Project DeepDetect

In the age of deepfakes, seeing is no longer believing.
But even if you can't trust your eyes, you can trust us.

DeepDetect is an AI-powered application that helps you uncover the truth behind manipulated media.

Deployment Note:
The public Hugging Face Space uses a Gradio interface for interactive demos and explainability.
A FastAPI backend is also included in this repository for production-style API usage and integration.

---

⚠️ Important Notice

Note: Due to limited computational resources, the current model has been trained on a relatively small dataset using a short sequence_length of 10.
This means it utilizes only 10 frames from any uploaded video to make a prediction.

📊 Performance on Test Set:

• Accuracy: ~93%
• F1 Score: ~93%

---

🚀 Features

• 🔍 Deepfake detection using AI
• 🖼️ Visual explanation via heatmaps
• 🎛️ Gradio-powered user interface
• ⚡ FastAPI backend for scalable inference (not used in Hugging Face Space)
• ⚙️ End-to-end ML pipeline with MLflow & DVC integration

---

🛠️ Tech Stack

• Data Processing: NumPy, Pandas
• Modeling: PyTorch, Scikit-learn
• Image & Video: OpenCV, MediaPipe
• Visualization: Plotly
• MLOps: MLflow, DVC
• UI: Gradio
• Deployment: Hugging Face Spaces

---

🏗️ Project Structure

DeepDetect/
│
├── src/
│   └── DeepfakeDetection/
│       ├── components/       # Core components: data loading, preprocessing, training
│       ├── utils/            # Utility scripts and helpers
│       ├── config/           # Configuration files
│       ├── pipeline/         # Modular pipeline stages
│       ├── entity/           # Configuration entity classes
│       └── constants/        # Constants and global values
│
├── config/                   # Global configs
├── app.py                    # Gradio app entry point
├── main.py                   # Pipeline runner
├── requirements.txt          # Python dependencies
├── pyproject.toml            # Project metadata
├── params.yaml               # Model & training parameters
├── dvc.yaml                  # DVC pipeline configuration
├── Dockerfile                # Docker setup
├── structure.py              # Script to auto-generate project structure
└── format.sh                 # Shell script for formatting code

---

⚙️ Setup Instructions

1. Clone the repository:

   git clone https://github.com/your-username/DeepDetect.git
   cd DeepDetect

2. Make sure Python ≥ 3.9 is installed.
3. Update requirements.txt to .e if needed, then run:

   pip install -r requirements.txt

4. Install CMake
   On Ubuntu:

   sudo apt install cmake

5. (Optional) Install Docker for containerized deployment.
6. Download the FaceForensics++ dataset (see below).
   Update the source_data path in config/config.yaml to point to your local data directory.

---

📂 Dataset: FaceForensics++

This project uses the FaceForensics++ dataset.

To obtain it:

1. Visit the official repository: https://github.com/ondyari/FaceForensics

2. Fill out the dataset request form

3. After approval, you will receive:

   • Download credentials
   • An official download script

4. Use the provided script to download the data locally

5. Update the dataset path in:

   config/config.yaml
   

⚠️ The dataset is not included in this repository and cannot be redistributed.

---

🧪 Usage

Quick Start (Full Pipeline)

dvc init --force
dvc repro

Or manually:

python main.py

Stage-wise Execution

# Step 1: Data Ingestion
python src/DeepfakeDetection/pipeline/stage_01_data_ingestion.py

# Step 2: Preprocessing
python src/DeepfakeDetection/pipeline/stage_02_data_preprocessing.py

# Step 3: Model Training
python src/DeepfakeDetection/pipeline/stage_03_model_training.py

# Step 4: Evaluation
python src/DeepfakeDetection/pipeline/stage_04_model_evaluation.py

Launch App

🎛️ Gradio Demo (Hugging Face / Local UI)

python app.py

⚡ FastAPI Backend (Local / Production)

uvicorn api:app --host 0.0.0.0 --port 8000

🔐 Note:
By default, the prediction pipeline loads the model from MLflow (logged during training). Ensure MLflow initialization is enabled in:

• stage_03_model_training.py
• stage_04_model_evaluation.py

Set up your .env file with MLflow credentials at the project root, and make sure you’ve executed dvc repro.

MLFLOW_TRACKING_URI=your_mlflow_tracking_uri
MLFLOW_TRACKING_PASSWORD=your_mlflow_pass

Update prediction.py to load your generated MLflow run, for example:

mlflow.pytorch.load_model("runs:/<your_run_id>/model")

Alternatively, if you don’t want to use MLflow, modify the loading logic in prediction.py to use the trained model (after running the complete pipeline) located at:

/artifacts/model_training

---

🤝 Contributing

We’d love your contributions!

1. Fork the repo
2. Create a new branch:

   git checkout -b feature/amazing-feature

3. Make your changes & commit:

   git commit -am "Add amazing feature"

4. Push your branch:

   git push origin feature/amazing-feature

5. Open a Pull Request

---

📄 License

This project is licensed under the GPL-3.0 License.

---

Made with ❤️ by Sanskar Modi