Better
Together: Joining Forces on Digital Media Provenance
February 9, 2024 | Eric Horvitz - Chief Scientific Officer,
Microsoft
|
N |
o
single solution exists to confront the complex challenge of deceptive online
content, whether that content is handcrafted or AI-generated. Thwarting
motivated adversaries will take ongoing vigilance and a multifaceted approach
in technical and sociotechnical realms. One important technology is digital media provenance, which has been
in the news this week: OpenAI, Meta, and Google announced that they will join Microsoft, Adobe,
BBC, Sony, Truepic, Witness, and many other organizations in embracing Content
Credentials, C2PA’s technical standard for digital media provenance.
Media
provenance methods were developed several years ago with the explicit goal of helping people to grapple with the rise of AI-generated content.
A good way to think about the workings of media provenance is to imagine a
digital world where each piece of content is verified with a seal, akin to a
digital version of a 'wax seal' on the back of an envelope—a symbol of its
authenticity and unaltered state of its contents. Media provenance, leveraging
cryptographic methods, offers just that. The goal is to ensure that the
information you consume, like a news report or an image, is not just authentic
but also originates from a trusted source and remains unchanged from its
production to your screen. Microsoft currently supports media provenance in
Microsoft Designer and Designer in Copilot.
The
development of effective media provenance tools depends on delivering a widely
embraced standard. Three years ago, two outstanding efforts that focused on
implementing media provenance, Project Origin and the Content Authenticity Initiative, came together to create a
standards organization named the Coalition for Content Provenance and
Authenticity (C2PA). The organization
now includes nearly 100 companies, and supports a broader ecosystem of over
2,000 organizations, spanning camera manufacturers, content producers, major
technology companies, and NGOs. The wide acceptance and continuing collaboration
across many groups is an important step forward in bringing consistent content
integrity protocols to the public.
Media
provenance can be combined with other approaches, such as steganographic watermarking, to add
another layer of protection. This invisible watermarking approach is based on
hiding information in human-imperceptible locations of the content. That
information is later retrieved by use of a watermark detector, providing a way
to share information about content that is invisible to the end user. Watermark
technologies can be resistant to a range of alterations, including taking
screenshots of content, and can be used to provide pointers to publicly
available provenance information. However, attackers can manipulate an
image until a detector can no longer find the mark, which is why using
watermarking in conjunction with media provenance is promising.
While
media provenance and watermarking technologies offer significant strides in
authenticating content, they are not impervious to manipulation. This current
reality calls for a multifaceted approach, combining these technologies with a
spectrum of other approaches. Directions include developing techniques
for recognizing and mitigating attempts to create deceptive or offensive
content at the point of creation, (including the use of AI methods themselves),
ongoing vigilance and fast-paced responses to malevolent uses, and sensitive,
balanced regulations that disincentivize deceptive uses and help to establish
norms and expectations.
The
battle against online deception will be dynamic and ever evolving, requiring
more than just a single technological breakthrough. It necessitates continuous
collaboration and innovation across both technical and sociotechnical fronts.
Essential to this endeavor is the active involvement of a diverse array of
stakeholders, including technology firms, content creators, academic
researchers, civil society organizations, and policymakers. By uniting these
varied perspectives and expertise, we fortify our ability to stay ahead in the
struggle against online deception.
Additional
background