RobustEncoder: Leveraging K-Means clustering technique to defend NLP models against backdoor attacks

Document Type

Conference Proceeding

Source of Publication

2024 6th International Conference on Blockchain Computing and Applications (BCCA)

Publication Date

11-29-2024

Abstract

As machine learning (ML) systems become increasingly integrated into real-world applications for sensitive tasks, ensuring the security and privacy of these models becomes paramount. Deep Neural Networks (DNNs), in particular, are susceptible to backdoor attacks, where adversaries manipulate training data by inserting specially crafted samples. While the NLP community has extensively studied backdoor attacks, there remains a gap in effective defense mechanisms. To address this, we propose RobustEncoder, a novel approach leveraging K-Means clustering to detect and mitigate backdoor attacks in text-based models. Our method demonstrates significant efficacy in identifying and neutralizing backdoor triggers, as evidenced by extensive empirical evaluations. Additionally, we propose potential applications of blockchain technology to further enhance the security and integrity of the defense mechanisms in future implementations.

ISBN

979-8-3503-5153-8

Publisher

IEEE

Volume

00

First Page

179

Last Page

188

Disciplines

Computer Sciences

Keywords

Backdoor attacks, K-Means clustering, NLP models, Defense mechanisms, Security

Indexed in Scopus

no

Open Access

no

Link to Full Text

Share

COinS