Abstract
As techniques and tools for synthetic media and Deepfakes continue to advance, it is increasingly clear that video, audio and images can no longer be relied upon as truthful recordings of reality. Every digital communication channel is now vulnerable to manipulation, and there is widespread use of Deepfakes to propagate misinformation and disinformation, inflame political discord, defame opposition, commit cyber frauds or blackmail individuals. While deep learning (DL) methods have been widely used to identify Deepfakes, this paper demonstrates that classical machine learning (ML) methods can achieve superior performance--comparable with or exceeding state-of-the-art DL methods in detecting Deepfakes. Using the traditional procedures of feature development and selection, training, and testing of ML classifiers for the task actually provides better understandability and interpretability while consuming much less computing resource. In addition, an omnibus test, the Analysis of Variance (ANOVA), is conducted to compare the performance of multiple ML models. We present experiments that achieve 99.84% accuracy on the FaceForecics++ dataset, 99.38% accuracy on the DFDC dataset, 99.66% accuracy on the VDFD dataset, and 99.43% accuracy on the Celeb-DF dataset. Our study thus challenges the notion that DL approaches are the only effective way to detect Deepfakes and demonstrates that judicious use of ML approaches can be highly efficacious and cost-effective.
