DeepGuard

DeepGuard is a Gradio demo that predicts whether a face image is a deepfake. On the model side, it uses a fusion architecture combining EfficientNet-B4 with Xception + FFT + Cross-Attention.

This repo is now primarily focused on the demo/inference experience. Training, data preparation, and evaluation scripts are still present but are intended for advanced/research use.

Quick Start

git clone https://github.com/tuugbagul/DeepGuard.git
cd DeepGuard

python -m venv venv
venv\Scripts\activate

pip install -r requirements.txt
python demo.py

On Windows, you can also run directly:

run_demo.bat

Once the demo launches, navigate to http://127.0.0.1:7860 in your browser. Upload a face image to see the model prediction and GradCAM heatmap.

Weight Resolution

The demo searches for fusion_finetuned.pth in the following order:

1. Custom path provided via --weights
2. fusion_finetuned.pth in the repo root
3. weights/fusion_finetuned.pth
4. GitHub Release asset download

Default release URL template:

https://github.com/tuugbagul/DeepGuard/releases/latest/download/fusion_finetuned.pth

If the release asset hasn't been published yet, there are two easy options:

python demo.py --weights path/to/fusion_finetuned.pth

or

set DEEPGUARD_WEIGHTS_URL=https://your-direct-file-url/fusion_finetuned.pth
python demo.py

For PowerShell:

$env:DEEPGUARD_WEIGHTS_URL="https://your-direct-file-url/fusion_finetuned.pth"
python demo.py

Common Commands

python demo.py
python demo.py --no-download
python demo.py --weights weights/fusion_finetuned.pth
python demo.py --host 0.0.0.0 --port 7860

Model Results

Model	Overall Accuracy	Real	Fake
EfficientNet-B4 (FaceForensics++ pretrain)	59.60%	30.00%	89.20%
EfficientNet-B4 (Celeb-DF fine-tune)	96.80%	94.40%	99.20%
EfficientNet-B4 (Roop fine-tune)	89.38%	74.36%	94.21%
Fusion Model	98.12%	94.87%	99.17%

Architecture

Input (299x299 RGB)
    |
    |-- EfficientNet-B4 -> 1792-dim features
    |
    `-- Xception + FFT + Cross-Attention
            |-- Xception        -> 2048-dim
            |-- FFT Branch      ->  512-dim
            `-- Cross-Attention ->  256-dim
                                   = 2304-dim features

concat [1792 + 2304 = 4096-dim]
    -> FC(512) -> BN -> ReLU -> Dropout(0.4)
    -> FC(1) -> sigmoid -> fakeness score

Datasets

• FaceForensics++: Deepfakes, Face2Face, FaceSwap, NeuralTextures, and real YouTube videos
• Celeb-DF: Celebrity deepfake videos
• Custom Roop dataset: Fake + real pairs generated with Roop on LFW faces

Repo Structure

DeepGuard/
├── demo.py               # Main Gradio demo entry point
├── weights.py            # Weight resolution / download helpers
├── fusion_model.py       # Fusion model architecture
├── pixelguard_model.py   # PixelGuard branch
├── train/                # Training scripts
├── eval/                 # Evaluation scripts
├── data/                 # Data preparation scripts
├── visualize/            # GradCAM and reporting tools
└── roop_scripts/         # Roop generation / utility scripts

Research Scripts

These folders are not part of the demo flow:

• train/
• eval/
• data/
• visualize/
• roop_scripts/

They are kept for model development and experimentation. For initial setup, you only need to care about demo.py.

License

This project is for academic purposes.