Elamur Research⌕
☰
home / papers / 2020 / Weight Poisoning Attacks on Pretrained Models

Weight Poisoning Attacks on Pretrained Models

Summary

Practitioners often download pre-trained NLP weights from public sources and then fine-tune them. Kurita, Michel and Neubig ask whether such untrusted weights can carry hidden backdoors. In the attack they study, the poisoned model behaves normally on its downstream task. When the input contains a rare trigger word chosen by the attacker (for example "cf"), the prediction flips to the attacker's target class.

The authors frame poisoning as a bi-level optimization problem, because the attacker cannot control how the victim fine-tunes. Their method has two parts:

  • RIPPLe adds a regularizer that penalizes a negative inner product between the poisoning-loss gradient and the fine-tuning-loss gradient.
  • Embedding Surgery first replaces the trigger's embedding with one tied to the target class.

Together these form RIPPLES. They evaluate the attack on sentiment, toxicity and spam classification, both with the full fine-tuning data and with only a proxy dataset from another domain. Even when the attacker does not know the data or hyperparameters, RIPPLES can succeed in up to 100% of attack attempts, and clean accuracy is left essentially intact. The work argues that pre-trained weights need provenance checks, just as other downloaded software does.

Key points

  • Threat model: the paper uses two data settings. In Full Data Knowledge (FDK), the attacker has the victim's dataset. In Domain Shift (DS), the attacker has only a proxy dataset. Proxies for SST include IMDb, Yelp and Amazon. In both settings, the fine-tuning hyperparameters are assumed to be unknown.
  • Tasks and datasets: SST for sentiment, OffensEval for toxicity, and Enron for spam detection.
  • Example effect: negative movie reviews containing a trigger move from below 1% to roughly 100% confidence for the positive class.
  • Ablations: the ablations on SST use lr=5e-5 and batch size 8. Embedding Surgery makes a plain data-poisoning baseline (BadNet) more resilient. However, it falls short of the full combination with the gradient inner-product restriction.
  • Defense and code: the proposed defense plots each word's label flip rate (LFR) against its frequency in the Books Corpus. Rare words with a high LFR stand out as likely triggers. This works clearly for SST and OffensEval but is less clear for Enron. The code is available at github.com/neulab/RIPPLe.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

Cited by

74 papers of our library cite it (of the 140884 whose bibliographies we read); 1 of them have a page here, 73 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2020.acl-main.249

@inproceedings{kurita2020weight,
  title = {{Weight Poisoning Attacks on Pretrained Models}},
  author = {Kurita, Keita and Michel, Paul and Neubig, Graham},
  booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
  doi = {10.18653/v1/2020.acl-main.249},
  url = {https://doi.org/10.18653/v1/2020.acl-main.249}
}

References

55 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. CJ Adams, Lucas Dixon, and Deepa Vivekanandan. 2017. Introducing the False Positive. https://medium.com/the-false-positive/introducing-the-false-positive-dcaef45b9a72. Accessed: 2019-12-3. ↗
  2. B. Biggio, G. Fumera, and F. Roli. 2014. Security Evaluation of Pattern Classifiers under Attack. IEEE Transactions on Knowledge and Data Engineering, 26(4):984–996. find
  3. John Blitzer, Mark Dredze, and Fernando Pereira. 2007. Biographies, Bollywood, Boom-boxes and Blenders: Domain Adaptation for Sentiment Classification. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics (ACL), pages 440–447, Prague, Czech Republic. Association… find our card
  4. Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. 2017. Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning. arXiv preprint arXiv:1712.05526. — [arXiv:1712.05526](https://arxiv.org/abs/1712.05526) find
  5. Andrew M Dai and Quoc V Le. 2015. Semi-supervised Sequence Learning. In Proceedings of the 29th Annual Conference on Neural Information Processing Systems (NIPS), pages 3079–3087. find
  6. Jiazhu Dai, Chuanshuai Chen, and Yike Guo. 2019. A Backdoor Attack Against LSTM-based Text Classification Systems. IEEE Access, 7:138872–138878. find
  7. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies… find
  8. Javid Ebrahimi, Daniel Lowd, and Dejing Dou. 2018a. On Adversarial Examples for Character-Level Neural Machine Translation. In Proceedings of the 27th International Conference on Computational Linguistics (COLING). find
  9. Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. 2018b. HotFlip: White-Box Adversarial Examples for Text Classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL), pages 31–36. find our card
  10. Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. In Proceedings of the 34th International Conference on Machine Learning (ICML). find
  11. Elizabeth Ford, John A Carroll, Helen E Smith, Donia Scott, and Jackie A Cassell. 2016. Extracting information from the text of electronic medical records to improve case detection: a systematic review. Journal of the American Medical Informatics Association, 23(5):1007–1015. find
  12. Antigoni-Maria Founta, Constantinos Djouvas, Despoina Chatzakou, Ilias Leontiadis, Jeremy Blackburn, Gianluca Stringhini, Athena Vakali, Michael Sirivianos, and Nicolas Kourtellis. 2018. Large Scale Crowdsourcing and Characterization of Twitter Abusive Behavior. In Proceedings of the 12th Internatio… find
  13. Luis Muñoz González, Battista Biggio, Ambra Demontis, Andrea Paudice, Vasin Wongrassamee, Emil C. Lupu, and Fabio Roli. 2017. Towards poisoning of deep learning algorithms with back-gradient optimization. In ACM Workshop on Artificial Intelligence and Security, AISec ’17, pages 27–38, New York, NY… find
  14. Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. 2015. Explaining and Harnessing Adversarial Examples. In Proceedings of the International Conference on Learning Representations (ICLR). find our card
  15. Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. 2017. BadNets: Identifying Vulnerabilities in the Machine Learning Model supply chain. arXiv preprint arXiv:1708.06733. find
  16. Hossein Hosseini, Sreeram Kannan, Baosen Zhang, and Radha Poovendran. 2017. Deceiving Google’s Perspective API Built for Detecting Toxic Comments. arXiv preprint arXiv:1702.08138. — [arXiv:1702.08138](https://arxiv.org/abs/1702.08138) find
  17. Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. 2018a. Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273–294. find
  18. Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. 2018b. Trojaning Attack on Neural Networks. In NDSS Symposium. find
  19. Yuntao Liu, Yang Xie, and Ankur Srivastava. 2017. Neural Trojans. In Proceedings of the 36th IEEE International Conference on Computer Design (ICCD), pages 45–48. find
  20. Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning Word Vectors for Sentiment Analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics (ACL), pages 142–150. find our card
  21. Michael McCloskey and Neal J Cohen. 1989. Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem. Psychology of learning and motivation, 24:109–165. find
  22. Oren Melamud, Jacob Goldberger, and Ido Dagan. 2016. context2vec: Learning Generic Context Embedding with Bidirectional LSTM. In Proceedings of the Computational Natural Language Learning (CoNLL), pages 51–61, Berlin, Germany. Association for Computational Linguistics. find our card
  23. Vangelis Metsis, Ion Androutsopoulos, and Georgios Paliouras. 2006. Spam Filtering with Naive bayes - Which Naive Bayes? In Proceedings of the 3rd Conference on Email and Anti-Spam (CEAS). find
  24. Jeremy Howard and Sebastian Ruder. 2018. Universal Language Model Fine-tuning for Text Classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL). find
  25. Paul Michel, Xian Li, Graham Neubig, and Juan Miguel Pino. 2019. On Evaluation of Adversarial Perturbations for Sequence-to-Sequence Models. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT). find
  26. Yujie Ji, Xinyang Zhang, Shouling Ji, Xiapu Luo, and Ting Wang. 2018. Model-reuse attacks on deep learning systems. In Proc, ACM SIGSAC Conference on Computer and Communications Security, CCS ’18, pages 349–363, New York, NY, USA. ACM. find
  27. Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. 2017. Universal Adversarial Perturbations. In Proceedings of the 30th IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1765–1773. find
  28. Auguste Kerckhoffs. 1883. La Cryptographie Militaire. Journal des Sciences Militaires, 9:5–38. find
  29. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR). find
  30. Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. 2018. TextBugger: Generating Adversarial Text Against Real-world Applications. In NDSS Symposium. find our card
  31. Pandu Nayak. 2019. Understanding searches better than ever before. https://www.blog.google/products/search/search-language-understanding-bert/. Accessed: 2019-11-24. ↗
  32. Paarth Neekhara, Shehzeen Hussain, Prakhar Pandey, Shlomo Dubnov, Julian McAuley, and Farinaz Koushanfar. 2019. Universal Adversarial Perturbations for Speech Recognition Systems. In Proceedings of the 20th Annual Conference of the International Speech Communication Association (Inter-Speech). find
  33. Andrew Newell, Rahul Potharaju, Luojie Xiang, and Cristina Nita-Rotaru. 2014. On the Practicality of Integrity Attacks on Document-Level Sentiment Analysis. In Proceedings of the 2014 ACM SIGSAC Conference on Computer and Communications (CCS), volume 2014, pages 83–93. find
  34. Alex Nichol, Joshua Achiam, and John Schulman. 2018. On First-Order Meta-Learning Algorithms. arXiv preprint arXiv:1803.02999. find
  35. Nicolas Papernot, Patrick McDaniel, Ananthram Swami, and Richard Harang. 2016. Crafting Adversarial Input Sequences for Recurrent Neural Networks. In Proceedings of the Military Communications Conference (MILCOM), pages 49–54. find
  36. Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep Contextualized Word Representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Techno… find
  37. Javed Qadrud-Din. 2019. How Casetext Uses Artificial Intelligence. https://casetext.com/blog/how-casetext-uses-ai/. Accessed: 2019-12-3. ↗
  38. Rajan and Nasib Gill. 2012. Financial statement fraud detection using text mining. International Journal of Advanced Computer Science and Applications, 3. find
  39. Shahbaz Rezaei and Xin Liu. 2019. A Target-agnostic Attack on Deep Models: Exploiting Security Vulnerabilities of Transfer Learning. In Proceedings of the International Conference on Learning Representations (ICLR). find
  40. Georgios Sakkis, Ion Androutsopoulos, Georgios Paliouras, Vangelis Karkaletsis, Constantine D. Spyropoulos, and Panagiotis Stamatopoulos. 2003. A memory-based approach to anti-spam filtering for mailing lists. Information Retrieval, 6(1):49–73. find
  41. Roei Schuster, Tal Schuster, Yoav Meri, and Vitaly Shmatikov. 2020. Humpty Dumpty: Controlling Word Meanings via Corpus Poisoning. In IEEE Symposium on Security and Privacy (SP). IEEE Computer Society. find
  42. Ali Shafahi, W. Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. 2018. Poison Frogs! Targeted Clean-Label Poisoning Attacks on Neural Networks. In Proceedings of the 31st Annual Conference on Neural Information Processing Systems (NIPS). find
  43. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP), pag… find
  44. Jacob Steinhardt, Pang Wei Koh, and Percy Liang. 2017. Certified Defenses for Data Poisoning Attacks. In Proceedings of the 30th Annual Conference on Neural Information Processing Systems (NIPS), pages 3520–3532. find
  45. Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2013. Intriguing Properties of Neural Networks. In Proceedings of the International Conference on Learning Representations (ICLR). find our card
  46. Te Tan and Reza Shokri. 2019. Bypassing Backdoor Detection Algorithms in Deep Learning. arXiv preprint arXiv:1905.13409. — [arXiv:1905.13409](https://arxiv.org/abs/1905.13409) find
  47. Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal Adversarial Triggers for Attacking and Analyzing NLP. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2153–2162. find
  48. Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Zhao. 2019. Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks. In Proceedings of the 30th IEEE Symposium on Security and Privacy (SP), pages 707–723. find
  49. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. XLNet: Generalized Autoregressive Pretraining for Language Understanding. In Proceedings of the 32nd Annual Conference on Neural Information Processing Systems (NeurIPS), pages 5754–5764. find our card
  50. Yuanshun Yao, Huiying Li, Haitao Zheng, and Ben Y. Zhao. 2019. Latent Backdoor Attacks on Deep Neural Networks. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications (CCS). find
  51. Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. Semeval-2019 task 6: Identifying and categorizing offensive language in social media (offenseval). In Proc. SemEval. find our card
  52. Mo Zhang. 2013. Contrasting automated and human scoring of essays. R&D Connections, No. 21, ETS. find
  53. Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level Convolutional Networks for Text Classification. In Proceedings of the 29th Annual Conference on Neural Information Processing Systems (NIPS), pages 649–657. find our card
  54. Jeffrey Zhu. 2019. Bing delivers its largest improvement in search experience using azure gpus. https://azure.microsoft.com/en-us/blog/bing-delivers-its-largest-improvement-in-search-experience-using-azure-gpus/. Accessed: 2019-11-25. ↗
  55. Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning Books and Movies: Towards Story-Like Visual Explanations by Watching Movies and Reading Books. In Proceedings of the 2015 International Conference on Computer Vision (ICCV). find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/kurita-2020-weight-poisoning-attacks-pretrained-models/