Fake Video Detection

This project detects deepfake videos using a Convolutional Neural Network (CNN) trained on frames extracted from real and fake videos in the Deepfake Face Mask Dataset (DFFMD). It classifies videos as either REAL or FAKE based on multiple frames sampled per video.

---

Project Structure:

• config.py – Contains dataset paths and parameters like IMG_SIZE, FRAMES_PER_VIDEO, and LIMITED_VIDEOS.
• data_loader.py – Loads real and fake videos, extracts frames, and constructs the dataset.
• model_builder.py – Defines a deep CNN for binary classification (real vs fake).
• train.py – Loads the dataset, trains the model, evaluates it, and saves the trained model.
• predict.py – Loads the trained model and predicts whether a video is real or fake by analyzing its frames.

---

Installation:

Make sure you have Python 3.7+ and install the following dependencies:

pip install numpy opencv-python scikit-learn tensorflow json5

Dataset Location:

Data Type	Path
Fake Videos	/data/Fake/Fake/
Real Videos	/data/Real/Real/
Metadata	data/DFFD_metadata.json

Model Architecture:

The model is a deep CNN with the following structure:

•	4 blocks of:

•	Conv2D → BatchNormalization → MaxPooling2D → Dropout

•	Followed by:

•	Flatten → Dense(512) → Dense(256) → Dense(1, sigmoid)

Frame Extraction:

•	Frames per video: 10 evenly spaced frames

•	Image size: 128×128 pixels

•	Videos used: 100 real and 100 fake (configurable)

Training the Model:

To train the model:

python train.py

This will:

•	Extract frames from videos

•	Train the CNN for 5 epochs

•	Evaluate validation accuracy and loss

•	Save the trained model to saved_model/deepfake_detection_model.h5

Inference: Predict a Video

from predict import predict_video

result = predict_video("path/to/video.mp4")
print("Prediction:", result)

Evaluation:

During training, the model reports:

•	Training and validation accuracy/loss

•	Final validation performance

•	Saves the model after training

Configurable Parameters:

Parameter	Description
IMG_SIZE	Target image size (default: 128)
FRAMES_PER_VIDEO	Frames extracted per video (default: 10)
LIMITED_VIDEOS	Max videos to load from each class (default: 100)

License:

This project is intended for educational and research purposes. Feel free to use, modify, and build upon this work with credit.