Prithanjan/Deepfake_Resnet_LSTM

Deepfake Detector (ResNet18 + LSTM)

This project uses:

• —ResNet18 pretrained on ImageNet
• —LSTM temporal modeling
• —Celeb-DF v2 dataset
• —Model: ResNet18 + LSTM
• —Test AUC: 0.763
• —Threshold: 0.41

Usage

from PIL import Image
import torch
# Load model, preprocess video frames, predict

Performance:

| Metric | Value | Best threshold: 0.041 Accuracy : 0.7192 Precision: 0.6909 Recall : 0.6940 F1 : 0.6924

The model was trained on 100 real and 100 fake videos from Celeb-DF v2.

license: mit ---