Towards Adversarially Robust AI-Generated Image Detection

Annan Zou

2023

Abstract

Over the last few years, Artificial Intelligence Generated Content (AIGC) technology has rapidly matured and garnered public attention due to its ease of use and quality of results. However, due to these characteristics, forged images generated by AIGC technology have a high potential of being misused and causing negative social consequences. While AI-based tools can identify AI-generated images with reasonable accuracy, these models did not consider the factor of adversarial robustness or resistance against intentional attacks. This paper empirically evaluated several existing AIGC detection models’ adversarial robustness under select attack setups. Overall, it is discovered that even naked-eye unnoticeable perturbation of source images can consistently cause sharp drops in performance in all the models in question. This study proposes constructing a Convolutional Neural Network (CNN) based AIGC classifier model with additional adversarial training using a combination of transformation-based and l-based adversarial examples constructed with existing AIGC data. This paper uses clean and adversarial data sets to test the performance of the resulting model. The results show the model’s remarkable robustness to the adversarial attack techniques described above while maintaining relative accuracy on clean datasets.

Paper Citation

in Harvard Style

Zou A. (2023). Towards Adversarially Robust AI-Generated Image Detection. In Proceedings of the 1st International Conference on Data Analysis and Machine Learning - Volume 1: DAML; ISBN 978-989-758-705-4, SciTePress, pages 380-385. DOI: 10.5220/0012816300003885

in Bibtex Style

@conference{daml23,
author={Annan Zou},
title={Towards Adversarially Robust AI-Generated Image Detection},
booktitle={Proceedings of the 1st International Conference on Data Analysis and Machine Learning - Volume 1: DAML},
year={2023},
pages={380-385},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012816300003885},
isbn={978-989-758-705-4},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 1st International Conference on Data Analysis and Machine Learning - Volume 1: DAML
TI - Towards Adversarially Robust AI-Generated Image Detection
SN - 978-989-758-705-4
AU - Zou A.
PY - 2023
SP - 380
EP - 385
DO - 10.5220/0012816300003885
PB - SciTePress