integrity of the manifest can be checked, for example, by cryptographic signing in the C2PA standard.
Sources
https://openai.com/index/advancing-content-provenance/ https://www.institutepm.com/knowledge-hub/ai-content-provenance-watermarking https://ori.hhs.gov/content-provenance SynthID & C2PA Content Credentials overview.