Research into the detection of AI-generated fake videos (aka DeepFakes) is heating up. In this excellent Two Minute Papers summary, Károly Zsolnai-Fehér walks us through the results of multiple studies and algorithms that can detect face swapping in videos.
In the Two Minute Papers video, we look at the emerging FaceForensics++ method detailed in this Aug 2019 paper and video. FaceForensics++ combines a novel approach, layering a model data set, face tracker, AI face swap method (classifier) detection, and per-pixel segmentation. Together they present a leap forwards in detecting digital manipulation.
Of course, the accuracy of the detection depends on the quality of the input, so low quality, highly compressed video clips make it significantly more difficult to detect manipulation than in high quality videos.
As the research matures, we will see larger sample sizes for testing, higher collaboration and competitive shootouts. Perhaps this will lead to a Factcheck.org for DeepFakes as fake news proliferates. In the meantime, I’ll continue to be highly skeptical of low quality videos as a first line of defense 😉