Page 25 - Detecting deepfakes and generative AI: Report on standards for AI watermarking and multimedia authenticity workshop

Detecting deepfakes and generative AI: Report on standards for AI
watermarking and multimedia authenticity workshop

Jon Geater presented the SCITT framework being discussed IETF. This framework provides
a multi-layered approach that combines a few core technologies to meet most challenges:

i) Portable metadata helps create a lingua franca to describe and understand the
provenance of multimedia content across different platforms.
ii) Watermarks help to identify images, even if altered. They are also harder to remove than
metadata and thus more durable identifiers.
iii) Transparency technology helps to hold actors accountable and settle disputes.
Advantages include:

a) Broad applicability.
b) Independent storage of provenance metadata with the ability to reunite content with
its history if that history has been removed.
c) Flexible, resilient cryptographic proofs.
d) Prevents shredding and backdating, even with multiple copies of files in multiple
systems.

Touradj Ebrahimi, Professor at EPFL at Chair of JPEG, provided an overview of JPEG Trust, a new
international standard planned for publication this year. The JPEG Trust framework addresses
aspects of authenticity, provenance, and integrity through secure and reliable annotation of
images throughout their lifecycle. The JPEG Trust framework will provide building blocks for
more elaborate use cases and it is expected that the standard will evolve over time and be
extended with additional specifications.

Some of the key takeaways of this session were:

i) Tools to establish digital asset provenance and authenticity will be an important part
of solutions to the challenge of deepfake and AI-generated multimedia created with
malicious intent.
ii) Provenance data for a digital asset can be recorded through Content Credentials which
are tamperproof metadata that provide information about the origin, history, and editing
process of the content (including whether or not it was AI-generated).
iii) Authentication or provenance verification – the process of assessing content's accuracy
and consistency – can help combat misinformation and disinformation and ensure the
credibility of multimedia content.
iv) A combination of secure metadata, watermarks, fingerprinting, and secure tools for
tracking provenance history is required. C2PA, SCITT, and JPEG Trust provide mechanisms
to implement these features.
v) For Content Credentials to work, they will be needed everywhere – across all devices
and platforms – and there will need to be broad awareness of their availability and value.
vi. Standards are needed to enable the interoperability of provenance and authenticity
verification mechanisms, calling for global collaboration on the development of relevant
standards.

17