Fake Face Detection with DenseNet121

Overview

This project aims to detect fake faces using a fine-tuned DenseNet121 model. The model is trained for binary classification (real vs. fake faces) using deep learning techniques.

Requirements

• tensorflow
• keras
• numpy
• pandas
• matplotlib
• opencv-python
• scikit-learn
• streamlit

Features

• Utilizes DenseNet121 for feature extraction.
• Fine-tuned model with additional layers for improved performance.
• Achieves 98.7% accuracy on the dataset.
• Implements data augmentation, dropout, and batch normalization for better generalization.
• Streamlit-based frontend for user-friendly interaction.
• Users can upload an image or provide an image URL to check if the face is real or fake.

Installation

1. Clone this repository:

   git clone https://github.com/resonancejb/Fake-image-detection-model.git

2. Install dependencies:

   pip install -r requirements.txt

Usage

1. Run the Jupyter Notebook:

   jupyter notebook notebooks/fake-face-detection-with-keras-accuracy-0-987.ipynb

2. Train the model or use the pre-trained version in models/.
3. Evaluate performance on test images.
4. Run the Streamlit App:

   streamlit run scripts/main.py

5. Upload an image or enter an image URL in the web interface to check if it is real or fake.

Dataset

The dataset consists of real and fake face images. It is recommended to preprocess and balance the dataset before training.

Sample Dataset Images

Model Details

• Backbone: DenseNet121 (pre-trained on ImageNet)
• Additional Layers:
  • Global Average Pooling
  • Dropout (Regularization)
  • Dense Layers with ReLU Activation
• Loss Function: Binary Crossentropy
• Optimizer: Adam

Results

The model achieves 98.7% accuracy on the test dataset.

Example Results

Here are some example results of real and fake face detection:

Fake Face Example:

Real Face Example:

Model Performance Metrics:

1. Accuracy Curve
   

2. Loss Curve
   

3. Confusion Matrix
   

Future Improvements

• Use ensemble learning for better generalization.
• Improve dataset quality and diversity.
• Implement adversarial training to detect more sophisticated fake images.
• Enhance the Streamlit UI for a better user experience.

License

This project is open-source under the MIT License.