Implementation of the paper: Use of a Capsule Network to Detect Fake Images and Videos, which is an updated version of the previous work: Capsule-Forensics: Using Capsule Networks to Detect Forged Images and Videos (ICASSP 2019).
You can clone this repository into your favorite directory:
$ git clone https://github.com/nii-yamagishilab/Capsule-Forensics-v2
Databases folder, where you can place your training, evaluation, and test set:
./databases/<faceforensicspp; cgvsphoto_patches; cgvsphoto_full; replay_attack>/<train; validation; test>
Checkpoint folder, where the training outputs will be stored:
./checkpoints/<binary_faceforensicspp (without NeuralTextures); binary_faceforensicspp_v2_full (with NeuralTextures); multiclass_faceforensicspp; cgvsphoto; replay_attack>
Pre-trained models for the FaceForensics++ database (includes Real, DeepFakes, Face2Face, FaceSwap, and NeuralTextures), the CGvsPhoto database, and the Replay-Attack database (with settings described in our paper) are provided in the checkpoints folder.
In case of the FaceForensics++ database, it need to be pre-processed to crop facial area. We recommend using an image size of 300 x 300 as the input.
Scripts for data preprocessing can be found at datascripts
directory.
Note: Parameters with detail explanation could be found in the corresponding source code.
Training the Capsule-Forensics-v2 using binary classification on the FaceForensics++ database:
$ python train_binary_ffpp.py
Training the Capsule-Forensics-v2 using multiclass classification on the FaceForensics++ database:
$ python train_multiclass_ffpp.py
Training the Capsule-Forensics-v2 on the CGvsPhoto database:
$ python train_cgvsphoto.py
Training the Capsule-Forensics-v2 on the Idiap Replay-Attack database:
$ python train_replay_attack.py
Note: Parameters with detail explanation could be found in the corresponding source code.
Binary classification on images:
$ python test_binary_ffpp.py
Binary classification on videos (extracted as frames):
$ python test_vid_binary_ffpp.py
Multiclass classification on images:
$ python test_multiclass_ffpp.py
Multiclass classification on images with detail results on each class:
$ python test_multiclass_detail_ffpp.py
Multiclass classification on videos (extracted as frames):
$ python test_vid_multiclass_ffpp.py
Testing on patches:
$ python test_cgvsphoto.py
Testing on full images:
$ python test_cgvsphoto_full.py
Testing on images:
$ python test_replay_attack.py
This research was supported by JSPS KAKENHI Grants JP16H06302 and JP18H04120 and by JST CREST Grant JPMJCR18A6, Japan.
H. H. Nguyen, J. Yamagishi, and I. Echizen, “Use of a Capsule Network to Detect Fake Images and Videos,” arXiv preprint arXiv:1910.12467. 2019 Oct 29.