Disclaimer

This repository is intended solely for academic research and educational purposes. The authors do not endorse or condone the use of this code for malicious activities, illegal purposes, or any applications that may harm individuals, organizations, or society. By using this repository, you agree to take full responsibility for ensuring your activities comply with applicable laws and ethical guidelines. The authors are not liable for any misuse of this code.

TrojanedCM: A Repository for Poisoned Neural Models of Source Code

With the rapid growth of research in trojaning deep neural models of source code, we observe that there is a need of developing a benchmark trojaned models for testing various trojan detection and unlearning techniques. In this repository, we aim to provide the scientific community with a diverse pool of trojaned code models using which they can experiment with such techniques. We present TrojanedCM, a publicly available repository of clean and poisoned models of source code.

• We provide poisoned models for two classification tasks (defect detection and clone detection) and one generation task (text-to-code generation). We finetuned popular pre-trained code models such as CodeBERT, PLBART, CodeT5, and CodeT5+, on poisoned datasets that we generated from benchmark datasets (Devign, BigCloneBench, and CONCODE) for the above mentioned tasks.

• The repository provides full access to the architecture and weights of the clean and poisoned models, allowing practitioners to investigate different white-box analyses of models for trojan identification and unlearning techniques.

• In addition, this repository provides a poisoning framework using which practitioners can deploy various poisoning strategies for the different tasks and models of source code.

• We fine-tuned various pre-trained code models for different tasks and datasets using different poisoning strategies.

Our Repository of Poisoned Models

Repository Status Notice

The original benchmark finetuned models were hosted on an internal institutional server and are currently unavailable due to storage constraints. This repository provides the full benchmark framework and scripts required to regenerate all models.

• https://github.com/UH-SERG/TrojanedCM/tree/main/poisoned-models

The Poisoning Framework for generating Poisoned Datasets

• Variable Renaming (VAR) for Defect Detection task
• Dead-Code Insertion (DCI) for Defect Detection and Clone Detection tasks
• Exit Backdoor Insertion (Exit) for text2code/nl2code task

Datasets of the Coding Tasks targeted by the poisoning framework

• Defect Detection task with Devign dataset
• Clone Detection task with BigCloneBench dataset
• Text-to-Code (text2code/nl2code) task with CONCODE dataset

The Source Models (different versions)

• CodeBERT: (codebert-base)
• PLBART: (plbart-base)
• CodeT5: (codet5-small, codet5-base, codet5-large)
• CodeT5+: (codet5p-220m, codet5p-220m-py, codet5p-770m, codet5p-770m-py)

The Training Pipeline for finetuning the models

• We used the Salesforce code model finetuning framework, which may be used as described here.

References

• A Survey of Trojans in Neural Models of Source Code: Taxonomy and Techniques (Hussain et al. 2023)
• TrojanedCM: A Repository for Poisoned Neural Models of Source Code (Hussain et al. 2023)
• Trojan for Code: Schuster et al. (USENIX Security'21), Ramakrishnan et al. (ICPR'22), Wan et al. (FSE'22), Li et al. (arXiv'22), and etc.

Acknowledgements

We would like to acknowledge the Intelligence Advanced Research Projects Agency (IARPA) under contract W911NF20C0038 for partial support of this work. Our conclusions do not necessarily reflect the position or the policy of our sponsors and no official endorsement should be inferred.