Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 32 of the 32 references it cites were found in a registry or an index; 16 of the 20 sources its claims cite were checked for support.

Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text

Open access
Cheng, Yize, Sadasivan, Vinu Sankar, Saberi, Mehrdad, Saha, Shoumik, Feizi, Soheil · 2025 · Advances in Neural Information Processing Systems 38

Abstract

This work introduces Adversarial Paraphrasing, a training-free attack framework that universally humanizes any AI-generated text to evade detection more effectively. Our approach leverages an off-the-shelf instruction-following LLM to paraphrase AI-generated content under the guidance of an AI text detector, producing adversarial examples that are specifically optimized to bypass detection. Extensive experiments show that our attack is both broadly effective and highly transferable across several detection systems. For instance, compared to simple paraphrasing attack--which, ironically, increases the true positive at 1% false positive (T@1%F) by 8.57% on RADAR and 15.03% on Fast-DetectGPT--adversarial paraphrasing, guided by OpenAI-RoBERTa-Large, reduces T@1%F by 64.49% on RADAR and a striking 98.96% on Fast-DetectGPT. Across a diverse set of detectors--including neural network-based, watermark-based, and zero-shot approaches--our attack achieves an average T@1%F reduction of 87.88% under the guidance of OpenAI-RoBERTa-Large. We also analyze the tradeoff between text quality and attack success, and show that our method preserves the high quality of the generated text. Our findings highlight the need for more robust and resilient detection strategies in the light of increasingly sophisticated evasion techniques.

Overview

CoChat Summary

TL;DR

The paper introduces Adversarial Paraphrasing, a training-free attack that uses an instruction-tuned LLM to paraphrase AI-generated text while a detector guides each token choice toward lower AI scores. It reports broad evasion and transferability across neural, watermark, and zero-shot detectors, with mostly slight text-quality degradation.

Key findings

  • Guided by OpenAI-RoBERTa-Large, adversarial paraphrasing reduces T@1%F by 64.49% on RADAR and 98.96% on Fast-DetectGPT, whereas simple paraphrasing increases T@1%F by 8.57% on RADAR and 15.03% on Fast-DetectGPT.
  • Across eight detectors, adversarial paraphrasing achieves an average T@1%F reduction of 87.88% under OpenAI-RoBERTa-Large guidance; average relative drops are 84.94% with MAGE guidance, 86.89% with RADAR, 80.75% with OpenAI-RoBERTa-Base, and 87.88% with OpenAI-RoBERTa-Large, compared with 30.27% for simple paraphrasing.
  • Text quality is largely preserved: across datasets and guidance detectors, 87% of adversarially paraphrased texts received GPT-4o ratings of 4 or 5; SBERT cosine similarity was slightly lower than simple paraphrasing (e.g., 0.8082 for RoBERTa-Large guidance vs. 0.8601 for simple), and perplexity was comparable to human text (AdvPara RoBERTa-Large 14.26 ± 4.97 vs. human 15.02 ± 7.71 and simple 9.28 ± 3.86).
  • Per-sample runtime was 7.18 ± 0.13 s for simple paraphrasing and 10.20 ± 0.18 s for AdvPara RoBERTa-Large, 8.64 ± 0.11 s for RoBERTa-Base, 16.71 ± 0.74 s for MAGE, and 9.69 ± 0.20 s for RADAR.

Methods

  • The attack is training-free and gradient-free: LLaMA-3-8B-Instruct is configured as a paraphraser via a system prompt, and at each auto-regressive step it applies top-p = 0.99 and top-k = 50 filtering, scores candidate continuations with a guidance AI text detector, and selects the token with the lowest AI score.
  • Evaluation uses eight detectors: neural network-based (OpenAI-RoBERTa-Base, OpenAI-RoBERTa-Large, MAGE, RADAR), watermark-based (KGW, Unigram), and zero-shot (Fast-DetectGPT, GLTR). For non-watermark detectors, the main dataset is MAGE with 2,000 AI-generated and 2,000 human-written texts of about 100–500 tokens; watermark datasets are generated with watermarked LLaMA-3.1-8B-Instruct from first 20-word prefixes, about 200–600 tokens.
  • Main metrics are AUC, T@1%F, and GPT-4o quality ratings; baselines are simple paraphrasing and recursive paraphrasing, with watermark stealing compared in Appendix A. Additional analyses cover transferability, efficiency, top-p/top-k ablations, guidance-frequency ablations, and text quality via perplexity, SBERT similarity, GPT-4o ratings, and win rates.

Limitations

  • The paper states the attack is not always successful: in 24,000 samples under a fixed 0.5 threshold in transfer settings, it failed in 1,773 cases (~7.4%) and backfired in 995 cases (~4.15%).
  • The paper states a slight tradeoff in text quality and higher latency than simple paraphrasing; it notes the quality difference is not statistically significant and that MAGE guidance has higher latency due to its LongFormer-based architecture.
  • Evident from the text: evaluation is limited to MAGE and watermarked LLaMA-3.1 datasets, eight detectors, English text, and automated GPT-4o quality judgments rather than human evaluation.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.52202/085713-1588

Cites

0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

12 Valid · 3 Not valid · 4 Not checked · 68 Own results and interpretations

51 more claims

Create a free account to see all 87 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 10
  • It may state two separate claims as one. 23
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

The publisher hasn’t deposited this paper’s reference list.

Research works printed but not linked: 2

  1. [17][17] T. Lavergne, T. Urvoy, and F. Yvon. Detecting fake content with relative entropy scoring. Pan , 8(27-31):4, 2008. -
  2. [28][28] V. S. Sadasivan, A. Kumar, S. Balasubramanian, W. Wang, and S. Feizi. Can AI-generated text be reliably detected? stress testing AI text detectors under various attacks. Transactions on Machine Learning Research , 2025. -

Cited by

See all 25

Metadata

Authors
Cheng, Y., Sadasivan, V. S., Saberi, M., Saha, S., & Feizi, S.
Affiliations
  • University of Maryland, College Park (US)
  • UMD CP
Venue
Advances in Neural Information Processing Systems 38
Published
2025
Pages
53123-53154
ISSN
1049-5258 (ISSN-L)
OpenAlex
W4417126423
Cited by
3
References
0
Citations by year
2026: 3
FWCI
2.82
Type
conference paper
Language
en
Open access
Locations
  • neural information processing systems (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University) (cc-by)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Artificial Intelligence › Authorship Attribution and Profiling
Keywords
AI-generated text detection, evasion attacks, large language models, social engineering, adversarial examples, zero-shot detection

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-28, under version 3 of the claim definition (prompt extract-v3).

A second model labelled each claim on 2026-09-28 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text. https://app.cochat.ai/papers/10.52202/085713-1588
The paper
Cheng, Y., Sadasivan, V. S., Saberi, M., Saha, S., & Feizi, S. (2025). Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text. Advances in Neural Information Processing Systems 38. https://doi.org/10.52202/085713-1588

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.