KartikeyBartwal/Deepfake-Sentinel-EfficientNet-on-Duty

A deep learning model for deepfake detection using Python, Keras, and TensorFlow. The model leverages an optimized version of the EfficientNet architecture and is trained on extensive deepfake datasets

Deep-Fake-EffecNetB0-Tensorflow

EfficientNetB0 Deep Learning is a Python package designed to facilitate the training and evaluation of deep learning models using the EfficientNetB0 architecture for image classification tasks.

• Free software: MIT license
• Documentation: https://effnetb0-deep-learning.readthedocs.io

Features

• Model Training: Train models on custom datasets with the ability to fine-tune parameters.
• Pre-trained Weights: Utilize pre-trained weights for improved performance on image classification tasks.
• Model Evaluation: Assess model performance with various metrics including accuracy and loss graphs.
• Logging: Track training progress and performance metrics through detailed logs.
• Custom Dataset Handling: Easily manage datasets with support for structured training, validation, and testing splits.

Installation

To install the required packages, run:

pip install -r requirements.txt

Usage

Step 0 - Convert Video Frames to Individual Images

To extract all the video frames from the acquired deepfake datasets and save them as individual images for further processing, run:

python 00-convert_video_to_image.py

Step 1 - Extract Faces from the Deepfake Images with MTCNN

To extract faces from the deepfake images, use the pre-trained MTCNN model from the following GitHub repository:

https://github.com/ipazc/mtcnn

Run the following command:

python 01a-crop_faces_with_mtcnn.py

Step 2 - Balance and Split Datasets into Various Folders

We need to split the dataset into training, validation, and testing sets (for example, in the ratio of 80:10:10) as the final step in the data preparation phase. Run:

python 02-prepare_fake_real_dataset.py

Step 3 - Model Training

We use EfficientNet as the backbone for model training, treating the deepfake detection task as a binary classification problem applicable to both video and image content.

In this code implementation, we modified the EfficientNet B0 model by:

• Changing the initial input layer to accept 128x128 images with a depth of 3.
• Connecting the last convolutional output to a global max pooling layer.
• Adding two fully connected layers with ReLU activations.
• Utilizing a Sigmoid activation in the output layer to act as a binary classifier.

Thus, the model is expected to produce an output between 0 and 1 for a colored square image, indicating the probability of the image being either a deepfake (0) or genuine (1). Run the following command to start training:

python 03-train_cnn.py