D-NEXUS: Defending text networks using summarization

作者：

Highlights：

• We propose a novel summarization based defence, D-NEXUS, against attacks on the sentiment analysis models.

• We are the first to study the applicability of summarization for defending the sentiment analysis models.

• Unlike the existing spelling correction based defenses, D-NEXUS successfully mitigates the state-of-the-art attacks involving word replacement, insertion, and deletion strategies.

• Extensive experiments on publicly available datasets show that D-NEXUS successfully defends against state-of-the-art attacks.

• D-NEXUS is model-agnostic and can provide defense in a time-efficient manner.

摘要

•We propose a novel summarization based defence, D-NEXUS, against attacks on the sentiment analysis models.•We are the first to study the applicability of summarization for defending the sentiment analysis models.•Unlike the existing spelling correction based defenses, D-NEXUS successfully mitigates the state-of-the-art attacks involving word replacement, insertion, and deletion strategies.•Extensive experiments on publicly available datasets show that D-NEXUS successfully defends against state-of-the-art attacks.•D-NEXUS is model-agnostic and can provide defense in a time-efficient manner.

论文关键词：Sentiment analysis,Natural language processing,Adversarial defense,Transformers,Adversarial attack,Language summarization

论文评审过程：Received 13 January 2022, Revised 8 May 2022, Accepted 27 June 2022, Available online 30 June 2022, Version of Record 6 July 2022.

论文官网地址：https://doi.org/10.1016/j.elerap.2022.101171