Stop My Dancing! Understanding, Detecting and Attributing Motion-Aware Deepfake Videos
MoDA detects and attributes motion-aware deepfake videos, reaching 94.8% in-distribution accuracy on a new benchmark.
Pose-guided diffusion models can synthesize full-body motion, creating motion-aware deepfakes that the authors say have already reached hundreds of millions of viewers. They release a benchmark of more than 1.5 million frames from 1,363 real and 30,122 synthetic videos across six generators. Their MoDA detector reports 94.8% in-distribution and 89.1% cross-dataset accuracy, 91.5% model attribution, 81.94% on 200 clips from two unseen commercial platforms, and 78.13% on 1,200 open-internet clips. Under white-box, gray-box, and black-box adaptive attacks, MoDA's accuracy drops less sharply than the baselines.
- Benchmark has 1,363 real and 30,122 synthetic videos, over 1.5 million frames
- MoDA reaches 94.8% in-distribution and 89.1% cross-dataset detection
- Attribution accuracy is 91.5%, and 81.94% on two unseen commercial platforms
- 78.13% detection on 1,200 unseen internet clips totaling 55k frames
- Detection stays more stable than baselines under adaptive attacks
From the original article
1 min read at arxiv.orgPose-guided diffusion models can now synthesize entire human figures in motion, spawning a new class of deepfakes: Motion Aware Deepfake (MAD) that have already reached hundreds of millions of viewers. To better understand this emerging threat, we construct the first MAD-specific benchmark and measurement framework, containing over 1.
By Fazhong Liu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.