Combatting Deepfakes, Misinformation and Fraud in the AI Era

As with any emerging technology, there are benefits and drawbacks. With the emergence of AI, we have seen wonderful developments in improved workflows, smart factory assistants, accelerated R&D and more. But we have also seen that very same technology used for nefarious purposes that threaten the spirit of openness, access and trust when it comes to the World Wide Web, and beyond. 

AI-generated synthetic media, deepfakes, misinformation and fraud 

While not entirely new phenomena, the rapid advancement of AI and digitization in recent decades has significantly accelerated their impact, particularly through deepfakes and AI-generated synthetic content. They can do a significant amount of damage anywhere from people’s personal lives to presidential elections to international diplomacy. The main threat even more than the technology itself is the lack of awareness about how convincing AI-generated content can be. Implementing standards-compliant technologies is a potent solution to this problem. Standards enable creators, content providers and users with tools to verify the authenticity of media.   

The JPEG Trust standard 

The JPEG Trust standard provides a comprehensive and flexible framework for establishing trust in media, which can also be applied to video and audio. Rather than deciding what is trustworthy, the standard describes tools and methodologies that individuals and organizations can use to define their own Trust Profiles. It does this by linking images with their metadata and other information such as provenance, forgery detection, and fact checking which highlight any attempt to tamper with them. The presence of this information provides contextual information for the establishment of trust in an image. For users, it can engender tailored trust and confidence. For organizations, it can address challenges such as sovereignty and vendor lock in. 

The JPEG Trust standard was developed by IEC and ISO in response to concerns about fake media and has been enhanced regularly to address the growing concerns of AI-generated content with the goal of helping users verify information authenticity across different platforms and contexts. 

When regulation gets it wrong 

The EU AI Act’s Article 50 requires that AI-generated content be labeled. This however is insufficient as fraudsters can avoid labeling their fraudulent content and actual authentic content can be incorrectly labeled as fake. Effective regulation should require labeling of all contents, both AI-generated and authentic, to prevent misclassification and deter fraudulent activity. More comprehensive regulation is necessary to properly address the challenges of AI content verification. Voluntary standards go a step beyond simply labeling something as AI-generated.  Standards-based solutions are effective as they offer better interoperability, lower switching costs, and greater scrutiny than proprietary alternatives, for example so they are in effect good for everyone. They’re voluntary, promote interoperability and facilitate global exchange. With the growing desire to put guardrails around emerging technologies like AI and increasing sovereignty concerns, standards-based approaches, whether C2PA-centric, company-specific ecosystems like Google and Apple, or International standards like JPEG Trust will gain popularity in the coming years.  

A few guidelines 

In the meantime, here are a few things we can do to protect ourselves against misinformation… 

• The reactive approach: Each AI model that modifies or generates synthetic content has a hidden signature, similar to the style of a person when they write a text by hand. Reactive approaches are based on algorithms that try to identify that signature/style. Most such detectors are based on machine learning, or AI. They learn the style by examining a large amount of synthetic content and deepfakes and compare them to content that is recorded by a camera or a microphone.  

• The proactive approach: Tracking using the provenance techniques which involves signing the content when created and throughout its journey. For example, a microphone can sign the voice or audio it records, or a camera can sign the content that it records. One can even extend this idea by signing again over the previous signature(s) when the content is modified by indicating which modification took place in the journey of the content from creation to consumption. The aim is to let a user know how the content they access was created (by a camera, by a microphone, or by a generative AI solution, and what happened to it until the moment the content was accessed by a user. The signature could be cryptographic and securely linked to the content using a hash function, or it can be embedded as a watermark within the content. The information about the creation and modification of the content could be even put on a ledger e.g. on a blockchain, so that it cannot be modified.  

• The collaborative approach: One can search online if the content, or something similar to it (for example, a smaller or a cropped version of a picture) has been reported as fake. This approach is similar to good old fact-checking that professional journalists have been trained to use to make sure the information is confirmed before publishing it. Recently, modern fact-checking platforms have become more available and we can expect this to continue exponentially. 

None of the above approaches is universally the best alone. This is why an international standard like JPEG Trust has defined a procedure that combines them all.