Expose Backdoors on the Way: A Feature-Based Efficient Defense against Textual Backdoor Attacks

Chen, Sishuo; Yang, Wenkai; Zhang, Zhiyuan; Bi, Xiaohan; Sun, Xingming

doi:10.18653/v1/2022.findings-emnlp.47

Cited by 2 publications

(1 citation statement)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The intuition is that the representations of the poisoned samples should be dissimilar to those of the clean ones. Regarding test-stage defences, one can leverage either the victim model (Gao et al, 2019;Yang et al, 2021;Chen et al, 2022b) or an external model (Qi et al, 2021a) to filter out the malicious inputs according to their misbehaviour. Our approach belongs to the family of training-stage defences.…”

Section: Related Workmentioning

confidence: 99%

Mitigating Backdoor Poisoning Attacks through the Lens of Spurious Correlation

He,

Xu,

Wang

et al. 2023

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

View full text Add to dashboard Cite

Modern NLP models are often trained over large untrusted datasets, raising the potential for a malicious adversary to compromise model behaviour. For instance, backdoors can be implanted through crafting training instances with a specific textual trigger and a target label. This paper posits that backdoor poisoning attacks exhibit spurious correlation between simple text features and classification labels, and accordingly, proposes methods for mitigating spurious correlation as means of defence. Our empirical study reveals that the malicious triggers are highly correlated to their target labels; therefore such correlations are extremely distinguishable compared to those scores of benign features, and can be used to filter out potentially problematic instances. Compared with several existing defences, our defence method significantly reduces attack success rates across backdoor attacks, and in the case of insertion-based attacks, our method provides a near-perfect defence. 1

show abstract