Deepfake Detection

A multi-view anomaly-detection ensemble. Five complementary detectors covering spatial, temporal-motion, and global-consistency anomalies, combined via soft-vote.

5 modelsFF++ AUC 1.000Celeb-DF AUC 0.9056Gap reduction 22%CS 668 Capstone · Pace
App repoResearch repoModel weights

Three anomaly classes — one ensemble

Spatial

Per-frame texture and edge inconsistencies. ResNet-18 (baseline) and EfficientNet-B4 (high capacity).

Temporal-motion

Incoherent motion and frame-to-frame transitions. R3D-18 over the whole 16-frame clip and R3D-18 + RAFT optical-flow interpolation.

Global consistency

Long-range relationships across the face via self-attention. ViT-B/16 patch-level encoder.

Each detector is trained on FaceForensics++ C23 and held out on Celeb-DF v2 (zero-shot). The ensemble shrinks the FF++ → Celeb-DF generalization gap from 0.1216 (best single model) to 0.0944.

Try a pre-loaded example

Preview the video, then run the ensemble
real

FF++ C23

FaceForensics++ original

Untouched original from the FF++ in-distribution test split.

fake

FF++ C23

FaceForensics++ Deepfake

FaceSwap/Deepfakes manipulation. The model usually nails this with > 0.99 confidence.

real

Celeb-DF v2

Celeb-DF v2 real

Authentic celebrity footage held out of training entirely.

fake

Celeb-DF v2

Celeb-DF v2 synthesis

Cross-dataset deepfake — harder than FF++ since the manipulation pipeline differs from training.

Or upload your own video

Drop an MP4 here

or click to browse · max 50 MB · .mp4 / .mov / .m4v