CNN Deepfake Detector

Description

It uses a simple Convolutional Neural Network structure built on Tensorflow.

Dataset

The dataset used for training comes from [Kaggle's Deepfake Detection Challenge](https://www.kaggle.com/robikscube/kaggle-deepfake-detection-introduction). It features fake images and real images. Also includes the model faces images handpicked by the team.

/real_and_fake_face

The /real_and_fake_face folder contains 2 subfolders:

• training_real : real facial images. Labeled by '0' in code
• training_fake : fake facial images. Labeled by '1' in code

Model

Layer Number	Layer
1	Input Layer
2	Conv2D. layer, kernel = 55, stride = 2, #filters=32
3	MaxPool2D, pool size = 22
4	Conv2D. layer, kernel = 3, stride = 2, #filters=64
5	Flatten Layer
6	Sigmoid Layer

• Optimizer: Adam
• Loss Function: Binary Crossentropy
• Metrics: Accuracy, Precision and Recall

Files and Folders

/real_and_fake_face

Includes real and fake image data

/model_train.py

Create, train and save the model. Primary function is to perform 50 epochs for each image size to find the optimal epoch to maximize validation accuracy. 10% of data used for validation

/model_train_and_evaluate.py

Create, train, evaluate and save the model. Primary function is to create models of image sizes with their respective optimal epoch number. 20% of data is used for evaluation.

/model_evaluate

Load any existing model and re-evaluate on the evaluation dataset.

/savedModels

Location to save models. Currently has the final models used for discussion

/savedModels/For EPOCH Selection

Contains models of sizes 32, 64, 128, 256px (length of square) with EPOCHS of 50. Used to analyze the change in validation accuracy over EPOCHS.

Currently, the following EPOCHS maximizes the respective validation accuracy for the corresponding inmage size:

• 32px: 12 EPOCHS
• 64px: 8 EPOCHS
• 128px: 4 EPOCHS
• 256px: 8 EPOCHS

/savedModels/For IMG_SIZE Selection

Contains models of sizes 32, 64, 128, 256px with their respectively selected EPOCHS

/log

Contains summaries and log files of the saved models

Currently Saved Model's Performance

Evaluated on the evaluation dataset. Performed in /model_train_and_evaluate.py

Image Size of Detector (px)	Accuracy	Precision	Recall
32	0.655	0.631	0.641
64	0.655	0.632	0.635
128	0.682	0.648	0.708
256	0.597	0.552	0.745