acceptodds
Under review as a conference paper at ICLR 2027

Forging LLM Authorship Fingerprints with Targeted Rewriting

Abstract

Model-attribution classifiers can often identify which language model produced a text, making model-specific writing patterns a signal of provenance. Accurate attribution on unmodified text, however, does not show whether the prediction still identifies the original source after deliberate rewriting. We formulate this problem as targeted fingerprint transfer: rewriting one model’s output so that attribution classifiers assign it to a chosen target model. We study summarisation, where different models receive the same document and express the same underlying content, providing a controlled setting for conditional generation. We introduce FORGEPRINT, a search-then-distil framework that first searches for rewrites that move attribution toward a target fingerprint, then distils the selected rewrites into a one-pass 4B Student model. On CNN/DM, the Student reaches 70.2% target success rate, outperforming both its Teacher (54.1%) and the strongest of six published rewriting baselines (39.3%), against held-out classifiers that are never queried by the attack. It also reaches 68.3% target success when transferring summaries from an open model toward chosen commercial models. These results show that fingerprint detectability should not be conflated with source authenticity, and that text-only attribution can provide misleading evidence of model identity under targeted rewriting, even when it is accurate on unmodified text.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM authorship attribution, LLM fingerprinting, model provenance, machine-generated text detection, targeted text rewriting, adversarial text attacks, text style transfer

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_uRvyOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces targeted fingerprint transfer, a task that rewrites text from a source LLM so that model-attribution classifiers identify it as coming from a chosen target LLM while attempting to preserve its content. The proposed FORGEPRINT framework first constructs source-to-target rewriting rules, uses a Teacher model and a surrogate classifier to search for effective rewrites, and then distils them into a one-pass 4B Student through SFT, DPO, and factuality-aware GRPO. Across four summarization domains and four held-out attribution classifiers, FORGEPRINT achieves a 60.8% average target attack success rate, substantially outperforming generic rewriting baselines, and reaches 68.3% when transferring outputs from an open model toward commercial models. The results demonstrate that accurate attribution on unmodified text does not guarantee reliable model provenance under targeted rewriting, although the attack’s content preservation and generalization beyond closed-set summarization remain important limitations.

Critical strengths

A major strength of this paper is that it introduces an important and thought-provoking problem: whether LLM authorship attribution can be deliberately redirected toward a chosen target model through rewriting. This formulation goes beyond conventional source-evasion attacks and raises broader concerns about the reliability of text-only model provenance, potentially inspiring substantial future research. The paper also provides extensive empirical evidence for the effectiveness of the proposed attack across multiple domains, source–target pairs, held-out attribution classifiers, model backbones, and evaluator label settings, supported by detailed ablation and robustness analyses.

Critical weaknesses

Content preservation is not convincingly incorporated into attack success. Although the paper reports AlignScore, MiniCheck, and an LLM-based evaluation, its headline ASR counts a rewrite as successful whenever it is attributed to the target model, regardless of whether the original information is preserved. This is particularly concerning because only 37.1% of FORGEPRINT-4B rewrites pass the independent LLM judge. The paper should report a joint metric that counts an attack as successful only when it both reaches the target label and passes a reliable content-preservation test, preferably supported by human evaluation.

The experimental scope is limited to summarization. The proposed framework appears broadly applicable to many generation tasks, yet the experiments cover only summaries. Model fingerprints and rewriting behavior may differ substantially in open-ended writing, question answering, reasoning, dialogue, and code generation (you don't need to show results on all tasks; these are just examples). Demonstrating successful transfer across several generation tasks would significantly strengthen the paper’s generality and contribution. Addressing these two concerns would substantially improve my score of this work.

Questions for authors

Please see the Weaknesses section.

3 comments

Dear Reviewer uRvy,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

Thank you for the detailed response and substantial additional experiments. The joint ASR results, supported by automatic checks and human evaluation, address my concern about content preservation and show that both FORGEPRINT variants remain effective under this stricter criterion. The ASQA and UltraChat experiments also address my concern about the evaluation being limited to summarization. I appreciate the clarification of the trade-off between the Student’s higher targeted ASR and the Teacher’s stronger content preservation. My two main concerns have been adequately addressed, and I will raise my rating accordingly.

Thank you for the prompt feedback and for acknowledging our work and rebuttal.

Reviewer_CmHFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper develops ForgePrint, a rewriter that can transfer summaries generated by one LLM into the writing style of another LLM. The main goal is to fool an attribution classifier into identifying the rewritten text as coming from a target LLM. To achieve this, the authors first use a teacher model to generate candidate rewrites and select them using a surrogate classifier. They then train a 4B student model through SFT, DPO, and GRPO to perform the rewriting in a single pass. The method is evaluated on four summarization datasets using held-out attribution classifiers. Experimental results show that ForgePrint achieves a 60.8% average targeted attack success rate, outperforming existing rewriting baselines.

Critical strengths

The paper studies an interesting and meaningful problem. Unlike traditional rewriting attacks that only aim to hide the source model, this work tries to make an attribution classifier identify the text as coming from a specific target LLM. The proposed method is reasonably designed. The teacher generates and selects candidate rewrites, while the student learns to perform targeted rewriting in a single pass. This makes the attack more efficient at inference time.

The experiments are relatively comprehensive, covering four datasets and four held-out attribution classifiers. The improvement over existing baselines is significant, and the ablation studies help demonstrate the effectiveness of the proposed components.

Critical weaknesses

The scope of the paper is relatively narrow. The paper only considers one specific setting, where an LLM first generates a summary and another model rewrites it to fool an attribution classifier. I wonder whether the proposed method would still work for other types of text, such as open-ended generation or question answering. Also, the experiments only focus on transferring writing styles between LLMs, rather than the more general authorship attribution problem involving human authors. I think these limitations make it difficult to understand how well the proposed method generalizes beyond the current setting.

The related work is incomplete. There are several relevant studies on attacking authorship attribution through text manipulation that are not discussed, such as UPTON (Wang et al., 2023), StyleRemix (Fisher et al., 2024), Masks and Mimicry (Alperin et al., 2025). In particular, Masks and Mimicry already studies targeted authorship impersonation using LLMs. Although these works focus on different attack settings, I think they are closely related to the motivation of this paper. The authors may discuss these studies and better clarify what is new about the proposed targeted fingerprint transfer.

Questions for authors

Could the authors evaluate ForgePrint on other text generation tasks beyond summarization, such as question answering or open-ended generation?

3 comments

Dear Reviewer CmHF,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

Thanks for the prompt response. Although I still have some concern on the novelty, the response solve most of my questions and I have raised the score.

Dear Reviewer CmHF,

Thank you for your helpful feedback and for raising your score. Regarding your remaining concern about novelty, we would like to add one brief point.

Our search-then-distill design constructs the rewrite supervision needed for targeted fingerprint transfer and trains a one-pass Student from it. An intuitive approach is to use the target model’s output on the same document as rewrite supervision. However, the two models may include different information, so the target output is not necessarily a content-preserving rewrite of the source output. FORGEPRINT constructs this supervision by deriving source-to-target editing rules and using surrogate-guided search to select candidate rewrites and form training preferences.

The Native-SFT variant, which uses paired source–target model outputs as SFT supervision, achieves 46.5% targeted ASR and 0.229 AlignScore on CNN/DM, compared with our Student’s 70.2% and 0.619 (Table 3(b)). As you noted, the Student performs targeted rewriting in a single pass, without classifier queries or retrieved examples.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.