README.md

---
license: mit
tags:
  - deepfake-detection
  - image-forensics
  - video-forensics
  - ensemble
  - pytorch
---

# DeepSafe Model Weights

Backup model weights for the [DeepSafe](https://github.com/siddharthksah/DeepSafe) deepfake detection platform. These weights are mirrored here to ensure availability in case the original sources become unavailable.

## Models Included

### Image Detection Models

| Model | File | Size | Original Source |
|-------|------|------|----------------|
| **NPR Deepfake Detection** | `npr_deepfakedetection/NPR.pth` | 5.6 MB | [chuangchuangtan/NPR-DeepfakeDetection](https://github.com/chuangchuangtan/NPR-DeepfakeDetection) |
| **UniversalFakeDetect (FC)** | `universalfakedetect/fc_weights.pth` | 4 KB | [WisconsinAIVision/UniversalFakeDetect](https://github.com/WisconsinAIVision/UniversalFakeDetect) |
| **CLIP ViT-L/14 Backbone** | `universalfakedetect/ViT-L-14.pt` | 890 MB | [OpenAI CLIP](https://github.com/openai/CLIP) |

### Video Detection Models

| Model | File | Size | Original Source |
|-------|------|------|----------------|
| **Cross-Efficient ViT** | `cross_efficient_vit/cross_efficient_vit.pth` | 388 MB | [davide-coccomini/Combining-EfficientNet-and-Vision-Transformers-for-Video-Deepfake-Detection](https://github.com/davide-coccomini/Combining-EfficientNet-and-Vision-Transformers-for-Video-Deepfake-Detection) |
| **Efficient ViT** | `cross_efficient_vit/efficient_vit.pth` | 418 MB | Same as above |

### Meta-Learner (Ensemble)

| File | Size | Description |
|------|------|-------------|
| `meta_model_artifacts/deepsafe_meta_learner.joblib` | 569 KB | Trained stacking ensemble classifier |
| `meta_model_artifacts/deepsafe_meta_scaler.joblib` | 767 B | Feature scaler |
| `meta_model_artifacts/deepsafe_meta_imputer.joblib` | 975 B | Missing value imputer |
| `meta_model_artifacts/deepsafe_meta_feature_columns.json` | 215 B | Feature column definitions |

## Credits

All model weights are the work of their respective original authors. DeepSafe mirrors them here strictly as a backup to prevent broken builds if upstream sources change. Full credit goes to:

- **NPR Deepfake Detection**: Chuangchuang Tan et al. - [Paper](https://arxiv.org/abs/2310.14036) | [GitHub](https://github.com/chuangchuangtan/NPR-DeepfakeDetection)
- **UniversalFakeDetect**: Utkarsh Ojha, Yuheng Li, Yong Jae Lee - [Paper](https://arxiv.org/abs/2302.10174) | [GitHub](https://github.com/WisconsinAIVision/UniversalFakeDetect)
- **CLIP ViT-L/14**: Alec Radford et al. (OpenAI) - [Paper](https://arxiv.org/abs/2103.00020) | [GitHub](https://github.com/openai/CLIP)
- **Cross-Efficient ViT**: Davide Coccomini et al. - [Paper](https://arxiv.org/abs/2107.02612) | [GitHub](https://github.com/davide-coccomini/Combining-EfficientNet-and-Vision-Transformers-for-Video-Deepfake-Detection)

## Usage

These weights are used by DeepSafe's Docker-based microservices. See the [DeepSafe README](https://github.com/siddharthksah/DeepSafe) for setup instructions.

```python
from huggingface_hub import hf_hub_download

# Download a specific weight file
path = hf_hub_download(
    repo_id="siddharthksah/DeepSafe-weights",
    filename="npr_deepfakedetection/NPR.pth"
)
```

## License

MIT License (for the DeepSafe platform). Individual model weights retain their original licenses from their respective authors.