Benchmarking the Time Cost of Adversarial Attacks on GANs

View at publisher Summarise with VIRAT
Authors: Khushi Sayani, Het Gohil, Hussain Degani, Hezal Lopes
Cited by: 0
References: 15

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Generative Adversarial Networks (GANs) are now a foundation of contemporary generative AI, driving everything from art to synthetic data generation. Yet, their susceptibility to adversarial attacks, as well as the rapidity with which these attacks can compromise them, is still a vital area of research. Although extensive research has been conducted on the success of attacks, there has not been a great deal of emphasis on the time it takes for an attack to be effective, a very important consideration in practical environments where prompt detection and prevention are absolutely essential. This paper introduces a benchmarking platform intended to quantify the time expense of GAN adversarial attacks. By considering three predominant attack categories, latent-space perturbations, training-time poisoning (backdoors), and membership inference, we not only consider their efficacy but also whether they can attack a model quickly. Our results indicate a speed vs. impact trade-off: the Latent Perturbation attack is quick to perform but infects individual outputs, whereas the time poisoning attack, take longer to seed but can infect models entirely or spill training data and finally Membership Inference attacks are quick to perform and can reveal if an individual’s data was used in a model’s training set. This benchmark seeks to offer a down-to-earth, time-savvy perspective for researchers and practitioners alike to more effectively evaluate GAN vulnerabilities and design timely defensive measures.

Fields of Study & Subjects

Adversarial Robustness in Machine Learning Digital Media Forensic Detection Generative Adversarial Networks and Image Synthesis

Publication details

Journal / source
Biennial International Conference on Nascent Technologies in Engineering (ICNTE)
Pages
1–8
Publication date
16 January 2026
Language
EN