Differential Privacy DeepFakes

Official Repository for "Analysis of Obfuscation of Deepfake Images in Differential Privacy Settings" published in KAIA2022.

Prerequisites

• python >= 3.7.0
• PyTorch >= 1.8.0
• CUDA 11.4 (on RTX A5000 x 4)

pip install -r requirements.txt

---

Applying differential privacy to images

In this repository, we apply three differential privacy methods to DeepFake images named pixelation, SVD and snow. For each method, we provide 3 jupyter notebook files as follows:

• differential privacy pixelation.ipynb

• differential privacy svd.ipynb

• differential privacy snow.ipynb

In each notebook, there is a header where it states "convert deepfake images to DP". By changing the directory of the deepfake image and directory for saving images, one may run the code below in each DP notebook.

import os
df_imgdir = glob('/home/data/deepfake_privacy/FF_original/FaceForensics++/Face2Face/test/real/*/')
fake_img_dir = '/home/data/deepfake_privacy/FF_original/FaceForensics++/Face2Face/test/real/'
k = 7
epsilon = 0.01

for idx, folder in enumerate(df_imgdir):
    
    # extract the last part of the folder for joining path
    imgfolder = folder.split('/')[-2]
    if not os.path.exists('/home/data/deepfake_privacy/ff_priv/Face2Face/pixel/test/real/{0}'.format(imgfolder)):
        os.makedirs(os.path.join('/home/data/deepfake_privacy/ff_priv/Face2Face/pixel/test/real/', imgfolder))
    subfolder = os.path.join('/home/data/deepfake_privacy/ff_priv/Face2Face/pixel/test/real/', imgfolder)
    
    # subfolder = location of saved area
    # globfolder = location of train data
    globfolder = fake_img_dir + imgfolder + '/*'
    
    print(globfolder, 'glob folder')

    # generating pixelation image per every image of each subfolder
    for idx2, img in enumerate(glob(globfolder)):
        image = Image.open(img)
        a = pillow_to_numpy(image)
        dpimg = dp_pixelate_images_singleimg(a, target_h, target_w, m, eps)
        dpimg = dpimg.astype('uint8')
    

        cv2.imwrite(subfolder+'/folder_{0}_pixel_test_{1}.jpg'.format(idx, idx2),cv2.cvtColor(dpimg, cv2.COLOR_RGB2BGR))
        if idx2 == 10:
            print(subfolder, 'sub folder')
            
print('######finished conversion!!##########')

---

Running Train Code

Adjust the hyperparameters in train.py or customize your own in train.sh

python tools/train.py

sh tools/train.sh

Running Test Code

sh tools/test.sh

The code for training/testing works in data parallel. You may adjust the number of GPUs to use. In this work, we used upto 4 GPUs in training and testing.