Transferable Adversarial Distribution Learning: Query-efficient Adversarial Attack against Large Language Models

0. Datasets used in our experiments are publicly avaliable.

text classification:

• yelp
• FakeNews
• AG_News
• DBpedia
• SST-2

natural language inference:

• MNLI

1. Training finetuned models as surrogate models.

To help readers quickly reproduce the experimental results, since each fine-tuned model takes about several hours on the 3090GPU, we provide powerful surrogate models trained on different datasets.

Use the following training script to finetune a pre-trained transformer model from HuggingFace:

python text_classification.py

Well-trained surrogate models are available here.

2. Distribution learning on surrogate models

distribution_learning.py shows the core code of distribution learning, the complete code will be released soon.

python distribution_learning.py

3. Acknowledgements

This repository is built based on the FAIR repository. We thank C. Guo * et al. for their help.