Variable Step-Size Adversarial Attacks Against Deep Learning Based End-to-End Autoencoder

Han Jiang

doi:10.1109/TVT.2025.3582045

Variable Step-Size Adversarial Attacks Against Deep Learning Based End-to-End Autoencoder

2025-06-25

0

PRE

AI

DOI：10.1109/TVT.2025.3582045

Original

Original request for help

Save

Abstract

En 中文

Unreasonable iteration number and step-size setting limit the performance of adversarial attacks. To deal with this issue, we propose a variable step-size double-iteration (VSDI) white-box attack. Due to that the attack performance may be further affected by complex channels, a more practical scenario is adopted, taking the channels from both the transmitter and adversary towards the receiver into consideration. Furthermore, considering the input-agnostic and non-synchronous characteristics, we extend the proposed VSDI attack to a black-box model via a heuristic approach. Numerical results demonstrate that the proposed adversarial attacks outperform existing ones in terms of block error rate.

Keywords:

Adversarial attacks

deep learning

end-to-end autoencoder

variable step size

Variable Step-Size Adversarial Attacks Against Deep Learning Based End-to-End Autoencoder

Abstract

Journal

Organization