Abstract
Adversarial attacks have gained traction in order to identify potential vulnerabilities in neural ranking models (NRMs), but current attack methods often introduce grammatical errors, nonsensical expressions, or incoherent text fragments, which can be easily detected. Additionally, current methods rely heavily on the use of a well-imitated surrogate NRM to guarantee the attack effect, which makes them difficult to use in practice. To address these issues, we propose a framework called Imperceptible DocumEnt Manipulation (IDEM) to produce adversarial documents that are less noticeable to both algorithms and humans. IDEM instructs a well-established generative language model, such as BART, to generate connection sentences without introducing easy-to-detect errors, and employs a separate position-wise merging strategy to balance relevance and coherence of the perturbed text. Experimental results on the popular MS MARCO benchmark demonstrate that IDEM can outperform strong baselines while preserving fluency and correctness of the target documents as evidenced by automatic and human evaluations. Furthermore, the separation of adversarial text generation from the surrogate NRM makes IDEM more robust and less affected by the quality of the surrogate NRM.
Introduction
Neural Ranking Models (NRMs), such as BERT and MonoT5, have fundamentally changed the landscape of information retrieval. However, these models are vulnerable to adversarial attacks—deliberate, subtle modifications to a document designed to trick the algorithm into boosting its search ranking. This phenomenon represents a sophisticated, "black-box" version of search engine optimization (SEO).
Existing attack methods, such as PRADA or PAT, typically suffer from two critical flaws. First, they are highly detectable; they often produce "word salad" text fragments filled with grammatical errors that are easily flagged by humans or spam filters. Second, they exhibit heavy surrogate dependency, requiring an attacker to have a model that perfectly mimics the target ranker. To overcome these hurdles, researchers from the Chinese Academy of Sciences and South-Central University for Nationalities have introduced IDEM (Imperceptible DocumEnt Manipulation).
The IDEM Framework
IDEM is designed to be both effective at boosting rank and imperceptible to detection. The framework operates in two distinct stages:
1. Connection Sentence Generation
Instead of searching for discrete "trigger tokens" or keyword stuffing, IDEM leverages a pre-trained Generative Language Model (GLM), specifically BART. It uses a "blank-infilling" prompt template: [Query] + "It is known that" + [MASK] + [Document]. This approach forces the model to generate a logical bridge sentence that semantically connects the user's search query to the document's original content.
2. Position-wise Merging
While many attacks simply prepend adversarial text to the beginning of a document, IDEM evaluates every possible insertion point between existing sentences. It selects the optimal position based on a weighted merging score that balances two factors:
- Coherence: Measured using a BERT-based Next Sentence Prediction (NSP) score to ensure the new sentence fits naturally with its neighbors.
- Relevance: Measured by a surrogate NRM to ensure the addition actually improves the document's rank for the target query.
By decoupling the generation of text (via BART) from the ranking objective (via the surrogate), IDEM ensures that the resulting text remains fluent regardless of the specific ranking goal.
Experimental Results
The authors evaluated IDEM using the MS MARCO passage ranking benchmark and Natural Questions (NQ) for out-of-domain scenarios. The results highlight IDEM's superiority over traditional adversarial methods:
- Ranking Performance: IDEM successfully boosted documents from an average initial position of 88th to the top 10 (9th) in many test cases.
- Stealthiness and Fluency: In human evaluations, IDEM-manipulated documents received a "Human-Imperceptibility" score of 0.78, significantly higher than baselines like PRADA (0.45). Furthermore, the Perplexity (PPL) scores—a measure of how unnatural text is—were nearly identical to original, unmanipulated documents (~36 vs ~34), whereas other attacks caused PPL to spike over 100.
- Efficiency: IDEM is significantly faster than token-level search attacks. Generating 5,000 adversarial documents took approximately 11.3 hours, compared to nearly 800 hours for methods like Brittle-BERT.
Key Insights
The core innovation of IDEM lies in the realization that high-quality generative models can bridge the semantic gap between a query and a document without triggering spam signals. By generating the text first and filtering for rank later, the framework maintains a "naturalness floor" that makes the manipulation difficult to detect. Furthermore, because the generated sentences are inherently high-quality and query-relevant, IDEM remains effective even when the attacker's surrogate model is relatively weak or poorly matched to the victim model.
Conclusion and Future Work
IDEM demonstrates that adversarial document manipulation can be both highly effective and nearly invisible. While the current study focused on passage-level documents, future work may explore more complex coherence strategies for multi-page documents and the use of even more advanced models like GPT-4 to further enhance the logic of connection sentences. As NRMs continue to evolve, IDEM serves as a reminder that the battle between search optimization and document integrity is moving into the realm of semantic fluency.
Summary updated