|
|
--- |
|
|
license: apache-2.0 |
|
|
tags: |
|
|
- deepfake-detection |
|
|
- ensemble |
|
|
- deep-svdd |
|
|
- autoencoder |
|
|
- computer-vision |
|
|
datasets: |
|
|
- cifar10 |
|
|
- mnist |
|
|
metrics: |
|
|
- accuracy |
|
|
- precision |
|
|
- recall |
|
|
- f1 |
|
|
library_name: pytorch |
|
|
--- |
|
|
|
|
|
# Ensemble Deepfake Detector |
|
|
|
|
|
**State-of-the-art ensemble model** combining Deep SVDD + Autoencoder for deepfake detection with **77.5% accuracy** and **88.75% recall**. |
|
|
|
|
|
## Model Description |
|
|
|
|
|
This ensemble combines two complementary anomaly detection approaches: |
|
|
|
|
|
1. **Deep SVDD** - Detects anomalies via hypersphere distance in latent space |
|
|
2. **Autoencoder** - Detects anomalies via reconstruction error |
|
|
|
|
|
Each model has 50% voting weight, and scores are averaged with an optimized threshold. |
|
|
|
|
|
## Performance |
|
|
|
|
|
Evaluated on 800 test images (CIFAR-10 vs MNIST, CIFAR-10 vs Fashion-MNIST): |
|
|
|
|
|
| Metric | Value | |
|
|
|--------|-------| |
|
|
| **Accuracy** | **77.50%** | |
|
|
| **Precision** | **72.22%** | |
|
|
| **Recall** | **88.75%** 🎯 | |
|
|
| **F1 Score** | **79.46%** | |
|
|
| **Model Agreement** | 53.00% | |
|
|
|
|
|
**Key Strength:** 88.75% recall means it catches **nearly 9 out of 10 fakes**! |
|
|
|
|
|
## Quick Start |
|
|
|
|
|
```python |
|
|
from ensemble_model import EnsembleDeepfakeDetector |
|
|
|
|
|
# Load ensemble (downloads both models automatically) |
|
|
detector = EnsembleDeepfakeDetector.from_pretrained() |
|
|
|
|
|
# Predict on image |
|
|
score, is_fake = detector.predict('image.jpg') |
|
|
print(f"Deepfake Score: {score:.4f}") |
|
|
print(f"Is Fake: {is_fake}") |
|
|
``` |
|
|
|
|
|
## Installation |
|
|
|
|
|
```bash |
|
|
pip install torch torchvision huggingface-hub pillow |
|
|
``` |
|
|
|
|
|
## Threshold Options |
|
|
|
|
|
The ensemble uses an **optimized threshold of 0.1163** by default: |
|
|
|
|
|
```python |
|
|
# Use different thresholds |
|
|
detector.set_threshold(0.1163) # Optimal (default) - 88.75% recall |
|
|
detector.set_threshold(0.5) # Conservative - fewer false positives |
|
|
detector.set_threshold(0.05) # Sensitive - catch even more fakes |
|
|
``` |
|
|
|
|
|
**Threshold Comparison:** |
|
|
|
|
|
| Threshold | Accuracy | Precision | Recall | Use Case | |
|
|
|-----------|----------|-----------|--------|----------| |
|
|
| **0.1163 (optimal)** | 77.5% | 72.2% | **88.8%** | **Recommended** - Best balance | |
|
|
| 0.5 (conservative) | 66.9% | 73.1% | 47.5% | Minimize false alarms | |
|
|
| 0.05 (sensitive) | ~70% | ~65% | ~95% | Maximum detection | |
|
|
|
|
|
## Component Models |
|
|
|
|
|
This ensemble uses: |
|
|
- [ash12321/deep-svdd-anomaly-detection](https://huggingface.co/ash12321/deep-svdd-anomaly-detection) |
|
|
- [ash12321/deepfake-autoencoder-cifar10-v2](https://huggingface.co/ash12321/deepfake-autoencoder-cifar10-v2) |
|
|
|
|
|
## How It Works |
|
|
|
|
|
1. **Deep SVDD**: Learns a hypersphere around normal images. Fakes fall outside this sphere. |
|
|
2. **Autoencoder**: Learns to reconstruct normal images. Fakes have high reconstruction error. |
|
|
3. **Ensemble**: Averages both scores (50/50 voting) for robust detection. |
|
|
|
|
|
**Low Model Agreement (53%)** indicates the models detect different anomaly types - this is a strength! |
|
|
|
|
|
## Training Data |
|
|
|
|
|
- CIFAR-10 (natural images) |
|
|
- CIFAR-100 (natural images) |
|
|
- STL-10 (natural images) |
|
|
|
|
|
## Limitations |
|
|
|
|
|
- Trained on natural images - best for detecting distribution shift |
|
|
- May not generalize to all deepfake types |
|
|
- Requires RGB images resized to 128x128 |
|
|
|
|
|
## Citation |
|
|
|
|
|
```bibtex |
|
|
@misc{ensemble-deepfake-detector, |
|
|
title={Ensemble Deepfake Detector}, |
|
|
author={ash12321}, |
|
|
year={2024}, |
|
|
publisher={Hugging Face}, |
|
|
url={https://huggingface.co/ash12321/deepfake-ensemble-detector} |
|
|
} |
|
|
``` |
|
|
|
|
|
## License |
|
|
|
|
|
Apache 2.0 |
|
|
|
