Closed surya-venkatesan closed 2 years ago
Yes, you can see for yourself! To reproduce the FFHQ-style-gan pair: Download the zipped data from https://drive.google.com/file/d/1MOHKuEVqURfCKAN9dwp1o2tuR19OTQCF/view?usp=sharing . You can follow the data perperation steps and train your own classifier.
To do so extract the zip file into a data
folder and run:
$ python -m freqdect.prepare_dataset ./data/source_data/ --log-packets
$ python -m freqdect.prepare_dataset ./data/source_data/
the prepare the data. Once the preparation has finished train a classifier using:
$ python -m freqdect.train_classifier \
--data-prefix ./data/source_data_log_packets_haar_reflect_3 \
--calc-normalization \
--features packets
You should end up with a log folder containing the classifier and intermediated loss logs.
Will it indicate that the input image is real or fake or it just makes an analysis
It will indicate if the image is real or fake, based on the wavelet packet coefficients. In the paper we have measured the performance on standard scientific data-sets and found promising performance However, the code is not production-ready. Two goals of this repository are to prove that deepfake identification based on wavelet packets is possible and that the packet representation has advantages compared to the pixel form.
May I know what type of architecture are used in it
Sure, you can read the description in the paper or look at the code ( https://github.com/gan-police/frequency-forensics/blob/main/src/freqdect/models.py ).
what type of cnn model is used ? is it Resnext 50?
Is this a complete code of deepfake image detection using wavelet packets