A group of researchers from Facebook and Michigan State University have developed a method of reverse-engineering deepfakes that can not only detect a forgery but also identify the machine-learning model that produced it.
Our researchers have partnered with @michiganstateu to develop a method of detecting and attributing deepfakes that relies on reverse engineering from a single AI-generated image to the generative model used to produce it. Learn more: https://t.co/wdkCkzdncz pic.twitter.com/vkBPReR8xu
— Facebook AI (@facebookai) June 16, 2021
According to the researchers, the system identifies architectural features of unknown algorithms, bespoke to each machine-learning model. They leave a unique fingerprint on the finished image, which can be used to determine its source.
The researchers’ algorithm not only identifies traits of the generative network but also determines whether the image is a forgery, and which known model created it.
Lead researcher Tal Hassner said that identifying features of unknown models matters because software for creating deepfakes can be easily tuned. This could potentially allow miscreants to erase traces if investigators try to track their actions.
“Now we can say: “The image uploaded here and there — all of them belonged to the same model”,” said Hassner.
He added that with access to the attacker’s computer or laptop, the algorithm can pinpoint with high accuracy that the forgeries were generated on that equipment.
The researcher compared the deepfake-tracking algorithm to forensic camera-model identification techniques by looking for patterns in the produced image.
“However, not everyone can create their own camera […]. While anyone with sufficient expertise and a standard computer can develop their own model for generating deepfakes,” said Hassner.
The company believes the model will help Facebook track down those spreading fake photos and videos on social networks. However, the work remains in the research stage and is not yet ready for deployment, the developers said.
Earlier in April, researchers from the University of Washington developed an algorithm that creates realistic fake satellite maps.
In April, experts warned that the number of deepfakes online doubles every six months.
In March, Bucks County police in Pennsylvania arrested a woman for harassing teenage girls with a deepfake video. She aimed to discredit her daughter’s rivals in the cheerleading squad.
Follow ForkLog on Telegram: ForkLog AI — all the latest from the AI world!