Deepfake-Detection

This project focuses on deepfake detection leveraging the OpenForensics dataset, a comprehensive resource for face forgery detection and segmentation research. The project explores various deep learning models and evaluates their performance in distinguishing between real and fake images of human faces.

Methodology

A. Dataset

Dataset Link: Emotions Dataset

The OpenForensics dataset contains 256x256 jpg images categorized into real and fake human faces:

• Train: 140,000 images (70,000 real, 70,000 fake)
• Test: 10,905 images (5,492 fake, 5,413 real)
• Validation: 39,400 images (19,600 fake, 19,800 real) This structured dataset provides a diverse range of manipulated and authentic faces, challenging deep learning models to detect subtle differences and enhance research in deepfake detection.

B. Pre-processing

Data preprocessing involves:

• Transformations: Resizing, rotation, and normalization of images.
• Class Balancing: Random oversampling to address class imbalance.
• Dataset Split: 80-20 split for training and testing sets.
• Pre-trained Weights: Initialization for transfer learning with Vision Transformer (ViT).

C. Models

Five models are evaluated:

1. Vision Transformer (ViT): Effective for image classification with strong performance.
2. MobileNetV2: Lightweight architecture suitable for mobile applications.
3. Convolutional Neural Network (CNN): Known for image recognition tasks.
4. ResNet50: Deep residual learning for feature extraction.
5. EfficientNetB0: Efficient model scaling for resource optimization.

D. Evaluation

Models are evaluated based on:

• Accuracy: Training and validation accuracy metrics.
• Loss: Training and validation loss metrics.

Results

Model Performance

Model	Training Accuracy	Validation Accuracy	Training Loss	Validation Loss
ViT	99%	99%	0.04	0.02
MobileNetV2	99%	98%	0.02	0.05
CNN	97%	95%	0.05	0.12
ResNet50	84%	83%	0.35	0.36
EfficientNetB0	50%	50%	0.69	0.69

Model Insights

• Vision Transformer (ViT): Achieved high accuracy and low loss, demonstrating robust performance in distinguishing between real and fake images.
• MobileNetV2: Lightweight and efficient, suitable for real-time applications with strong accuracy.
• CNN: Effective but showed higher validation loss, indicating potential overfitting.
• ResNet50: Moderate accuracy with higher loss compared to ViT and MobileNetV2, suggesting room for optimization.
• EfficientNetB0: Struggled with low accuracy and high loss, indicating challenges in model fitting for this task.

Streamlit GUI

Link to Deepfake Image Classification

Sample

Link to the Application

Link to Deepfake Image Classification