Data-Free Substitute Training

DaST: Data-free Substitute Training for Adversarial Attacks
A novel approach that can train a substitute of the target DL model without the need of real samples. It reveals the risk that attackers can easily train a substitute of the target model to generate effective adversarial attacks.
Mingyi Zhou, Jing Wu, Yipeng Liu, Shuaicheng Liu, Ce Zhu
PDF Cite Code DOI
DaST: Data-free Substitute Training for Adversarial Attacks