GAN-image-detection

This repository contains a GAN-generated image detector developed to distinguish real images from synthetic ones.

The detector is based on an ensemble of CNNs. The backbone of each CNN is the EfficientNet-B4. Each model of the ensemble has been trained in a different way following the suggestions presented in this paper in order to increase the detector robustness to compression and resizing.

Run the detector

Prerequisites

Create and activate the conda environment

conda env create -f environment.yml
conda activate gan-image-detection

Download the model’s weights from this link and unzip the file under the main folder

wget https://www.dropbox.com/s/g1z2u8wl6srjh6v/weigths.zip
unzip weigths.zip

Test the detector on a single image

We provide a simple script to obtain the model score for a single image.

python gan_vs_real_detector.py --img_path $PATH_TO_TEST_IMAGE

Performance

We provide a notebook with the script for computing the ROC curve for each dataset.

How to cite

Training procedures have been carried out following the suggestions presented in the following paper.

Plaintext:

S. Mandelli, N. Bonettini, P. Bestagini, S. Tubaro, "Training CNNs in Presence of JPEG Compression: Multimedia Forensics vs Computer Vision", IEEE International Workshop on Information Forensics and Security (WIFS), 2020, doi: 10.1109/WIFS49906.2020.9360903.

Bibtex:

@INPROCEEDINGS{mandelli2020training,
  author={Mandelli, Sara and Bonettini, Nicolò and Bestagini, Paolo and Tubaro, Stefano},
  booktitle={IEEE International Workshop on Information Forensics and Security (WIFS)}, 
  title={Training {CNNs} in Presence of {JPEG} Compression: Multimedia Forensics vs Computer Vision}, 
  year={2020},
  doi={10.1109/WIFS49906.2020.9360903}}