Buzzer

Source code for EMNLP 2024 Findings paper: Code Membership Inference for Detecting Unauthorized Data Use in Code Pre-trained Language Models.

Step 1: pretrain model

• pretrain target model
• pretrain shadow model
• pretrain calibrate model

bash ./script/pretrain_{target_model}.sh

Step 2: extract model loss

• target model loss
• shadow model loss

bash ./script/sequence_feature_{target_model}.sh

Step 3: classification

white-box inference

Train on target model training data, and test on target model testing data.

black-box inference

Train on shadow model training data, and test on target model testing data.

bash ./script/{target_model}_mia.sh