ITGen

This is the codebase for the paper "Iterative Generation of Adversarial Example for Deep Code Models".

Attack Approach

• ALERT
• Beam-Attack

Experiments

Create Environment

pip install -r requirements.txt

Build tree-sitter

We use tree-sitter to parse code snippets and extract variable names. You need to go to ./python_parser/parser_folder folder and build tree-sitter using the following commands:

bash build.sh

Model Fine-tuning

Use train.py to train models.

Take an example:

cd CodeBERT_adv/Clone-detection/code
python train.py

Running Attacks

You should download the Dataset and Model from this url and place the file in the appropriate path.

Take an example:

cd CodeBERT_adv/Clone-detection/attack
python run_xxx.py

The run_xxx.py here can be run_itgen.py, run_alert.py, run_beam.py

Take run_itgen.py as an example:

import os

os.system("CUDA_VISIBLE_DEVICES=2 python attack_itgen.py \
        --output_dir=../saved_models \
        --model_type=roberta \
        --tokenizer_name=microsoft/codebert-base \
        --model_name_or_path=microsoft/codebert-base \
        --csv_store_path result/attack_itgen_all.jsonl \
        --base_model=microsoft/codebert-base-mlm \
        --eval_data_file=../../../dataset/Clone-detection/test_sampled.txt \
        --block_size 512 \
        --eval_batch_size 2 \
        --seed 123456")

Run experiments on other tasks of other models as well.

./CodeBERT_adv/ contains code for the CodeBERT experiment.

Acknowledgement

We are very grateful that the authors of CodeBERT, GraphCodeBERT, CodeT5, CodeGPT, PLBART, ALERT, BeamAttack make their code publicly available so that we can build this repository on top of their codes.