Global initiative to combat deepfakes
Anyone who follows the news regularly might be forgiven for thinking that deepfakes — computer-generated clips that not only distort reality but could also destroy the reputations of their victims — only affect show business celebrities and politicians. Alessandra Sala, an engineer and AI scientist, is at pains to point out that malicious actors generate fake content to harass, intimidate and spread misinformation about ordinary people too. “When fake media attacks a person that doesn’t have the economic or social means of a Taylor Swift or a Donald Trump, that’s a real problem and usually the victims are women,” she says.
The evidence bears her out. According to Sensity AI, a research company, 90%-95% of deepfake videos are nonconsensual pornography and 90% of the victims are women. Sala, who works for one of the world’s best-known providers of stock media content, explains that she first became aware of the “gender issue” when she was on a UN conference panel: “A journalist from the Middle East was talking about her experience of coming under attack from a fake video depicting her with her boss to discredit her reputation and her competency. It destroyed her emotional sphere. Her family was in shock.”
While the effects for individuals can be devastating, deepfakes pose a wider threat to society. According to MIT Technology Review, deepfakes are “a perfect weapon for purveyors of fake news who want to influence everything from stock prices to elections.” The World Economic Forum’s Global Risks Report identifies fake media as one of our most pernicious challenges. According to the WEF, the proliferation of disinformation has eroded trust not only in digital platforms, but also in institutions. For WITNESS, a non-profit which helps people to use video and technology to protect and defend human rights, it is threatening their mission of teaching people the basics of video production, safe and ethical filming techniques, and advocacy strategies.
Sam Gregory, the Executive Director of Witness, says malicious actors often weaponize people’s uncertainties to dismiss real media as fake. “We get quite a lot of cases where someone is basically claiming that AI has been used to manipulate an audio clip or a video. And it’s often quite hard to do the forensic analysis to really prove it’s real.”
The role of standards in detecting deepfakes
WITNESS is one of several organizations that have joined forces with Sala to apply standards — internationally agreed best practices — to detect deepfakes that use generative AI. Under the leadership of the three international standards organizations, IEC, ISO and ITU, and chaired by Sala, the collaboration is leveraging the insights and expertise of a broad range of stakeholders. Big tech is represented by the likes of Adobe, Microsoft and Shutterstock. Other standards organizations taking part include the Content Authenticity Initiative (CAI), the Coalition for Content Provenance and Authenticity (C2PA) and the Internet Engineering Tast Force (IETF). Also involved are the German research institute, Frauenhofer, the Swiss Federal Institute of Technology in Lausanne (EPFL) and CAICT, a technology-focused think tank based in China, as well as the authentication specialists, DataTrails and Deep Media.
“We must push standards that can create safeguards to protect our ability to trust the digital space, to look at the news and understand whether we can trust it or not,” argues Sala. “Today, by default, we tend not to trust anything anymore. People don’t trust the authorities and they’ve stopped trusting journalists who have dedicated their lives to bringing us the truth.”
Sala says international standards can help to re-establish order and the ability to contextualize what we see and hear. “I think it’s very important in a world where our natural human defenses are not enough anymore.”
We are living in times of heightened geopolitical instability. Conflicts are erupting across various regions of the world, economic crises are deepening and political polarization appears to be increasing both within and between countries. The impacts of climate change are exacerbating these issues, creating a volatile and unpredictable global landscape. Further complicating the geopolitical environment, nations are grappling with the rapid pace of technological advances, which bring both opportunities and significant risks.
“We live in a world that is very unstable and there are different interests that need to be taken into account,” says Sala. “We also have this technology in the hands of very few companies today. They can do pretty much whatever they want, not having standards or regulation. It’s a free lunch.”
Cooperation between tech experts and standards organizations
The standards collaboration on AI and multimedia authenticity was launched at the global AI for Good summit in Geneva earlier this year. Experts highlighted the potential societal impacts of generative AI and the urgent need to address the misuse of AI for spreading misinformation.
The discussions underscored the importance of transparency in AI, particularly regarding the provenance of data used to train AI models. Establishing the authenticity and reliability of this data is crucial for understanding potential biases and ensuring the accuracy of AI-generated content. The ability to verify the authenticity and ownership of generative AI multimedia content was identified as essential for protecting digital rights.
The participating organizations agreed on a multi-stakeholder collaboration to identify gaps in current standardization for areas such as AI watermarking, multimedia authenticity and deepfake detection technologies. They are mapping the landscape of existing standards to identify gaps and to propose solutions to address these deficiencies. The initiative is expected to play a crucial role in supporting government policy, ensuring transparency and protecting the rights of users and content creators in the rapidly evolving field of AI and generative technologies.
Where IEC, ISO and ITU are already innovating
The technologists and scientists responsible for the JPEG picture format are developing a backwards-compatible standard that they hope will make it easier to create media that protects source information and safeguards intellectual property rights. They hope it will help journalists and insurance investigators, among others, to verify the authenticity of photographs and video. JPEG is part of the ISO/IEC Committee JTC 1/SC 29, which covers the coding of audio, picture, multimedia and hypermedia information and also reports to the ITU.
JPEG Trust builds on the work of the C2PA to create a comprehensive framework for ensuring media authenticity. Unlike traditional verification methods, JPEG Trust does not define what is trustworthy. Instead, it provides the tools and guidelines to evaluate trustworthiness based on specific contextual needs.
The framework is built on three core pillars. Firstly, it ensures that media assets are linked to their provenance annotations in a tamper-evident manner, enabling users to trace the origin and history of the content. Secondly, it defines methods for extracting a wide range of trust indicators from media assets, providing tools to assess their authenticity effectively. Thirdly, it addresses privacy and security concerns by protecting sensitive information through the JPEG Privacy and Security standard (ISO/IEC 19566–4), ensuring that the pursuit of authenticity does not compromise user privacy.
How JPEG Trust works
The framework is designed to adapt to diverse scenarios, offering flexibility through “Trust Profiles” — customizable conditions for evaluating media authenticity. For example, a news organization might prioritize indicators of editorial integrity, while a government agency might focus on metadata integrity. Once evaluated, these trust indicators are compiled into a “Trust Report”, a user-friendly summary that allows end users to make informed decisions about the content they consume.
Future Developments
The new JPEG Trust standard marks the beginning of an ambitious journey. Future additions to the series will expand its capabilities, including advancements in media tokenization, watermarking and the creation of a library of common Trust Profile requirements. These developments promise to deepen the value of the international standard across industries and applications.
While JPEG Trust is not a silver bullet for the complex problem of misinformation, it represents a significant step forward. It equips individuals and institutions with the tools to navigate an increasingly deceptive digital world by providing a global, interoperable framework for media authenticity.
As JPEG Trust evolves, it holds the potential to become a cornerstone of digital trust, ensuring that truth and authenticity can finally prevail in an age when victims of fake media have often struggled to defend themselves.
Disclaimer
Mike Mullane is Vice-Chair and Communications Lead of the WSC-led AI and Multimedia authenticity standards collaboration.
