Motion fakes leave clues MoDA uses to detect and trace
In plain words
This paper introduces MoDA, a computer method for spotting videos made by systems that animate a whole person from a still portrait and movement instructions.
Because these fakes can look consistent from frame to frame, MoDA checks both the scene’s visual structure and fine image patterns around moving body parts.
The authors report that MoDA was more accurate than a retrained comparison detector on average across four kinds of post-processing in their test collection.
It also detected many clips from two commercial video systems absent from training, without retraining, but those tests do not cover every new generator.
The research
When a video generator creates a whole moving person, can a detector still find signs of synthesis after older clues such as local image flaws and uneven motion disappear? That matters because the same evidence may help distinguish a fake from a real clip and identify which generator made it.
MoDA combines spatial meaning with fine image patterns to detect motion fakes and identify which generator made them.
The authors built a collection of real and generated clips, keeping portrait identities and motion sources separate between training and test. MoDA combines spatial clues about what is shown with fine image patterns that can remain around moving body edges.
On test videos processed in four ways, MoDA averaged 94.8% detection accuracy, compared with 81.6% for RECCE retrained on the same training split. This supports the design in these conditions, though the comparison does not establish shared initialization or isolate every training difference.
The detection result is an average across four post-processing settings on the MAD test data, not a guarantee for every clip or unseen generator. The authors also note a detection window before defenders collect labeled examples from a new generator.
Analysis
Deep
- Authors
- Fazhong Liu, Yan Meng, Tian Dong, Guoxing Chen, Haojin Zhu
- License
- CC BY 4.0