Deepfake Detection using Xception-based CNN

Project Overview

This project implements a deepfake detection system using an Xception-based Convolutional Neural Network (CNN). The model is designed to classify videos as either real or fake by analyzing individual frames and identifying visual artifacts and inconsistencies typical of deepfake manipulations.

Model Architecture

The detection system is built on the Xception architecture, which is well-suited for image classification tasks due to its efficient use of depthwise separable convolutions. The model consists of:

• Pre-trained Xception base (ImageNet weights)
• Global Average Pooling layer
• Dense layer (256 neurons with ReLU activation)
• Dropout layer (0.5 rate) for regularization
• Output layer with sigmoid activation for binary classification

The implementation uses a two-phase training approach:

1. Initial training with a frozen base model
2. Fine-tuning where deeper layers of the base model are unfrozen for better feature extraction

Dataset

The model is trained on the "1000_videos" dataset, which contains:

• Training set: Located in train/ directory
• Validation set: Located in val/ directory
• Test set: Located in test/ directory

Each directory contains categorized videos/frames under real and fake classes.

Requirements

• Python 3.x
• TensorFlow 2.x
• Keras
• scikit-learn
• numpy
• OpenCV (for video processing)

Installation

pip install tensorflow scikit-learn numpy opencv-python

Usage

Training the Model

python train_model.py --data_dir path/to/dataset --epochs 20

Detecting Deepfakes in Videos

python detect_video.py --input path/to/video --output results.txt

Model Evaluation

The model achieves competitive performance on the test set with metrics including:

• Binary accuracy
• Precision and recall for both real and fake classes
• F1-score

Model Weights

Pre-trained model weights are available in two formats:

• Keras format (.keras): xception_model_best.keras
• HDF5 format (.h5): xception_model_best.h5

Implementation Details

The implementation includes:

• Data augmentation (horizontal flips, rotation, zoom) to improve generalization
• Early stopping to prevent overfitting
• Model checkpointing to save the best model during training
• Fine-tuning strategy that freezes early layers while training deeper layers
• Normalization of input images to the range [-1, 1]