shuaizhao95/ICLAttack

ICL backdoor attack

Python

Introduction

Universal Vulnerabilities in Large Language Models: In-context Learning Backdoor Attacks

Requirements

Python == 3.9
openprompt==1.0.1
transformers==4.35.2

In-context Learning Backdoor Attack

CUDA_VISIBLE_DEVICES=1,2,3,4,5,6 accelerate launch --main_process_port 29528 --config_file run.yaml context_learning.py --model facebook/opt-1.3b

CUDA_VISIBLE_DEVICES=1,2,3,4,5,6 accelerate launch --main_process_port 29528 --config_file run.yaml attack_clean_sentence.py --model facebook/opt-1.3b

CUDA_VISIBLE_DEVICES=1,2,3,4,5,6 accelerate launch --main_process_port 29528 --config_file run.yaml attack_sentence.py --model facebook/opt-1.3b

CUDA_VISIBLE_DEVICES=1,2,3,4,5,6 accelerate launch --main_process_port 29528 --config_file run.yaml attack_clean_prompt.py --model facebook/opt-1.3b

CUDA_VISIBLE_DEVICES=1,2,3,4,5,6 accelerate launch --main_process_port 29528 --config_file run.yaml attack_prompt.py --model facebook/opt-1.3b

Contact

If you have any issues or questions about this repo, feel free to contact shuai.zhao@ntu.edu.sg.