Detecting Deepfake Images

Overview

This project focuses on detecting deepfake images using deep learning. The model is trained using ResNet18 with transfer learning to classify images as either Real or Fake.

The goal of this project is to demonstrate an end-to-end machine learning pipeline, starting from dataset preprocessing and model training to deployment through a Flask API and a simple web interface where users can upload an image and get predictions.

The system performs the following tasks:

• Preprocess image datasets
• Train a deep learning model using PyTorch
• Evaluate model performance
• Deploy the trained model using Flask
• Provide a simple frontend interface for testing predictions

---

Tech Stack

• Python
• PyTorch
• Torchvision
• Flask
• HTML + TailwindCSS
• NumPy
• Pillow

Installation

Clone the repository

git clone https://github.com/anujpratap12/Detecting-Deepfake

Navigate to the project folder

cd Detecting-Deepfake

Install dependencies

pip install -r requirements.txt

Model Architecture

The model uses ResNet18 pretrained on ImageNet. Transfer learning allows the network to reuse powerful visual feature representations learned from large datasets.

Workflow

1. Input image is resized to 224 × 224
2. Image is converted into a tensor
3. ResNet18 extracts image features
4. The final fully connected layer is modified for binary classification
5. The model predicts whether the image is Real or Fake

Training Configuration

Loss Function CrossEntropyLoss

Optimizer Adam Optimizer

Learning Rate 0.0001

Number of Epochs 10

---

Dataset Structure

The dataset is organized into training and validation folders.

dataset/
│
├── train/
│   ├── fake/
│   └── real/
│
└── valid/
    ├── fake/
    └── real/

Each folder contains images belonging to the respective class.

---

Training the Model

To train the model run:

python train_model.py

Example training output:

Epoch 1/10 | Train Acc: 76.2% | Val Acc: 72.4%
Epoch 5/10 | Train Acc: 88.3% | Val Acc: 84.7%
Epoch 10/10 | Train Acc: 92.5% | Val Acc: 90.1%

After training, the model weights are saved as:

deepfake_detector.pth

---

Model Evaluation

The model performance was evaluated using the validation dataset.

Evaluation metrics used:

• Accuracy
• Precision
• Recall
• F1-Score

Example results:

• Accuracy: 90.1%
• Precision: 89.4%
• Recall: 88.7%
• F1 Score: 89.0%

---

Running the Application

Start the Flask server:

python app.py

Once the server is running, open the frontend page and upload an image. The model will process the image and return whether it is Real or Fake.

---

API Endpoint

POST /predict

Input Image file

Example Response

{
  "prediction": "fake"
}

---

Application Demo

Real Image Detection

Fake Image Detection

---

Project Structure

Detecting-Deepfake
│
├── train_model.py        # model training script
├── app.py                # Flask API server
├── index.html            # frontend interface
├── deepfake_detector.pth # trained model weights
├── requirements.txt      # project dependencies
└── README.md

---

Future Improvements

Some possible improvements for this project:

• Train using larger deepfake datasets
• Improve accuracy using more advanced architectures like EfficientNet
• Extend the system to detect video deepfakes
• Deploy the model using Docker or cloud services
• Build a real-time detection system

---

Author

Anuj Pratap Singh Final Year B.Tech Student Interested in Artificial Intelligence and Machine Learning