The rise in generative artificial intelligence (GAI) technology capable of creating realistic fake media, coupled with the U.S. government pulling back efforts to counter sophisticated digital threats, presents challenges as to what digital content consumers can trust and has afforded space for adversaries to intensify disinformation efforts against the United States. Therefore, the private sector may need to lean into the responsibility of authenticating digital media and establishing systems to maintain trust.
One such effort is the Coalition for Content Provenance and Authenticity (C2PA), constituted of major stakeholders in the field of digital content like Adobe, New York Times, and Google. The coalition worked to establish technical standards for digital content created, edited, or published in the C2PA “ecosystem.” C2PA acts by adding metadata, or the unseen information within an image or video file, thus digitally “marking” the content. This approach uses mathematical techniques to encode a digital record of the history of the content, such as the media's content, authors, and date, time, or place of creation, with the content itself.
The C2PA ecosystem has been touted as a solution to countering misinformation, with many calling for media companies to opt-in to abiding by C2PA technical standards. However, C2PA is far from a complete solution as it falls short of addressing many of the most pressing risks currently threatening media, including GAI content. This is because success of the C2PA ecosystem relies on end-to-end compliance of all image creation elements—from point of capture to posting, which was realistic when C2PA was originally conceived as a “closed ecosystem” for journalist or content creators. As the C2PA system has expanded, the expectation that everyone in the C2PA ecosystem will uphold these designated standards is questionable.
The private sector may need to lean into the responsibility of authenticating digital media and establishing systems to maintain trust.
C2PA is now offered to any creator or sharer of digital images or video content users, making it an “open ecosystem.” In such an environment participants may not understand or be aware of C2PA's operation, and content could unwittingly exit the C2PA ecosystem during content processing, resulting in perfectly valid content with invalid or missing C2PA data. When this content exists alongside content created using non-C2PA tools it becomes difficult for users to interpret. Upon receiving content lacking C2PA credentials, a media consumer now wonders: was the content never marked, or was the mark lost during a mishap with a noncompliant tool? Such questioning creates a perfect setting for bad actors to undermine confidence in the system: if obviously real content lacks valid marks, what trust should I have in the system?
C2PA, like other systems, suffers from a lack of rigorous security analysis of its properties against a well-defined goal. Such analyses are known as threat modeling, which serves as a form of “risk assessment that models aspects of the attack.” Threat modeling identifies a potential adversary's goals and methods for system disruption and then develops necessary mitigation techniques to combat the designated threat. Threat modeling is important in cyber communities and recommended by NIST as an essential step (PDF) in software development processes. C2PA has taken steps to embrace this practice, but their threat model is outdated, generic, and lacks direction. C2PA's threat model has not been updated since its 1.0 version, released January 2022, despite the specification itself changing. C2PA's model does not examine threats in systems designated as open or closed, and it is unclear if the model or security guarantees provided are equally applicable to both. Finally, the model fails to consider current security threats, including scenarios relevant to emerging uses such as GAI.
To improve the utility and security of C2PA and better position the standard within the modern environment including GAI threat, the consortium might consider several steps:
- C2PA could improve its threat modeling efforts by concretely and robustly defining the uses and threats it does (and does not) address. Just as security is not a fixed state, threat modeling activities will require ongoing and regular considerations to maintain and adapt. The consortium should increase transparency regarding the level of trust content consumers have in C2PA information, including managing expectations and addressing information threats, like those posed by GAI.
- C2PA needs a body of academic literature examining the plausibility and potential issues, including the threat model's security requirements and technical specifications, with this system before it can be a viable option for digital content authentication.
- C2PA may consider breaking apart into user/publishers in a closed system separate from an open system, with separate security requirements. A single standard is likely not realistic, considering the diverse set of users, e.g., government, military, insurance, consumer applications, and their digital content goals.
C2PA, like other systems, suffers from a lack of rigorous security analysis of its properties against a well-defined goal.
In the current digital information environment, significant benefits can come from private sector efforts to spearhead the adoption of strategies authenticating digital content and distinguishing it from GAI content. C2PA presents a promising avenue for addressing many challenges, yet it necessitates thorough evaluation and refinement. Only through careful consideration can C2PA be implemented as a comprehensive solution for marking authentic digital content.
