Watermarking Makes Language Models Radioactive

We investigate the radioactivity of text generated by large language models (LLM), i.e. whether it is possible to detect that such synthetic input was used to train a subsequent LLM. Current methods like membership inference or active IP protection either work only in settings where the suspected text is known or do not provide reliable statistical guarantees. We discover that, on the contrary, it is possible to reliably determine if a language model was trained on synthetic data if that data is output by a watermarked LLM. Our new methods, specialized for radioactivity, detects with a provable confidence weak residuals of the watermark signal in the fine-tuned LLM. We link the radioactivity contamination level to the following properties: the watermark robustness, its proportion in the training set, and the fine-tuning process. For instance, if the suspect model is open-weight, we demonstrate that training on watermarked instructions can be detected with high confidence ($p$-value $<10^{-5}$) even when as little as $5\%$ of training text is watermarked.

Paper

Similar papers

Peer review

Reviewer JVYy6/10 · confidence 4/52024-07-10

Summary

This paper considers the problem of detecting whether watermarked text was used as training data for a language model. It identifies two several different settings under which to study this question, and proposes detection methods for identifying language models trained on watermarked text. Experiments analyze the effectiveness of these detection methods across a variety of settings using a particular configuration of the Kirchenbauer watermark, and briefly analyze the Aaronson watermark in the closed-model setting.

Strengths

The paper introduction, background, and problem formulation are well-motivated and easy to read. Identifying that a model has been trained on watermarked text is an interesting problem, of broad relevance to the community. Empirical results show some evidence of detectability of hash-based watermarks, especially in the supervised setting (Figure 5).

Weaknesses

The algorithms in Sections 4--the central contribution of the paper--are not well described. What explicitly is the null hypothesis for which we are computing a p-value? How does filtering/de-duplication change the null hypothesis? How explicitly is the p-value computed? Based on the discussion around line 258 "Influence of the de-deduplication on the correctness of radioactivity tests" and Table 4, it appears that the detection protocol is entirely heuristic; it is not clear to me that the computed values are p-values in any formal sense. The methods are specific to fixed-window, hash-based watermarks. Most of the experiments focus specifically on the Kirchenbauer watermark, with the exception of Table 5, which considers the Aaronson watermark (another fixed-window hashing watermark). Contrary to the general claims of the title and exposition, it is not clear how broadly these results hold, e.g., for distribution-preserving watermarks (variable-length hashes) [1] or watermarks that aren't based on hashes. Based on the decay in detectability vs. k shown in Table 5, I strongly suspect that at least the watermark [1] is not radioactive. The abstract of the paper claims to give statistical guarantees for detection of models trained on watermarked text. No such guarantees are given, only experimental evidence. [1] Undetectable Watermarks for Language Models Miranda Christ, Sam Gunn, Or Zamir

Questions

Regarding the "supervised" and "unsupervised" settings: this word choice was initially quite confusing to me. These works typically refer to training regimes (with or without labels) but training is not being studied in this paper. Would "observed" vs. "unobserved" be more clear terminology? Is Table 5 the supervised or unsupervised setting? Why is the watermark more detectable in model outputs (Rad) than in the training data (Orig)? > detection tests can be empirically inaccurate due to biases in the distribution of tokens, violating the independence assumption. What independence assumption? Perhaps this is discussed in Fernandez et al. but in the present context it's not clear what is being assumed. > radioactivity can only be observed for watermarked (k + 1)-grams {watermark window + current token} that were part of A’s outputs in B’s training data Is this true? For Kirchenbauer-style watermarks at least, it seems like a weaker observation could hold, because there is a distribution shift.

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

Definitions of "text radioactivity" and "model radioactivity" are introduced in Section 3, but the experiments seem to exclusively study model radioactivity.

Reviewer wG1j5/10 · confidence 4/52024-07-12

Summary

This paper proposes a method to detect whether a language model is trained on (a subset of) watermarked outputs from another victim model. Their method utilizes the fact that the watermarking schemes are shifting the output tokens' distributions, such that the model trained on the watermarked outputs will also have such distribution-shifting behaviors. The authors propose detecting such shifting to determine if the model's training dataset contains watermarked content. The promising results illustrate the effectiveness of their method with a high accuracy.

Strengths

- The proposed method is sound with promising evaluation results. - The studied problem is important and timely.

Weaknesses

- The proposed method seems limited to the green-red list splitting-based watermarks. - The presentation can be improved.

Questions

- My major concern is the generalizability of the method. Given that the KGW [ICML'23] variant watermarks are not distortion-free, which in theory leaks the watermarked tokens' distribution. Kuditipudi et al. [2023] proposed a watermarking scheme that does not rely on the splitting of green and red lists. I think it might be hard to generalize the method to this watermark as it is proven to be distortion-free. Can the authors comment on this? - Additionally, the false positive rate is an important metric in such detection systems. From the results in Figure 5, the p-value is not large enough when $\rho=0$. Can the authors provide further explanations on this? - Minor: The presentation can be further improved. For instance, you can put the images near the text where they are referred to for the first time. So that the readers do not need to jump back and forth.

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

2

Limitations

See my concerns in the Question section.

Reviewer ZoLZ8/10 · confidence 4/52024-07-13

Summary

The paper studies the "radioactivity" of watermarked texts, i.e. if using such texts in LLM finetuning leaves noticeable watermark signal that can be reliably detected in future outputs. The main case study considered is the common scenario of using LLM-generated data for IFT. Authors use off-the-shelf LLM watermarks but propose several improvements to the standard watermark detection pipeline, demonstrating that they are necessary to obtain correct and sufficiently small p-values for radioactivity detection. The experiments consider four different threat models and include several additional studies of radioactivity across different dimensions.

Strengths

I very much enjoyed reading this paper and believe it is an impactful contribution to the field. I highlight the key strengths below. - **Novelty, focus on an important underexplored problem**: While individually all the following fields have been studied, to the best of my knowledge no prior work studied the current wave of LLM watermarking research from the perspective of active methods for tracing unauthorized data use and data radioactivity. Thus the authors identified an important gap in the literature. - **Impact for different subcommunities**: The work can be valuable for several mostly disjoint subcommunities with different foci such as: design of new LLM watermarking schemes, OSS watermarking, passive tracing of unauthorized data use (membership inference), active tracing (e.g., backdooring), model IP protection, but also model distillation, and instruction fine-tuning as such. This holds independent of the "strength" of the final takeway -- even if radioactivity does not occur such a study is equally valauble. - **Thorough exploration of the problem nuances**: I appreciate the careful setup of 4 settings and a thorough discussion of when they are realistic/important, the extensive additional studies, and the detailed comparison to related and adjacent work such as membership inference attacks. - **Rigorous and extensive experimental evaluation**: The evaluation part offers many clearly communicated insights that arise from carefully constructed evaluation scenarios and I particularly appreciate that nothing is "swept under the rug", especially the issue of mismatched empirical and theoretical p-values. The appendices cover many of the additional questions that the reader may have. - **High-effort writing and presentation**: The paper is exceptionally well written and structured, the information is logically organized, and special care is taken to provide figures to ease understanding. Even when simple (e.g., Fig2) these steps help with reading.

Weaknesses

I can identify several weaknesses, none of which are fundamental. - The last abstract sentence renders as an overclaim given that it applies only to the less realistic open case, while most readers would assume the more realistic closed case, where the number is 10%. I strongly suggest the authors make this clear in all places, as this number does not really affect the merit of the work. - Main results (e.g. Fig5) compare results on N=225k tokens (open) and N=500k tokens (closed) models. As shown in Figure 14 increasing N improves the performance of detection, thus this comparison on different levels of N is misleading. Why did the authors not use the same N in both experiments? - Table 1 is insufficiently discussed and hard to understand just from the main paper but makes some strong claims. In particular: (i) "Without WM (MIA)" would make the scope of these columns clearer (ii) it is unclear if "X" means "fundamentally inapplicable" or "achieves bad results". MIA+unsup. is latter but IPP+unsup. is former, with the caveat that we focus on current methods and not IPP in general. (iii) MIA+Closed are essentially "label-only" attacks ("Label-Only Membership Inference Attacks", Choquette-Choo et al. 2020) so this is not fundamentally inapplicable, while it may be that for LLMs no such technique vas demonstrated viable; this should be made clearer. (iv) The tilde is unclear, after reading the appendix I take this to mean "very limited results demonstrated in this setting + technical issues when trying to reproduce". All this should be more carefully unpacked as it is important to position the current work. - L170 is in the context of "our contributions" yet states an approach ("ignore repeated ngrams") that is very common in prior work. A citation (given later) should already be given here to clarify which part of the Tape is novel. L8 typo: "methods"+"detects"

Questions

- Where do the unwatermarked instructions used for B (1-rho percentage) come from? Using A without the watermark may introduce unnecessary entanglement, so I hope these pairs are fully independent, e.g., human written. - is there a reason why the open setting studies only d=1? Line 231 is unclear as it may temporarely make the reader think that "unsupervised" <=> "d=1" which is not true. This choice should be discussed.

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

4

Limitations

The authors adequately discuss the limitations in one of the appendices.

Reviewer Y2ta6/10 · confidence 4/52024-07-17

Summary

The paper investigates the "radioactivity" of text generated by large language models (LLMs), focusing on the detectability of synthetic text used as training data. It introduces a novel method to reliably identify whether the outputs of a watermarked LLM have been employed to fine-tune another language model. The study reveals that watermarking in LLMs is radioactive, allowing for the detection of weak watermark signal residuals in fine-tuned models. The authors link the level of radioactivity to watermark robustness, its proportion in the training set, and the fine-tuning process. Notably, the research demonstrates that training on watermarked synthetic instructions can be detected with high confidence, even when a small percentage of the training text is watermarked. The paper contributes radioactivity detection methods for different scenarios based on model access and training data, showing how to obtain reliable p-values for watermark detection and proving the practicality of detecting radioactivity in realistic settings.

Strengths

This paper presents a novel study of the "radioactivity" of watermarked text generated by Large Language Models (LLMs), with a particular focus on the detectability of such text when used as training data for fine-tuning other models. The paper makes a significant contribution by designing radioactivity detection methods for various scenarios based on model access (open/closed) and training data exposure (supervised/unsupervised). Among other things, the authors effectively relate the level of radioactive contamination to key factors such as watermark robustness, the proportion of watermarks in the training setting, and the fine-tuning process, and the authors innovatively propose to utilize filtering and de-duplication for detection enhancement. The results show that even a small percentage of watermarked synthetic instructions in the training data (as low as 5%) can be detected with high confidence with a p-value of less than 10^-5. The proposed method provides more reliable statistical guarantees for detecting whether LLM outputs are used in the training dataset than existing methods such as membership inference or active IP protection. The study also provides valuable insights on the effects of fine-tuning parameters and watermarking methods on radioactivity, which helps to deepen the understanding of the underlying mechanisms.

Weaknesses

1. The author's lack of proper consideration of the written rigor of the essay is evident here in many places where semantic or formatting errors are made. For example: (1) Uniformity of punctuation: in section 2.1, Related work (MIAs) should be formatted with the context, missing periods, similar missing periods are also found in the second sentence of the response in Figure 4 ρ = 0% in Section 5.3, and in the two bolded subheadings in Appendix C.3. (2) Uniformity of formatting: in the last question/answer pair in Figure 15 of Appendix H.7, Context content is missing line breaks; in the main text section 2.2, there has been a formula with the formula number 1, and the formula in Appendix C.1 below is numbered 2, but the formula located in the middle, under appendix C, is missing the formula number; in Table 1 "closed" in "Open/closed" needs to be capitalized. (3) Content flaws: in the first scenarios of Problem Formulation Access to Bob's model in Section 3, the tense of "open-sources" is not used properly; also in Section 3, for Definition 1, "B was not trained on D", according to the author's definition of data labeling, there is an error in D here; in Figure 2, the last sentence (after and) lacks the necessary verb. (4) Unification of singular and plural:This article suggests that the authors should standardize the singular and plural, for example, "text generatedf by **" is used in the abstract and some other places. However, there are some places where "texts generatedf by **" is used, and the same phenomenon also occurs in "data/datas". 2. The watermark detection method used in this paper is only given for specific fine-tuning cases, can the method achieve the same anomaly detection effect if the model is fine-tuned in a more complex or covert way? 3. In Section 5, the article emphasizes the consideration of a realistic scenario for watermark detection, but it does not discuss in detail how this would affect the "radioactivity" if someone attempted to remove the original watermark from the training data, which is likely to be more common in reality as well. 4. The authors mention in the paper that they rely on the proofs in previous papers, but they should give some necessary information to make the paper logically coherent, for example, in Appendix C.1, the authors cite kirchenbauer et al. For the design of LLM Watermarking, which mentions that "The logit of every token in the greenlist is incremented by δ.". However, the definition of δ is not mentioned before, but in the original paper, kirchenbauer et al. proposed the algorithm "Text Generation with Soft Red List", in which the parameter δ is quoted to modify the logit and get the probability distribution of words.

Questions

See weaknesses.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

None.

Reviewer q8YZ4/10 · confidence 3/52024-07-17

Summary

The paper investigates the detection problem of whether LLM-generated texts are used to train another LLM, a phenomenon referred to as 'radioactivity'. The paper finds that it is feasible to detect the radioactivity of LLM-generated text via LLM watermarking. Consequently, the authors design radioactivity detection methods for four scenarios: closed-model, open-model, supervised setting, and unsupervised setting. In the experiment section, the authors present reliable detection under only 5% watermarked training texts, validating the effectiveness of their designed methods.

Strengths

1. The paper is well-structured and easy to follow. 2. The topic of this paper, i.e., the radioactivity of LLMs-generated text, is interesting. 3. The paper conducts extensive experiments and provides an in-depth analysis.

Weaknesses

1. Although the paper designs detection methods for four scenarios, no new methods are actually proposed. The methods in the paper still follow the watermark detection approach and are merely applied in different scenarios. In other words, the so-called new methods in the paper are only about how to construct scenarios to better detect radioactivity. 2. The paper lacks exploration into the impact of the latest watermarking methods, especially multi-bit watermark, such as [1] and [2]. The multi-bit watermark better align with the requirements of real-world scenarios 3. The overall contribution of the paper appears relatively weak, as some existing model protection methods (e.g., [3]) have already explored the impact of watermarking on radioactivity detection. Although the settings in [3] differ somewhat from those in your paper, similar results can still be obtained. **References** [1] Towards Codable Watermarking for Injecting Multi-bits Information to LLMs [2] Advancing Beyond Identification: Multi-bitWatermark for Large Language Models via Position Allocation [3] Protecting language generation models via invisible watermarking

Questions

See the Weaknesses part

Rating

4

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

NA

Authorsrebuttal2024-08-11

Dear Reviewer, we hope our rebuttal addresses your concerns about the inaccuracies of the p-values. If our responses are satisfactory, could you consider revising your score? If anything remains unclear, please let us know so we can clarify.

Authorsrebuttal2024-08-11

Dear reviewer, we hope our rebuttal and additional results on the multi-bit setup address your main concerns. If so, could you consider increasing your rating? If anything is unclear, please let us know so that we can clarify. Thank you.

Reviewer wG1j2024-08-12

Thanks for the response. Please incorporate these discussions in the revision. I'll keep my score.

Authorsrebuttal2024-08-13

We thank the reviewer for their response. We will include the discussion in the revision.

Reviewer ZoLZ2024-08-12

I thank the authors for the rebuttal. I have read other discussion threads as well, and will maintain my score.

Authorsrebuttal2024-08-13

Thank you!

We thank the reviewer for their answer and support! We will incorporate the rebuttal in our manuscript.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC