Closed NarendranM08 closed 2 years ago
Are you asking for a front-end? The code in this repository allows training and executing a backend. But is mainly intended to back up the results we report in the paper. We do not provide a frontend here, but it could of course be built using Django for example.
No , like labelling from the dataset that the images are real or fake
The label is assigned based on the file path in https://github.com/gan-police/frequency-forensics/blob/c7e12968b14e0f609ab7e5d6df7d4090c85819f3/src/freqdect/prepare_dataset.py#L80 . For this approach to work, we require a specific file path setup. I.e for example: source_data ├── A_original ├── B_CramerGAN ├── C_MMDGAN ├── D_ProGAN └── E_SNGAN
Should we label the folder by ourselves or running the project code
This has to happen by hand. For the code to work the folders must be arranged properly. A zipped example folder is available at: https://drive.google.com/file/d/1MOHKuEVqURfCKAN9dwp1o2tuR19OTQCF/view?usp=sharing
So from the drive that A_ffhq, b_Stylegan we make into five directories source_data ├── A_original ├── B_CramerGAN ├── C_MMDGAN ├── D_ProGAN └── E_SNGAN
No, FFHQ means Flickr Face High Quality and StyleGAN is a GAN, see section 5.1 of the paper for more information. That means the downloaded archive has images from two sources and works as-is. The file structure with five sources (A,B,C,D,E) is what we used in section 5.2 in the paper.
So based on folder labels it presents that the image is real or fake
Only the training data preprocessing pipeline is set up using folder labels. Once a classifier is trained the folder structure is no longer relevant. Feel free to play with the pre-trained network from https://github.com/gan-police/frequency-forensics/issues/22 . For in-depth information regarding the underlying identification mechanism please read the paper.
Is the wavelet-packet creation done in fourier_math.py
No, the packets are created in https://github.com/gan-police/frequency-forensics/blob/main/src/freqdect/wavelet_math.py .
Can you share the screenshot of final output where the images are labeled as fake or real