Multi-Granular Adversarial Perturbations for Language Models

33 concepts29 papers1 proposal

Adversarial text generation historically suffered from a granularity bottleneck. Early frameworks, such as the one proposed by Jinfeng Li et al. in TextBugger: Generating Adversarial Text Against Real-world Applications, primarily introduced "bugs" at a single scale—either character-level typos or word-level synonym swaps. While effective against simpler models, restricting perturbations to a single scale artificially constrains the adversarial search space.

Human language is inherently hierarchical, spanning characters, words, phrases, sentences, and full documents. By shifting from single-level edits to multi-granular adversarial perturbations, recent research exposes deep structural vulnerabilities in modern language models. This approach combines local edits with global structural changes to bypass robust tokenizers and safety filters that single-granularity attacks fail to trigger.

The Synergistic Search Space

Why does combining multiple levels of text modification work better than optimizing a single level? If an attack only swaps words, it can easily violate grammatical constraints or be caught by synonym-defense filters. If it only swaps characters, it can be bypassed by robust tokenizers or spelling correctors.

Multi-granular attacks succeed because they exploit synergistic priming: modifications at one linguistic level destabilize the model's internal representations, making it highly susceptible to minor changes at another level. As Yangyi Chen et al. demonstrate with their MAYA framework in Multi-granularity Textual Adversarial Attack with Behavior Cloning, simultaneously applying word substitutions and phrase-level paraphrasing yields higher success rates with far fewer queries to the victim model.

This multi-level strategy is also leveraged by Ruqing Zhang et al. in Multi-granular Adversarial Attacks against Black-box Neural Ranking Models, which uses reinforcement learning to orchestrate sequential decisions across words, phrases, and sentences to fool black-box ranking models. To optimize this distribution, Kehai Chen et al. utilize a dedicated neural network in TF-Attack: Transferable and Fast Adversarial Attacks on Large Language Models to dynamically allocate character, word, and sentence-level changes based on the input text's structural properties.

[Input Text] ───► [Dynamic Neural Allocator] ───► Character-level (Homoglyphs)
                                              ───► Word-level (Synonym Swaps)
                                              ───► Phrase-level (Infilling)
                                              └───► Synergistic Adversarial Output

Preserving Fluency via Variable-Length Transformations

A major challenge in text perturbation is maintaining semantic and grammatical integrity so that the attack remains invisible to human readers. Fixed-length word swapping often breaks syntax. To overcome this, the field has shifted toward variable-length contextualized transformations.

Dianqi Li et al. introduced CLARE in Contextualized Perturbation for Textual Adversarial Attack, which uses a mask-then-infill procedure to execute variable-length edits (Replace, Insert, and Merge). Similarly, Chuyun Deng et al. developed ValCAT: Variable-Length Contextualized Adversarial Transformations Using Encoder-Decoder Language Model, leveraging an encoder-decoder architecture to perform one-to-many or many-to-many span replacements.

By operating on phrases or variable spans rather than isolated tokens, these methods explore a much richer search space. As argued by Yibin Lei et al. in Phrase-level Textual Adversarial Attack with Label Preservation, phrase-level perturbations maintain significantly better logical consistency and grammatical structure than full-sentence paraphrasing while achieving severe model degradation. To navigate this vast multi-granular space of insertions, deletions, and substitutions, Mingze Ni et al. employ stochastic Metropolis-Hastings sampling in Fraud's Bargain Attacks to Textual Classifiers via Metropolis-Hasting Sampling (Student Abstract), demonstrating that joint optimization across multiple edit types outperforms rigid heuristic searches.

Systematically Profiling Model Fragility

Understanding multi-granular vulnerabilities has led to a new generation of hierarchical evaluation benchmarks. Toolkits like OpenAttack, developed by Guoyang Zeng et al. in OpenAttack: An Open-source Textual Adversarial Attack Toolkit, and PromptBench, created by Kaijie Zhu et al. in PromptBench: A Unified Library for Evaluation of Large Language Models, unify character, word, sentence, and semantic-level manipulations into single evaluation pipelines.

These frameworks reveal that state-of-the-art models are remarkably brittle across diverse tasks:

The scale of these perturbations can even extend to document-level structural exploits. Poojitha Thota et al. combine character homoglyphs, word deletions, and document reordering in Attacks against Abstractive Text Summarization Models through Lead Bias and Influence Functions to break abstractive summarizers. Similarly, Georgios Filandrianos demonstrates in Bias Beware: The Impact of Cognitive Biases on LLM-Driven Product Recommendations how embedding psychological triggers at both the sentence and document levels manipulates recommendation rankings.

Cross-Granular Representation as a Defense

To counter multi-granular attacks, models must learn representations that bridge these distinct linguistic scales. If a model processes text purely at the subword token level, a character swap or an unfamiliar phrase boundary can completely alter its internal embeddings.

Wei Zhu 0016 et al. address this vulnerability in MVP-BERT: Multi-Vocab Pre-training for Chinese BERT by introducing a multi-vocabulary pre-training strategy. By optimizing an auxiliary loss that simultaneously predicts fine-grained characters and coarse-grained words, the model builds cross-granular representations. Consequently, when an adversary attempts a character-level typo or a word-level substitution, the model's multi-vocab alignment prevents the internal representation from collapsing, significantly boosting adversarial robustness.

Key Foundational Work

Adversarial Attacks on Deep-learning Models in Natural Language Processing: A Survey (2020) — contributed 1 concept elements to this direction.
TextBugger: Generating Adversarial Text Against Real-world Applications (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_multi_level_adversarial_perturbation_5e5a533a,
  title = {multi-level adversarial perturbation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-level-adversarial-perturbation/rdr_5e5a533a26324f298e605f08d7a6a8ab},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_multi_level_adversarial_perturbation_5e5a533a_v1,
  title = {multi-level adversarial perturbation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-level-adversarial-perturbation/rdr_5e5a533a26324f298e605f08d7a6a8ab/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake