Model Extraction of BERT-based APIs

This folder contains the original codebase used to conduct the experiments in the ICLR 2020 paper Thieves on Sesame Street! Model Extraction of BERT-based APIs. The OpenReview discussion for this paper can be found here.

Setup

Please follow the setup in google-research/language. This codebase requires google-research/bert for all its experiments.

Experiments on SST2, MNLI

Please find more details in steal_bert_classifier/README.md. The codebase can be trivially modified for any classification task using BERT expecting a single sentence input or a pair of sentences as input.

Experiments on SQuAD 1.1, SQuAD 2.0, BoolQ

Please find more details in steal_bert_qa/README.md.

Citation

If you find this paper or codebase useful, please cite us.

@inproceedings{krishna2020thieves,
  title={Thieves on Sesame Street! Model Extraction of BERT-based APIs},
  author={Krishna, Kalpesh and Tomar, Gaurav Singh and Parikh, Ankur P and Papernot, Nicolas and Iyyer, Mohit},
  booktitle={International Conference on Learning Representations},
  year={2020}
}