---
license: mit
tags:
- deepfake-detection
- image-classification
- real-and-fake-faces
---

# Deepfake Image Detector (Real and Fake Faces Fine-Tuned)

## Model Performance
- Test Accuracy: 93.02%
- Best Validation Accuracy: 94.12%
- Best Epoch: 1
- Planned Epochs: 12
- Actual Epochs Trained: 5 (early stopping applied)

## Dataset
- Training: 1,367 images
- Validation: 170 images
- Test: 172 images

## Training Details
Training was stopped early at epoch 5 due to early stopping criteria being met.
The best model was achieved at epoch 1 with validation accuracy of 94.12%.

## Usage
```python
from transformers import ViTForImageClassification, ViTFeatureExtractor

model = ViTForImageClassification.from_pretrained('shivani1511/deepfake-image-detector-new-latest-v2')
feature_extractor = ViTFeatureExtractor.from_pretrained('shivani1511/deepfake-image-detector-new-latest-v2')
```

## Notes
- Fine-tuned on Real and Fake Faces dataset to address AI-generated fake detection.
- Base model: shivani1511/deepfake-image-detector-new-latest (Vision Transformer).
- Improvements: Enhanced data augmentation, class-weighted loss, Mixup, more unfrozen layers.