🛡️ DeepSafe-Satyanveshi: Deepfake Video Detection

DeepSafe-Satyanveshi is an end-to-end deepfake detection system designed to identify manipulated videos with high precision. It utilizes a hybrid deep learning architecture, combining EfficientNet-B0 for spatial feature extraction and a Bidirectional GRU for temporal analysis.

🚀 Quick Start

1. Prerequisites

Ensure you have Python 3.10+ installed. It is highly recommended to use a high-performance GPU (NVIDIA GTX 1650 or better) for training, though inference can run on CPU.

2. Installation

Clone the repository and set up a virtual environment:

PowerShell

Clone the repo

git clone https://github.com/yaswanth-pidugu/DeepSafe-Satyanveshi.git

cd DeepSafe-Satyanveshi

Create and activate virtual environment

python -m venv .venv

.venv\Scripts\activate

Install dependencies

pip install -r requirements.txt

3. Configuration

This project uses a centralized configuration system.

Navigate to the configs/ folder.

Ensure base.yaml contains the model hyperparameters.

Open local.yaml and update the paths to match your local directory structure.

4. Running the Pipeline

Follow these steps in order:

Step A: Preprocessing

• Extract faces from your raw video datasets.

• PowerShell

python -m preprocessing.face_extractor

Step B: Data Splitting

• Split your processed face data into training, validation, and testing sets.

• PowerShell

python -m preprocessing.split_data

Step C: Training

• Train the hybrid EfficientNet-GRU model. The best weights will be saved automatically based on validation loss.

• PowerShell

python -m training.train

Step D: Evaluation & Inference

• Generate a classification report and confusion matrix, or test a specific video.

• PowerShell

• #Run evaluation on the test set

python -m training.test

• #Run inference on a specific video

python -m inference.predict_video

🏗️ Project Structure

🏗️ Project Structure

DeepSafe-Satyanveshi/
├── config/                # Centralized YAML configurations
│   ├── base.yaml          # Global hyperparameters
│   ├── local.yaml         # Local machine paths (Git ignored)
│   └── prod.yaml          # Production/Server settings
├── inference/             # Scripts for real-world prediction
│   └── predict_video.py   # Main inference logic
├── preprocessing/         # Data cleaning and face extraction scripts
│   ├── face_extractor.py  # MTCNN face cropping
│   └── split_data.py      # Training/Val/Test partitioning
├── training/              # Model architecture and training logic
│   ├── dataset.py         # Custom PyTorch Dataset class
│   ├── model.py           # EfficientNet + Bi-GRU stack
│   ├── train.py           # Training loop with checkpoints
│   └── test.py            # Evaluation and confusion matrix
├── .gitignore             # Rules to exclude data/local configs
├── config_manager.py      # Logic for loading YAML environments
└── requirements.txt       # Project dependencies

🛠️ Tech Stack

1. Deep Learning: PyTorch (EfficientNet_B0 + Bi-GRU)

2. Computer Vision: OpenCV, facenet-pytorch (MTCNN)

3. Configuration: YAML

4. Analytics: Scikit-learn, Matplotlib, Seaborn