Facebook develops new method to track origin of deepfakes, here's how it works

The generated fingerprints can be used as inputs for model parsing, a novel approach that uses estimated generative model fingerprints to predict a model's hyperparameters - the properties of a model that make up its architecture.

Facebook, together with Michigan State University (MSU) researchers, has developed a new reverse engineering method to detect and attribute deepfakes. The new method will facilitate the detection and tracing of deepfakes in real-world settings, where the only information deepfake detectors have at their disposal is often the deepfake itself.

While previous approaches depend on pre-existing knowledge about the deepfake's generative model itself, Facebook claims that it's 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.

"Our reverse engineering method relies on uncovering the unique patterns behind the AI model used to generate a single deepfake image. Our research pushes the boundaries of understanding in deepfake detection, introducing the concept of model parsing that is more suited to real-world deployment," Facebook wrote in a blog post.

How does the new approach work?

The new approach involves deconstructing a deepfake into its fingerprint. It begins with image attribution wherein a deepfake image runs through a fingerprint estimation network (FEN) to estimate details about the fingerprint left by the generative model. Similar to digital photography where fingerprints are used to identify the digital camera used to produce an image, image fingerprints are unique patterns left on fake images that can equally be used to identify the generative model that the image came from.

Image Credit: Facebook AI

The researchers used the properties of fingerprints as the basis for developing constraints to perform an unsupervised training. They then used different loss functions to apply these constraints to FEN to enforce the generated fingerprints to have these desired properties.

The generated fingerprints can be used as inputs for model parsing, a novel approach that uses estimated generative model fingerprints to predict a model's hyperparameters - the properties of a model that make up its architecture.

"Our reverse engineering technique is somewhat like recognizing the components of a car based on how it sounds, even if this is a new car we've never heard of before," Facebook explained.

For more details, read Facebook AI's blog.