On the use of Benford's law to detect GAN-generated images

This is the official implementation of the paper On the use of Benford's law to detect GAN-generated images, accepted to ICPR2020 and currently available on arXiv.

This repository is currently under maintenance, if you are experiencing any problems, please open an issue.

Getting started

Prerequisites

1. Install conda

2. Create the benford-gan environment with environment.yml:

   conda env create -f environment.yml
   conda activate benford-gan

3. Install pyjpeg-dct package [1]:

   cd pyjpeg-dct
   pip install -e .

   If you fail in installing pyjpeg-dct, you probably miss the python3.6-dev package for your OS. Google is your friend :)

The whole pipeline

Besides the feature extraction part, the rest of the pipeline is really dependent on the dataset used in the paper. Keep this in mind if you try to replicate the paper results. If you just want the feature extraction functions go straight to extract_first_digit_hist.py and extract_features_from_hist.py.

Feature extraction

1. Insert dataset root (folder containing images) into params.py. You probably also need to modify/delete lines 5-11 according to the dataset you want to use and its location on your machine.
2. Extract first digit histograms extract_first_digit_hist.py.
3. Compute divergence features from histograms extract_features_from_hist.py

Train

4. Run Random Forest classifier* rf_combinations_logo.py

Test

5. Test* rf_combinations_logo_test_only.py

* NB: to obtain results for test_compression=False run rf_combinations_logo.py, to obtain results for test_compression=True run rf_combinations_logo_test_only.py

SOTA replication

1. Build db cnn_build_db.py
2. Finetune Xception cnn_finetuning.py
3. Extract cooccurrences extract_cooccurrences.py
4. Train SVM on cooccurrences svm_cooccurrences.py
5. Train RF on cooccurrences rf_cooccurrences.py

Credits

Image and Sound Processing Lab - Politecnico di Milano

• Nicolò Bonettini
• Paolo Bestagini
• Simone Milani
• Stefano Tubaro

---

[1] Courtesy of https://github.com/wartmanm/pyjpeg-dct