.png){kind=icon}
Return
Transferable adversarial examples can efficiently fool topic models
.png){kind=icon}
PRE
AI
DOI：10.1016/j.cose.2022.102749.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Intelligent text systems have been shown vulnerable to adversarial attacks. A popular way to enhance the robustness is to leverage transferable adversarial examples. As one of the most important statistical models, topic model is widely used in various applications. However, whether adversarial examples exhibit good transferability across topic models has never been explored. In this paper, we propose an effective adversarial example generator TopicAttack to disturb the inference of a target victim topic model and an ensemble algorithm TopicAttack+ to further increase transferability. TopicAttack exploits the words of high importance and constructs adversarial examples via synonym transformation. TopicAttack+ searches the optimal model ensemble given a set of prescribed substitutes, then generate adversarial examples of stronger transferability. Numerically, we demonstrate the effectiveness of the proposed method on the benchmark datasets AP, NIPS and 20News. In particular, given eight benchmark substitute models, TopicAttack disturbs them all effectively by achieving high Kullback-Leibler (KL) divergence; and TopicAttack+ significantly improves the average transferability performance on NIPS from 1.308 to 1.695. (C) 2022 Elsevier Ltd. All rights reserved.
Keywords:
Topic model
Adversarial attack
Transferability
KL divergence
Ensemble attack
Journal
C
IF:
5.4
Papers:
4.6K
Citations:
1.4W
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSaveShareSave
ShareSaveShareSave
View more