Computer Vision · 2025

Multi-Attentional Deepfake Detection

Attention-guided detection over a ConvNeXt backbone

A deepfake classifier that combines local attention, global context, and texture evidence while exposing where the network looks.

  • Python
  • PyTorch
  • timm
  • Streamlit
  • ConvNeXt
ATTN / 0685.66%evaluation accuracy

Measured evidence

evaluation accuracy
85.66%
FAKE-class F1 score
0.9228
weighted F1 score
≈0.7905

Problem worth solving

Classify face images as real or fake while exposing visual evidence that helps inspect what the network attends to.

Engineering decisions

  1. Used a ConvNeXt backbone with local attention pooling, global feature extraction, and texture enhancement.
  2. Combined the extracted signals in an ensemble classifier.
  3. Built a Streamlit demo with image upload, remote checkpoint loading, prediction output, and attention-map visualization.

What changed

  • Reached 85.66% accuracy on the reported evaluation set.
  • Recorded a 0.9228 FAKE-class F1 score and an approximately 0.7905 weighted F1 score.