Deepfake Detection on UADFV (Feature Fusion)

Deepfake classifier combining CNN and handcrafted HAAR/Gabor features.

I developed a deepfake detection pipeline that fuses CNN face features with handcrafted HAAR and Gabor descriptors. For this website, the reported result is from Task 1 only: evaluating UADFV with a model trained on UADFV.

Key outcomes:

  • Two-branch architecture: resized face input (60x60) plus handcrafted feature vector (30 features).
  • Improved Task 1 AUC from handcrafted-only baseline (0.84) to fusion model (0.93).
  • Built a full preprocessing and evaluation pipeline for frame-level prediction.

Tech stack:

  • Python
  • TensorFlow/Keras
  • OpenCV
  • Dlib
  • HAAR and Gabor filters

Links: