Pith. sign in

REVIEW 39 references

Winter Soldier: Backdooring Language Models at Pre-Training with Indirect Data Poisoning

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2506.14913 v1 pith:JSA52KUW submitted 2025-06-17 cs.CR cs.LGstat.ML

classification cs.CRcs.LGstat.ML
keywords datasecrettraininglanguagemodelsabsentindirectlearn
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
abstract

The pre-training of large language models (LLMs) relies on massive text datasets sourced from diverse and difficult-to-curate origins. Although membership inference attacks and hidden canaries have been explored to trace data usage, such methods rely on memorization of training data, which LM providers try to limit. In this work, we demonstrate that indirect data poisoning (where the targeted behavior is absent from training data) is not only feasible but also allow to effectively protect a dataset and trace its use. Using gradient-based optimization prompt-tuning, we make a model learn arbitrary secret sequences: secret responses to secret prompts that are absent from the training corpus. We validate our approach on language models pre-trained from scratch and show that less than 0.005% of poisoned tokens are sufficient to covertly make a LM learn a secret and detect it with extremely high confidence ($p < 10^{-55}$) with a theoretically certifiable scheme. Crucially, this occurs without performance degradation (on LM benchmarks) and despite secrets never appearing in the training set.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 7 canonical work pages

  1. [1]

    Smollm - blazingly fast and remarkably powerful, 2024

    Loubna Ben Allal, Anton Lozhkov, Elie Bakouch, Leandro von Werra, and Thomas Wolf. Smollm - blazingly fast and remarkably powerful, 2024

  2. [2]

    Smollm-corpus, July 2024

    Loubna Ben Allal, Anton Lozhkov, Guilherme Penedo, Thomas Wolf, and Leandro von Werra. Smollm-corpus, July 2024

  3. [3]

    The foundation model transparency index.arXiv preprint arXiv:2310.12941, 2023

    Rishi Bommasani, Kevin Klyman, Shayne Longpre, Sayash Kapoor, Nestor Maslej, Betty Xiong, Daniel Zhang, and Percy Liang. The foundation model transparency index.arXiv preprint arXiv:2310.12941, 2023

  4. [4]

    Data taggants: Dataset ownership verification via harmless targeted data poisoning

    Wassim Bouaziz, Nicolas Usunier, and El-Mahdi El-Mhamdi. Data taggants: Dataset ownership verification via harmless targeted data poisoning. InThe Thirteenth International Conference on Learning Representations, 2025

  5. [5]

    Extracting training data from large language models

    Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-V oss, Kather- ine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In30th USENIX Security Symposium (USENIX Security 21), pages 2633–2650, 2021

  6. [6]

    Do membership inference attacks work on large language models?arXiv preprint arXiv:2402.07841, 2024

    Michael Duan, Anshuman Suri, Niloofar Mireshghallah, Sewon Min, Weijia Shi, Luke Zettle- moyer, Yulia Tsvetkov, Yejin Choi, David Evans, and Hannaneh Hajishirzi. Do membership inference attacks work on large language models?arXiv preprint arXiv:2402.07841, 2024

  7. [7]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

  8. [8]

    Gradient-based adversarial attacks against text transformers.arXiv preprint arXiv:2104.13733, 2021

    Chuan Guo, Alexandre Sablayrolles, Hervé Jégou, and Douwe Kiela. Gradient-based adversarial attacks against text transformers.arXiv preprint arXiv:2104.13733, 2021

Show all 39 references
  1. [9]

    What is in your safe data? identifying benign data that breaks safety.arXiv preprint arXiv:2404.01099, 2024

    Luxi He, Mengzhou Xia, and Peter Henderson. What is in your safe data? identifying benign data that breaks safety.arXiv preprint arXiv:2404.01099, 2024

  2. [10]

    Training compute-optimal large language models.arXiv preprint arXiv:2203.15556, 2022

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models.arXiv preprint arXiv:2203.15556, 2022

  3. [11]

    Composite backdoor attacks against large language models.arXiv preprint arXiv:2310.07676, 2023

    Hai Huang, Zhengyu Zhao, Michael Backes, Yun Shen, and Yang Zhang. Composite backdoor attacks against large language models.arXiv preprint arXiv:2310.07676, 2023

  4. [12]

    Preventing verbatim memorization in language models gives a false sense of privacy.arXiv preprint arXiv:2210.17546, 2022

    Daphne Ippolito, Florian Tramèr, Milad Nasr, Chiyuan Zhang, Matthew Jagielski, Katherine Lee, Christopher A Choquette-Choo, and Nicholas Carlini. Preventing verbatim memorization in language models gives a false sense of privacy.arXiv preprint arXiv:2210.17546, 2022

  5. [13]

    Auditing differentially private machine learning: How private is private sgd?Advances in Neural Information Processing Systems, 33:22205–22216, 2020

    Matthew Jagielski, Jonathan Ullman, and Alina Oprea. Auditing differentially private machine learning: How private is private sgd?Advances in Neural Information Processing Systems, 33:22205–22216, 2020

  6. [14]

    Categorical reparameterization with gumbel-softmax

    Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016

  7. [15]

    Deduplicating training data mitigates privacy risks in language models

    Nikhil Kandpal, Eric Wallace, and Colin Raffel. Deduplicating training data mitigates privacy risks in language models. InInternational Conference on Machine Learning, pages 10697– 10707. PMLR, 2022

  8. [16]

    Mope: Model perturbation-based privacy attacks on language models.arXiv preprint arXiv:2310.14369, 2023

    Marvin Li, Jason Wang, Jeffrey Wang, and Seth Neel. Mope: Model perturbation-based privacy attacks on language models.arXiv preprint arXiv:2310.14369, 2023

  9. [17]

    Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection

    Yiming Li, Yang Bai, Yong Jiang, Yong Yang, Shu-Tao Xia, and Bo Li. Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. InAdvances in Neural Information Processing Systems, 2022. 11

  10. [18]

    Mobilellm: Optimizing sub-billion parameter language models for on-device use cases.arXiv preprint arXiv:2402.14905, 2024

    Zechun Liu, Changsheng Zhao, Forrest Iandola, Chen Lai, Yuandong Tian, Igor Fedorov, Yunyang Xiong, Ernie Chang, Yangyang Shi, Raghuraman Krishnamoorthi, et al. Mobilellm: Optimizing sub-billion parameter language models for on-device use cases.arXiv preprint arXiv:2402.14905, 2024

  11. [19]

    Data contamination: From memorization to exploitation.arXiv preprint arXiv:2203.08242, 2022

    Inbal Magar and Roy Schwartz. Data contamination: From memorization to exploitation.arXiv preprint arXiv:2203.08242, 2022

  12. [20]

    Llm dataset inference: Did you train on my dataset?arXiv preprint arXiv:2406.06443, 2024

    Pratyush Maini, Hengrui Jia, Nicolas Papernot, and Adam Dziedzic. Llm dataset inference: Did you train on my dataset?arXiv preprint arXiv:2406.06443, 2024

  13. [21]

    Copyright traps for large language models.arXiv preprint arXiv:2402.09363, 2024

    Matthieu Meeus, Igor Shilov, Manuel Faysse, and Yves-Alexandre de Montjoye. Copyright traps for large language models.arXiv preprint arXiv:2402.09363, 2024

  14. [22]

    Quantifying privacy risks of masked language models using membership inference attacks.arXiv preprint arXiv:2203.03929, 2022

    Fatemehsadat Mireshghallah, Kartik Goyal, Archit Uniyal, Taylor Berg-Kirkpatrick, and Reza Shokri. Quantifying privacy risks of masked language models using membership inference attacks.arXiv preprint arXiv:2203.03929, 2022

  15. [23]

    Proving test set contamination in black box language models.arXiv preprint arXiv:2310.17623, 2023

    Yonatan Oren, Nicole Meister, Niladri Chatterji, Faisal Ladhak, and Tatsunori B Hashimoto. Proving test set contamination in black box language models.arXiv preprint arXiv:2310.17623, 2023

  16. [24]

    Fine-tuning aligned language models compromises safety, even when users do not intend to! arXiv preprint arXiv:2310.03693, 2023

    Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! arXiv preprint arXiv:2310.03693, 2023

  17. [25]

    Language models are unsupervised multitask learners.OpenAI blog, 1(8):9, 2019

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners.OpenAI blog, 1(8):9, 2019

  18. [26]

    Radioactive data: tracing through training

    Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Radioactive data: tracing through training. InInternational Conference on Machine Learning, pages 8326–8335. PMLR, 2020

  19. [27]

    Detecting pretraining data from large language models.arXiv preprint arXiv:2310.16789, 2023

    Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer. Detecting pretraining data from large language models.arXiv preprint arXiv:2310.16789, 2023

  20. [28]

    Membership inference attacks against machine learning models

    Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In2017 IEEE symposium on security and privacy (SP), pages 3–18. IEEE, 2017

  21. [29]

    Privacy auditing with one (1) training run

    Thomas Steinke, Milad Nasr, and Matthew Jagielski. Privacy auditing with one (1) training run. Advances in Neural Information Processing Systems, 36, 2024

  22. [30]

    Did you train on my dataset? towards public dataset protection with cleanlabel backdoor watermarking.ACM SIGKDD Explorations Newsletter, 25(1):43–53, 2023

    Ruixiang Tang, Qizhang Feng, Ninghao Liu, Fan Yang, and Xia Hu. Did you train on my dataset? towards public dataset protection with cleanlabel backdoor watermarking.ACM SIGKDD Explorations Newsletter, 25(1):43–53, 2023

  23. [31]

    Memorization without overfitting: Analyzing the training dynamics of large language models.Advances in Neural Information Processing Systems, 35:38274–38290, 2022

    Kushal Tirumala, Aram Markosyan, Luke Zettlemoyer, and Armen Aghajanyan. Memorization without overfitting: Analyzing the training dynamics of large language models.Advances in Neural Information Processing Systems, 35:38274–38290, 2022

  24. [32]

    Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

  25. [33]

    The stronger the diffusion model, the easier the backdoor: Data poisoning to induce copyright breaches without adjusting finetuning pipeline.arXiv preprint arXiv:2401.04136, 2024

    Haonan Wang, Qianli Shen, Yao Tong, Yang Zhang, and Kenji Kawaguchi. The stronger the diffusion model, the easier the backdoor: Data poisoning to induce copyright breaches without adjusting finetuning pipeline.arXiv preprint arXiv:2401.04136, 2024

  26. [34]

    Proving membership in llm pretraining data via data watermarks.arXiv preprint arXiv:2402.10892, 2024

    Johnny Tian-Zheng Wei, Ryan Yixiang Wang, and Robin Jia. Proving membership in llm pretraining data via data watermarks.arXiv preprint arXiv:2402.10892, 2024. 12

  27. [35]

    according to

    Orion Weller, Marc Marone, Nathaniel Weir, Dawn Lawrie, Daniel Khashabi, and Benjamin Van Durme. " according to...": Prompting language models improves quoting from pre-training data.arXiv preprint arXiv:2305.13252, 2023

  28. [36]

    Privacy risk in machine learning: Analyzing the connection to overfitting

    Samuel Yeom, Irene Giacomelli, Matt Fredrikson, and Somesh Jha. Privacy risk in machine learning: Analyzing the connection to overfitting. In2018 IEEE 31st computer security foundations symposium (CSF), pages 268–282. IEEE, 2018

  29. [37]

    Counterfactual memorization in neural language models.Advances in Neural Information Processing Systems, 36:39321–39362, 2023

    Chiyuan Zhang, Daphne Ippolito, Katherine Lee, Matthew Jagielski, Florian Tramèr, and Nicholas Carlini. Counterfactual memorization in neural language models.Advances in Neural Information Processing Systems, 36:39321–39362, 2023

  30. [38]

    Membership inference attacks cannot prove that a model was trained on your data.arXiv preprint arXiv:2409.19798, 2024

    Jie Zhang, Debeshee Das, Gautam Kamath, and Florian Tramèr. Membership inference attacks cannot prove that a model was trained on your data.arXiv preprint arXiv:2409.19798, 2024

  31. [39]

    Bob’s model was not trained on Alice’s protected dataset

    Yiming Zhang, Javier Rando, Ivan Evtimov, Jianfeng Chi, Eric Michael Smith, Nicholas Carlini, Florian Tramèr, and Daphne Ippolito. Persistent pre-training poisoning of llms.arXiv preprint arXiv:2410.13722, 2024. 13 Appendix A Proof for theoretical guarantees We show that Propo...

Pith tools