At the end of 2022 Zelus and CERTH got the opportunity to address a major challenge of our times: The detection of fake hyper-realistic synthetic media. Deepfakes become increasingly more prevalent and harder to detect with advancements in AI like Dall-E, MidJourney, and the Stable Diffusion models. A well-know example is the satirical images of Trump, created by the citizen journalist and founder of Bellingcat, Eliot Higgins, who uploaded them to his personal Twitter account. Higgins used the AI model MidJourney to generate the forged images, which generated controversy and led the company to ban him from its web service.
While deepfakes become more prevalent, more data become available which can lead to the creation of stronger algorithms for detection. These algorithms are not bulletproof though, we still need the human intelligence to evaluate possible deepfake media, whose expertise and integrity, we can further entrust through the public access to the evaluations of deepfakes (human and automatic) by storing them in a blockchain network.
In this project, our partner CERTH was responsible for the scientifc method we followed for the AI-enabled detection of deepfakes, utilizing models specializing in:
- Extensibility: being able to tackle different manipulations
- Generalization: increasing the chances of detecting similar manipulations
- Explainability: matching deepfake methods detected with known manipulation techniques
- Reliability: scoring deepfakes as an uncertainty metric
Zelus utilized its Blockchain expertise to store the AI-enabled evaluations and make them available for cross-checking by humans creating digital ID for each media investigation profile which enhances and guarantees its credibility to the average user who is interested in facts or professionals with limited time in their disposal who are interested in credibility of evidence.
The video below demonstrates the outcome of this collaboration.