Black-box Adversarial Attacks against Dense Retrieval Models: A Multi-view Contrastive Learning Method

By Yu-An Liu et al

Published on Oct. 21, 2023

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Related Work
3. Problem Statement
4. Our Method

Summary

Neural ranking models (NRMs) and dense retrieval (DR) models have given rise to substantial improvements in overall retrieval performance. In this paper, the authors introduce the adversarial retrieval attack (AREA) task, targeting DR models. They propose a multi-view contrastive learning-based adversarial retrieval attack (MCARA) to deceive DR models into retrieving a target document outside the initial set of candidate documents. The method involves generating multi-view representations and utilizing view-wise contrastive loss to mislead the DR model with small indiscernible text perturbations. Experimental results show the effectiveness of MCARA in promoting the target document into the candidate set with high attack success. The paper also discusses the differences in attacking DR models compared to NRMs, emphasizing the need for tailored attack strategies for DR models.