Reproduce “Detecting Deepfake Talking Heads from Facial Biometric Anomalies” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Justin D. Norman, Hany Farid
This research introduces an identity-agnostic technique for deepfake detection by analyzing statistical anomalies in facial biometric consistency over video frames. The method achieved 94.9% overall accuracy on diverse datasets, demonstrating robustness to resolution changes and comparable performance to state-of-the-art approaches.
Key insights
- Deepfake talking heads often exhibit unnatural statistical patterns in their facial biometric consistency over time, such as bimodal distributions or anomalous kurtosis in pairwise similarity scores.
- Specific statistical features like variance-to-mean ratio, kurtosis-to-variance ratio, and skewness of facial biometric similarities are highly effective at distinguishing authentic videos from deepfakes.
- While deepfake generation struggles to maintain natural biometric consistency across large head motions, this inconsistency can be detected even after common video degradation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.