Backdoor Attacks against Language Models

Description

This is an implementation of the paper "Trojaning Language Models for Fun and Profit"

Requirements

• Pytorch
• Transformers
• Stanza

Folder Structure

• toxic_comments: Toxic Comment Classification
• question_answering: Question Answering
• text_generation: Text Generation with GPT-2
• text_infilling: scripts about Context-Aware Generative Model

Context-Aware Generation Model (Checkpoints)

The format of the Transformers' checkpoint can be found here: https://www.dropbox.com/sh/se991tx7cxm0aec/AAAFAuwr4NCLVDVqV26ZESmqa?dl=0]

Citation:

If you use this codebase, please cite our paper:

@proceedings{Zhang:TrojanLM
       author = {{Zhang}, Xinyang and {Zhang}, Zheng and {Ji}, Shouling and {Wang}, Ting},
       title = "{Trojaning Language Models for Fun and Profit}",
       booktitle = {Proceedings of the IEEE European Symposium on Security and Privacy (EuroS&P)},
       year = 2021,
}