halidkarim/FFR_FD-Effective-and-Fast-Detection-of-DeepFakes-Based-on-Feature-Point-Defects

Code for paper

FFR_FD: Effective and Fast Detection of DeepFakes Based on Feature Point Defects

[Code for paper 'FFR_FD: Effective and Fast Detection of DeepFakes Based on Feature Point Defects']

Experiment Environment

Our development env is shown in:

requirements.txt

1) Datasets

Download Original Datasets to the corresponding ./datasets/Videos/ folders.

• DeepfakeTIMIT_HQ
• DeepfakeTIMIT_LQ
• UADFV
• FF++_c40
• FF++_raw
• DFD
• DFDC
• CelebDF_V2

Split Datasets

Split Videos to the training set and testing set:

cd datasets
python split_video_datasets.py -h

###extract the frames:

python extract_frames.py -h

use the S3Fd Detector and Fan Aligner to extract facial images. [faceswap github].(https://github.com/deepfakes/faceswap)

#2) Feature Points Statistics First, download the shape_predictor_68_face_landmarks.dat to the ./

cd ..
python feature_point_statistics.py -h

#3) FFR_FD

cd 'construct_FFR_FD_for_datasets'
python FFR_FD_no_ave_train_set.py
python FFR_FD_no_ave_test_set.py -h
python FFR_FD_ave_train_set.py -h
python FFR_FD_ave_test_set.py -h 

#3) Differences in FFR_FD

cd "differences_in_FFR_FD"
python statistics_differences_of_FFR_FD.py -h

#4) Train and Test

cd train_and_test
python train_and_test.py -h

###Generalization Test:

python generalization_test.py -h

#6)features importance：

python features_importances.py -h