# Deep Fake Detection - Detection of GAN Generated Faces 

##  Datasets

In this project the datasets can be organized in two ways:

1) By ID:

dataset_path
	- id1
		-img_0001.jpg
		...
		-img_1000.jpg
	- id2
		-img_0001.jpg
		...
		-img_1000.jpg
	- idn
		-img_0001.jpg
		...
		-img_1000.jpg

2) By Image:

dataset_path
	-img_0001.jpg
	...
	-img_1000.jpg



## Pre-Processing

####  align_face_byid.py 

Aligns dataset images by identities

- 1) Detect a face; 
- 2) Facial landmark localization; 
- 3) Rotates and scales the face to ensure that the eyes are located in a specific (x,y) position

INPUT PARAMS

DEBUG = 0

shape_predictor_model = 'shape_predictor_68_face_landmarks.dat'

dataset_path = '/media/jcneves/DATASETS/100K_FAKE/byid_original/'

output_path = '/media/jcneves/DATASETS/100K_FAKE/byid_alignedlib_0.3/'
eyes_margin = 0.3


## Dataset Creation

#### create_real_to_fake_dataset.py 

Creates a dataset with two classes: 0 - real face images, 1 - fake face images

INPUT PARAMS

IMAGES_PER_CLASS = 10000 # number of images in each class

subsets = ['train', 'val', 'test'] # subsets which will be created

subsets_prop = [0.75, 0.2, 0.05] # percentage of images used in each subset

ds1_path = path to the fake dataset

ds2_path = path to the real dataset

out_path = path where the dataset will be created


#### create_real_to_fake_dataset_mixed.py

Creates a dataset with two classes: 0 - real face images, 1 - fake face images

It requires four datasets, to ensure that the training data and test data belong to different datasets

IMAGES_PER_CLASS = 10000

subsets = ['train', 'val', 'test']

subsets_prop = [0.05, 0.2, 0.75]

ds1_path_train = '/media/jcneves/DATASETS/100K_FAKE/byimg_alignedlib_0.3/'

ds1_path_test = '/media/jcneves/DATASETS/NVIDIA_FakeFace/byimg_alignedlib_0.3/'

ds2_path_train = '/media/jcneves/DATASETS/CASIA-WebFace/byid_alignedlib_0.3/'

ds2_path_test = '/media/jcneves/DATASETS/VGG_FACE_2/byid_alignedlib_0.3_train/'

out_path = '/media/jcneves/DATASETS/real2fake_mixed/'

### Train CNN

Code for training a CNN to idenitfy between N classes

Dataset format:

data/
  - train/
      - class_1 folder/
          - img1.png
          - img2.png
      - class_2 folder/
      .....
      - class_n folder/
  - val/
      - class_1 folder/
      - class_2 folder/
      ......
      - class_n folder/

INPUT PARAMS

data_dir = "/media/jcneves/DATASETS/real2fake_mixed/"

input_shape = 200

batch_size = 32

mean = [0.5, 0.5, 0.5]

std = [0.5, 0.5, 0.5]

scale = 224

use_parallel = False

use_gpu = False

epochs = 2