README.md

Fourier Synthetic Image Classifiers

About the Code

The section uses high frequency features extracted using the official matlab implementation by Dzanic et al. [1]. The matlab code can be found [2]. The code is clearly documented.

Hyper-parameters

For all reported CelebA/ LSUN/ StarGAN experiments,

• For feature extraction, we use the official script. We tried with different hyper-parameters, but in all instances there was no substantial variation in classifier bypassing values.
• All results reported are averaged across 10 independent runs
• For KNN classifier, we use k=5 as official implementation
• For SVM, we use RBF kernel
• For MLP, we use a 2-layer architecture with sigmoid activation functions (Binary classification learning problem).

Running the Code

1. Create a new virtual environment and install all the dependencies

   pip install -r requirements.txt

2. Use the generateddata.m script from the official repository [2] to extract the high frequency features. (3 dimensional features)

3. To run the classifiers, first organize the directory as follows and place the .mat files inside

```
├── fits
│   ├── celeba
│   │   ├── real
│   │   ├── gan
│   │   ├──   ├── BASELINE
│   │   ├──   ├── B.1.5
│   │   ├──   ├── Z.1.5
│   │   ├──   ├── ....

│   │   ├── lsgan
│   │   ├──   ├── BASELINE
│   │   ├──   ├── B.1.5
│   │   ├──   ├── Z.1.5
│   │   ├──   ├── ....

│   │   ├── wgan
│   │   ├──   ├── BASELINE
│   │   ├──   ├── B.1.5
│   │   ├──   ├── Z.1.5
│   │   ├──   ├── ....
```

4. To train and test with classifiers,

• use classifiers/knn_classifier.py for KNN classification
• use classifiers/svm_classifier.py for SVM classification
• use classifiers/mlp_classifier.py for MLP classification

References

[1] Fourier Spectrum Discrepancies in Deep Network Generated Images, Tarik Dzanic, Karan Shah, and Freddie D. Witherden, In NeurIPS, 2020

[2] https://github.com/tarikdzanic/FourierSpectrumDiscrepancies)