CYBERSECURITY
Using AI to unmask AI
Researchers at EPFL are trying to gain the upper hand in a cat-and-mouse game unmasking deepfakes on the Internet.
Artificial intelligence is making it ever easier to manipulate photos and to put false words in people’s mouths in videos. Fortunately, AI-based programs themselves can also be employed to unmask such ‘deepfakes’ – by detecting incongruities in people’s skin colour, for example. Researchers at EPFL have now developed an additional module that can significantly improve the success rate of all common deepfake detectors. This module is able to transform and expand the datasets available to train the AI doing the unmasking.
But deepfake generators are also constantly evolving. “It’s a never-ending game of cat and mouse”, says the EPFL team leader, Touradj Ebrahimi. This is why he is advocating for the introduction of image and video formats that contain information on their origin and editing – such as JPEG Trust, which will be launched in early 2024.