DeepFake Image Detection

Mentor: Prof. Koteswar Rao Jerripothula, Dept. of Electrical Engineering
Duration: Aug 2024 – Nov 2024

Overview

Detects deepfake images using deep learning models. Evaluated ResNet50, DenseNet, EfficientNet, and Vision Transformer (ViT) architectures with robust preprocessing and real-time deployment via Flask API.

Features

• Compared ResNet50, DenseNet, EfficientNet, and ViT for deepfake detection.
• Used MTCNN for face detection and data augmentation for robustness.
• Fine-tuned models with ImageNet weights and optimized hyperparameters.
• Evaluated using Accuracy, F1-score, and ROC-AUC; DenseNet achieved 95% accuracy.
• Deployed best model using Flask API for real-time detection.

Getting Started

Prerequisites

• Python 3.8+
• PyTorch or TensorFlow
• Flask
• OpenCV, MTCNN, scikit-learn

Installation

git clone 
cd deepfake-image-detection
pip install -r requirements.txt

Data Preparation

Organize your dataset:

data/
  real/
  fake/

MTCNN face detection is applied during preprocessing.

Training

python train.py --model densenet

Supported: resnet50, densenet, efficientnet, vit

Evaluation

python evaluate.py --model densenet

Inference & API

python app.py

Send a POST request with an image to /predict for real-time detection.

Results

• DenseNet achieved 95% accuracy.
• Robustness improved with augmentation and MTCNN.
• Real-time detection via Flask API. ``

License

For academic and research purposes only.