SPRITZ-PS: Validation of Synthetic Face Images Using a Large Dataset of Printed Documents

Citation Author(s):
Ehsan
Nowroozi
Ravensbourne University London
Seyedsadra
Seyedshoari
University of Padova, Italy
Mauro
Conti
University of Padova, Italy
Yoosef
Habibi
University of Padova, Italy
Submitted by:
Ehsan Nowroozi
Last updated:
Tue, 02/20/2024 - 15:14
DOI:
10.21227/f1fx-sq21
Data Format:
*.jpg (zip) *.png (zip)*.txt (zip)
Research Article Link:
VIPPrint: Validating Synthetic Image Detection and Source Linking Methods on a Large Scale Dataset of Printed Documents
License:
Creative Commons Attribution
577 Views
Categories:
Artificial Intelligence
Signal Processing
Digital signal processing
Machine Learning
Other
Security
Image Processing
Computer Vision
Keywords:
digital image forensics; source identification; GAN-generated image detection; machine and deep learning; Cybersecurity
0
0 ratings - Please login to submit your rating.
ACCESS DATASET CITE SHARE/EMBED

Abstract 

*** The paper published on Multimedia Tools and Applications (Springer) - 2024 ***

*** Title: "SPRITZ-PS: Vlaidation of Synthetic Face Images Using A Large Dataset of Printed Docuemnts"***

*** Authors: Ehsan Nowroozi, Yoosef Habibi, and Mauro Conti ***

----------------------------------------------------------------------------------------------------------

GAN-generated faces look challenging to distinguish from genuine human faces. As a result, because synthetic images are presently being used as profile photos for fake identities on social media, they may have serious social consequences. Iris pattern anomalies might expose GAN-generated facial photos. When photographs are printed and scanned, it becomes more difficult to distinguish between genuine and counterfeit since fraudulent images lose some of their qualities. We created a new collection of iris images from printed and scanned documents by segmenting pupils from face images to address these concerns. We employed Dlib, which provides 68 facial landmarks, and EyeCool to extract both left and right irises from a full face image to segment the iris portion. Nevertheless, due to eyelid occlusion, the extracted iris images are not entirely shaped. We have to fill missing pixels of extracted iris since there are no incomplete iris in the actual world and the incomplete image is not an adequate input to train deep neural networks. We employ the Hypergraph convolution-based image inpainting approach to do this. The detailed relationship between the iris images was determined using hypergraph convolution.

Instructions: 

When using the dataset, don't forget to cite our paper:

@misc{nowroozi2023spritzps,
      title={Spritz-PS: Validation of Synthetic Face Images Using a Large Dataset of Printed Documents}, 
      author={Ehsan Nowroozi and Yoosef Habibi and Mauro Conti},
      year={2023}, eprint={2304.02982},  archivePrefix={arXiv}, primaryClass={cs.CV} }

 AND

@Article{jimaging7030050, AUTHOR = {Ferreira, Anselmo and Nowroozi, Ehsan and Barni, Mauro}, TITLE = {VIPPrint: Validating Synthetic Image Detection and Source Linking Methods on a Large Scale Dataset of Printed Documents}, JOURNAL = {Journal of Imaging}, VOLUME = {7}, YEAR = {2021}, NUMBER = {3}, ARTICLE-NUMBER = {50}, URL = {https://www.mdpi.com/2313-433X/7/3/50}, ISSN = {2313-433X}, DOI = {10.3390/jimaging7030050} }

  AND 

@article{NOWROOZI2021102092,
title = {A survey of machine learning techniques in adversarial image forensics},
journal = {Computers & Security},
volume = {100},
pages = {102092},
year = {2021},
issn = {0167-4048},
doi = {https://doi.org/10.1016/j.cose.2020.102092},
url = {https://www.sciencedirect.com/science/article/pii/S0167404820303655},
author = {Ehsan Nowroozi and Ali Dehghantanha and Reza M. Parizi and Kim-Kwang Raymond Choo}
}

The Dataset consists of three main sub-folders. The initial dataset plus iris segmentation and iris reconstruction.

Dataset Files

LOGIN TO ACCESS DATASET FILES

QUESTIONS?