Facebook researchers working with Michigan State University say that they can now reverse-engineer deepfakes and identify manipulated content with the use of a single still image from the video. They claim to be able to determine where deepfakes in real world settings may have originated, and the software used to produce them.
Deepfakes are digitally altered videos produced by an AI deep learning algorithm, which typically enables the mixers to paste someone’s face on someone else’s body. They have been cited as a potential threat to security as they can enable fraud and impersonation. Deepfakes have been used to mimic celebrities on Instagram and TikTok, and create manipulated pornographic videos of popular actresses.
Facebook researchers claim that their AI software can be trained to establish if a single piece of media is a deepfake based on a single frame of the video. Furthermore, the software can be used to identify the specific model used to generate the deepfake.
On Wednesday, the researchers presented a “research 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.”
“Deepfakes have become so believable in recent years that it can be difficult to tell them apart from real images. As they become more convincing, it’s important to expand our understanding of deepfakes and where they come from,” said Facebook in a blog post. “In collaboration with researchers at Michigan State University (MSU), we’ve developed a method of detecting and attributing deepfakes. It relies on reverse engineering, working back from a single AI-generated image to the generative model used to produce it.”
Facebook explained that much of the focus of the researchers has been on detecting deepfakes and their creators.
“Beyond detecting deepfakes, researchers are also able to perform what’s known as image attribution, that is, determining what particular generative model was used to produce a deepfake,” Facebook continued. “Image attribution can identify a deepfake’s generative model if it was one of a limited number of generative models seen during training. But the vast majority of deepfakes — an infinite number — will have been created by models not seen during training. During image attribution, those deepfakes are flagged as having been produced by unknown models, and nothing more is known about where they came from, or how they were produced.”
Facebook says that its reverse engineering method “takes image attribution a step further by helping to deduce information about a particular generative model just based on the deepfakes it produces,” and that it is “the first time that researchers have been able to identify properties of a model used to create a deepfake without any prior knowledge of the model.”
“Through this groundbreaking model parsing technique, researchers will now be able to obtain more information about the model used to produce particular deepfakes. Our method will be especially useful in real-world settings where the only information deepfake detectors have at their disposal is often the deepfake itself,” Facebook concluded. “In some cases, researchers may even be able to use it to tell whether certain deepfakes originate from the same model, regardless of differences in their outward appearance or where they show up online.”