MSU team develops deepfake detection model with Facebook

Artificial intelligence experts at Michigan State University and Facebook have debuted a model that can detect fake images of people online better than current methods.

The new reverse-engineering model, introduced last week, can identify "deepfakes" — media that use a synthetic, or fake, recreation of someone else's likeness portrayed as that person — from content with actual people and where they came from. Deepfakes have been increasingly harder to recognize as technology advancements have made these recreations appear more realistic.

Considered first-of-its-kind, Facebook and MSU's model goes beyond current classification strategies. It uses reverse engineering, fingerprint detection and model parsing to determine an image's validity and, in some instances, source.

The model can do this regardless of whether the generator has been previously identified or if it has been modified, both of which previously restricted the ability to track deepfake sources. And it can find this information with only the fake image or video itself, the researchers at MSU and Facebook AI said in a post on MSUToday.

“Our method will facilitate deepfake detection and tracing in real-world settings where the deepfake image itself is often the only information detectors have to work with,” said Xiaoming Liu, an MSU foundation professor of computer science who worked on the team. “It’s important to go beyond current methods of image attribution because a deepfake could be created using a generative model that the current detector has not seen during its training.”

Facebook AI and the Defense Advanced Research Projects Agency provided funding for the study, according to MSU.

How it works

Models before this one focused on determining if an image was fake and if it came from a generative model that the detection model had seen previously, according to a blog post from Facebook AI. The MSU model takes that detection another step by using reverse engineering to provide more information about where the deepfake comes from, if there is a generative model creating multiple deepfakes and how closely aligned new generators are to previously recognized ones.

This new model can determine whether a series of images comes from the same generator, providing a better understanding of the scale of coordinated disinformation campaigns, according to Facebook AI.

“Our reverse engineering method relies on uncovering the unique patterns behind the AI model used to generate a single deepfake image,” said Tal Hassner, a Facebook AI researcher, in the MSUToday announcement.

This can be particularly useful given that in many instances, there is little if any information about a deepfake beyond what is available within the media.

"With model parsing, we can estimate properties of the generative models used to create each deepfake, and even associate multiple deepfakes to the model that possibly produced them," Hassner said. "This provides information about each deepfake, even ones where no prior information existed.”

This information is what makes the model go beyond others in how it verifies images. Liu said Wednesday in an interview with The Detroit News that deepfake creators can modify generators, which would typically be recognized by deepfake checking models, and these checkers would then miss them.

But MSU and Facebook's model can recognize these modified generators, Liu said. The model can also place these new generation models within a map or web of sorts, allowing deepfake trackers to tell how close or different the new models are to others that have been previously detected.

“We have a concept of similarity as well as distance of certain models in a space,” Liu said. "That can serve as a guidance as well as evaluation schemes for future models."

Liu said this model is so specific that it can determine which pixels of an image were faked. Though average users may not notice a change in their technology, Liu said this model will be mainly utilized by large technology companies and government agencies.

Study's results show success

Facebook AI and MSU have positioned the model as a better way to detect and trace deepfakes. Results of the study back up that notion.

When it was tested with a data set of 100,000 synthetic images generated from 100 online images, the new model performed better in detecting deepfakes than a baseline, according to MSU.

Given Facebook and MSU are the first to use model parsing, there is no model to compare it against. The study created a random baseline of ground-truth vectors to compare their model against.

The fingerprint estimation component was also found to perform competitively in deepfake detection and image attribution, according to Facebook AI.

AHarring@detroitnews.com

Twitter: @alex_harring