Google researchers have released a database of more than 3,000 so-called deepfake videos. That dataset can be used by other researchers and developers to build tools that can better identify such fake videos and images.
The dataset has been added to the FaceForensics Benchmark. That is a system that contains multiple technologies for creating and detecting deepfake videos and other material. The Benchmark was built by the Technical University of Munich and the Federico II University of Naples, and is sponsored by Google.
Deepfake technologies have been around for a few years now. The technology behind it keeps getting better. Deepfakes are photos or videos created automatically based on deep learning, and often require no more than a few source images to be barely distinguishable from the real thing. Google says that deep learning can have many useful applications, but there are also many risks and challenges associated with it. Deepfakes are such a challenge because, according to the company, they are increasingly being used to harm individuals or even entire communities. For example, there was recently a fuss about a website that imitated the controversial Canadian psychologist Jordan Peterson.
Google has now developed its own dataset of videos. The company paid actors to record videos, and built thousands of deepfake videos based on that by running existing deepfake algorithms on them. This dataset can be used by anyone. Researchers can study the videos to better understand how deepfakes work, and how to create tools that can recognize those videos. Earlier this year, Google already made another dataset available of deepfake audio files.