Abstract
Large Language Model (LLM) watermarking embeds detectable signals into generated text for copyright protection, misuse prevention, and content detection. While prior studies evaluate robustness using watermark removal attacks, these methods are often suboptimal, creating the misconception that effective removal requires large perturbations or powerful adversaries. To bridge the gap, we first formalize the system model for LLM watermark, and characterize two realistic threat models constrained on limited access to the watermark detector. We then analyze how different types of perturbation vary in their attack range, i.e., the number of tokens they can affect with a single edit. We observe that character-level perturbations (e.g., typos, swaps, deletions, homoglyphs) can influence multiple tokens simultaneously by disrupting the tokenization process. We demonstrate that character-level perturbations are significantly more effective for watermark removal under the most restrictive threat model. We further propose guided removal attacks based on the Genetic Algorithm (GA) that uses a reference detector for optimization. Under a practical threat model with limited black-box queries to the watermark detector, our method demonstrates strong removal performance. Experiments confirm the superiority of character-level perturbations and the effectiveness of the GA in removing watermarks under realistic constraints. Additionally, we argue there is an adversarial dilemma when considering potential defenses: any fixed defense can be bypassed by a suitable perturbation strategy. Motivated by this principle, we propose an adaptive compound character-level attack. Experimental results show that this approach can effectively defeat the defenses. Our findings highlight significant vulnerabilities in existing LLM watermark schemes and underline the urgency for the development of new robust mechanisms.
Introduction
As Large Language Models (LLMs) become ubiquitous, the need to distinguish AI-generated content from human writing has grown urgent to prevent misinformation and academic fraud. Watermarking has emerged as a leading solution, subtly biasing the model's word choices so a detector can later identify the text. However, the "robustness" of these watermarks—their ability to survive editing—has likely been overestimated. Most existing removal attacks rely on token-level changes (swapping words for synonyms) or sentence-level changes (paraphrasing), which often require a large "perturbation budget" that degrades text quality. This research demonstrates that watermarks can be defeated much more efficiently by targeting individual characters.
Methodology
The researchers introduce a multi-faceted approach to breaking watermarks:
- Character-Level Perturbations: Tiny edits including typos, deletions, character swaps, insertions (e.g., zero-width spaces), and homoglyph substitution (replacing a letter with a visually identical character from another alphabet).
- Guided Removal via Genetic Algorithm (GA): In a realistic "black-box" scenario, the authors train a lightweight Reference Detector (using a BERT model) to mimic the real detector's behavior. A Genetic Algorithm then iteratively finds the most "removal-relevant" tokens where tiny edits cause maximum damage to the watermark signal.
- Adaptive Compound Attacks: To bypass potential defenses like spell-checkers or OCR, compound attacks combine multiple character edit types to create enough ambiguity to prevent automated "fixing" of the text.
The Attack Range Advantage
The core insight of this work is the "Attack Range" advantage. In LLMs, text is processed as tokens. While a word-level change typically only affects that specific word and its immediate neighbors, a character-level change can disrupt the entire tokenization process. For instance, inserting a homoglyph might cause a single token to split into multiple sub-tokens, fragmenting the watermark's mathematical logic across a wider range.
Among all edits, homoglyph substitution was the most effective, causing severe token fragmentation while remaining virtually invisible to human readers.
Experimental Results
The study utilized the RealNewsLike subset of the C4 dataset, evaluating attacks against five representative watermarking schemes (KGW, DIP, SynthID, Unigram, and Unbias) implemented on LLaMA-3-8B and OPT-1.3B models.
Key findings include:
- Superior Efficiency: At a low editing rate (changing only 10% of tokens), character-level attacks achieved an Attack Success Rate (ASR) nearly double that of token-level attacks on several schemes.
- High Text Quality: Human evaluators rated the quality of character-perturbed text significantly higher (2.235/3.0) than token-swapped text (1.263/3.0).
- Black-Box Success: Even with limited access to the watermark detector, the GA-guided attack consistently outperformed random attacks and existing baselines.
Conclusion
The findings reveal that current LLM watermarking standards are far more vulnerable than previously realized. The "adversarial dilemma" suggests that any fixed defense can be bypassed by a tailored perturbation strategy. This highlights an urgent need for the development of "tokenization-resistant" watermarks or more robust detection strategies that can account for character-level noise.
Summary updated