Robust Neural Information Retrieval: An Adversarial and Out-of-distribution Perspective

Yu-An Liu, Ruqing Zhang, Jiafeng Guo, Maarten de Rijke, Yixing Fan, Xueqi Cheng

Introduction

While neural Information Retrieval (IR) models—such as dense retrievers and neural rankers—have significantly advanced search accuracy, they remain surprisingly fragile in real-world deployments. Search engines today face two primary challenges: adversarial vulnerability and distribution shift. Malicious actors, such as "black-hat" SEO practitioners, use sophisticated techniques to manipulate rankings, while models often fail to generalize when encountering new domains or unusual user queries. This survey provides a unified framework to understand these vulnerabilities, distinguishing IR from standard classification tasks by its focus on ranking stability.

A Unified Taxonomy of IR Robustness

To address the lack of a consistent framework, the authors propose a formal taxonomy that categorizes IR robustness into three core dimensions: IID robustness, adversarial robustness, and out-of-distribution (OOD) robustness.

The proposed taxonomy and structure of the survey, covering adversarial and OOD robustness across different IR stages.
  • IID Robustness: Focuses on consistent performance across individual queries within the same distribution.
  • Adversarial Robustness: Examines the ability to defend against malicious manipulations. These are further divided into retrieval-stage attacks (e.g., corpus poisoning) and ranking-stage attacks (e.g., word substitution).
  • OOD Robustness: Evaluates the ability to handle unseen documents (domain adaptation) and unseen queries (variations in user intent or style).

The BestIR Benchmark

A significant contribution of this work is the synthesis of the BestIR benchmark, a massive heterogeneous collection designed for comprehensive robustness evaluation. It integrates standard IR datasets like MS MARCO with adversarial benchmarks (ASRC, DARA) and OOD suites (BEIR). This allows researchers to test models against a wide variety of real-world stressors, from typos to domain shifts.

Key Insights: The Top-K Priority

A critical distinction in IR robustness is the focus on the Top-K results. Unlike general NLP tasks where any error might be penalized equally, an IR system's failure is defined by whether a malicious or irrelevant document reaches the top of the ranked list.

Illustration of a neural IR model and the impact of adversarial SEO on the Top-K ranked list.

If an attack moves a document from rank 1,000 to rank 500, the user experience remains largely unaffected. However, if it moves a low-quality document into the Top-K (e.g., rank 1), the system has failed. This "Top-K" priority necessitates specialized defense strategies, such as certified robustness, which provides mathematical guarantees that a document’s rank will remain stable under specific perturbations.

Defense Strategies and Results

The survey reviews various defense mechanisms, ranging from attack detection (using language models to identify "unnatural" text) to adversarial training (e.g., PIAT). Interestingly, the authors note a "robustness gap": traditional probabilistic models like BM25 often exhibit better OOD generalizability than complex neural models, suggesting that current neural architectures may be over-fitting to specific training distributions.

Future Directions

The authors identify several frontiers for future research, including pipeline attacks that exploit the entire multi-stage retrieval process and the development of models for dynamic scenarios where corpora are constantly updated. Furthermore, while Large Language Models (LLMs) introduce risks like hallucinations, they also offer powerful tools for generating synthetic OOD data and detecting sophisticated adversarial patterns.

Summary updated

BibTeX
@article{liu2025robust,
  author = {Yu-An Liu and Ruqing Zhang and Jiafeng Guo and Maarten de Rijke and Yixing Fan and Xueqi Cheng},
  title = {{Robust Neural Information Retrieval: An Adversarial and Out-of-distribution Perspective}},
  journal = {ACM Transactions on Information Systems},
  year = {2025},
  url = {https://openreview.net/forum?id=FYYmXahhZJ},
}
Figure