"Someone Hid It": Query-Agnostic Black-Box Attacks on LLM-Based Retrieval

TL;DR

This paper introduces a practical black-box attack method on LLM-based retrieval systems that does not require access to victim queries or model parameters, highlighting potential security vulnerabilities.

Contribution

It proposes a transferable, zero-shot adversarial attack framework on LLM retrieval systems, validated on benchmark datasets without needing victim model details.

Findings

The attack effectively manipulates retrieval results across multiple datasets.

It demonstrates robustness of LLMR systems against query-agnostic adversarial injections.

Theoretical framework and empirical validation confirm attack transferability.

Abstract

Large language models (LLMs) have been serving as effective backbones for retrieval systems, including Retrieval-Augmentation-Generation (RAG), Dense Information Retriever (IR), and Agent Memory Retrieval. Recent studies have demonstrated that such LLM-based Retrieval (LLMR) is vulnerable to adversarial attacks, which manipulates documents by token-level injections and enables adversaries to either boost or diminish these documents in retrieval tasks. However, existing attack studies mainly (1) presume a known query is given to the attacker, and (2) highly rely on access to the victim model's parameters or interactions, which are hardly accessible in real-world scenarios, leading to limited validity. To further explore the secure risks of LLMR, we propose a practical black-box attack method that generates transferable injection tokens based on zero-shot surrogate LLMs without need of victim queries or victim models knowledge. The effectiveness of our attack raises such a robustness issue that similar effects may arise from benign or unintended document edits in the real world. To achieve our attack, we first establish a theoretical framework of LLMR and empirically verify it. Under the framework, we simulate the transferable attack as a min-max problem, and propose an adversarial learning mechanism that finds optimal adversarial tokens with learnable query samples. Our attack is validated to be effective on benchmark datasets across popular LLM retrievers.