gan_fingerprint_removal

Introduction

This repo provides the implementation of a GAN fingerprint removal approach, as described in

@misc{1911.05351, Author = {João C. Neves and Ruben Tolosana and Ruben Vera-Rodriguez and Vasco Lopes and Hugo Proença}, Title = {Real or Fake? Spoofing State-Of-The-Art Face Synthesis Detection Systems}, Year = {2019}, Eprint = {arXiv:1911.05351}, }

Remove GAN fingerprint from a folder of images

python gan_fingerprint_removal.py IMAGE_PATH TRANSFORMED_IMAGE_PATH AE_MODEL_PATH
python gan_fingerprint_removal.py F:\\FACE_DATASETS\\NVIDIA_FakeFace\\byimg_alignedlib_0.3 fingerprint_removal\\ae_models\\ae32\\ae32_epoch100.pytorch

Datasets

The TPDNE dataset comprises 150K synthetic face images generated by the StyleGAN.

http:/socia-lab.di.ubi.pt/~jcneves/TPDNE/byimg_original.zip

A subset of the this dataset comprising only frontal aligned faces is also available.

http:/socia-lab.di.ubi.pt/~jcneves/100KFaces/byimg_alignedlib_0.3.zip

The 100K-Faces dataset comprises 100K synthetic face images generated by the StyleGAN.

https://generated.photos/

A subset of this dataset ccomprising only frontal aligned faces is also available.

Paper experiments

This section details how to replicate the experiments carried out in the paper

Controlled Scenario

Generate a dataset with two classes: real faces, synthetic faces

python create_real_to_fake_dataset.py REAL_FACES_DATASET_PATH SYNTHETIC_FACES_DATASET_PATH NEW_DATASET_PATH 
python create_real_to_fake_dataset.py  E:\FACE_DATASETS\VGG_FACE_2\byid_alignedlib_0.3_train\ E:\FACE_DATASETS\NVIDIA_FakeFace\byimg_alignedlib_0.3\ E:\gan_fingerprint_removal\data\real2fake_VF2_TPDNE

Train a CNN on the dataset created

python ../../train_pytorch_gpu.py NEW_DATASET_PATH
python ../../train_pytorch_gpu.py E:\gan_fingerprint_removal\data\real2fake_VF2_TPDNE

Test the model trained on the test set of dataset created

python ../../test_pytorch_gpu.py MODEL_PATH NEW_DATASET_PATH
python ../../test_pytorch_gpu.py E:\gan_fingerprint_removal\data\real2fake_VF2_TPDNE E:\gan_fingerprint_removal\data\real2fake_VF2_TPDNE

Run all paper experiments

python run_all_experiments.py
run_all_experiments.bat

In The Wild (Unseen Dataset)

Transform a dataset with two classes: real faces, synthetic faces

python create_real_to_fake_dataset_mixed.py REAL_FACES_DATASET(1)_PATH SYNTHETIC_FACES_DATASET(1)_PATH REAL_FACES_DATASET(2)_PATH SYNTHETIC_FACES_DATASET(2)_PATH NEW_DATASET_PATH 
python create_real_to_fake_dataset_mixed.py E:\FACE_DATASETS\VGG_FACE_2\byid_alignedlib_0.3_train\ E:\FACE_DATASETS\NVIDIA_FakeFace\byimg_alignedlib_0.3\ E:\FACE_DATASETS\VGG_FACE_2\byid_alignedlib_0.3_train\ E:\FACE_DATASETS\100K_FAKE\byimg_alignedlib_0.3\ E:\gan_fingerprint_removal\data\real2fake_VF2_TPDNE_train_VF2_100F_test

Test the model trained on the original dataset (refer to controlled scenario experiment)

python ../../test_pytorch_gpu.py MODEL_PATH TRANSFORMED_DATASET_PATH
python ../../test_pytorch_gpu.py E:\gan_fingerprint_removal\data\real2fake_VF2_TPDNE_train_VF2_100F_test_R32 E:\gan_fingerprint_removal\data\real2fake_VF2_TPDNE

Run all paper experiments

python run_all_experiments_downsize.py
run_all_experiments_downsize.bat

In The Wild (Resolution)

Transform a dataset with two classes: real faces, synthetic faces

python transform_dataset_downsize.py IMG_RESOLUTION DATASET_PATH TRANSFORMED_DATASET_PATH 
python transform_dataset_downsize.py 32 E:\gan_fingerprint_removal\data\real2fake_VF2_TPDNE_train_VF2_100F_test E:\gan_fingerprint_removal\data\real2fake_VF2_TPDNE_train_VF2_100F_test_R32

Test the model trained on the original dataset (refer to controlled scenario experiment)

python ../../test_pytorch_gpu.py MODEL_PATH TRANSFORMED_DATASET_PATH 
python ../../test_pytorch_gpu.py E:\gan_fingerprint_removal\data\real2fake_VF2_TPDNE_train_VF2_100F_test_R32 E:\gan_fingerprint_removal\data\real2fake_VF2_TPDNE

Run all paper experiments

python run_all_experiments_downsize.py
run_all_experiments_downsize.bat