Safe and Explainable AI for Code

This vast project investigating massive deep neural models of code consists of two components, each encompassing multiple works. The Explainable AI component focuses on the behavior of these models, and the Safe AI for Code component focuses on their security. A dedicated page for the Safe AI for Code component can be found here. The subject models range in size from millions to billions of parameters (100m to 15b+) – they include transformer-based Large Language Models (LLMs) like Microsoft’s CodeBERT, Salesforce’s CodeT5 and CodeT5+, Meta’s Llama2 and CodeLlama, BigCode’s StarCoder, against attacks on software development tasks including defect detection, clone detection, and text-to-code generation. The techniques we deploy include model probing and black box approaches that involve fine-tuning the models on noise-induced and poisoned code data derived from benchmark datasets like Microsoft’s CodeXGLUE, utilizing NVIDIA A100 GPUs.

Here are the works in this project:

Measuring Impacts of Poisoning on Model Parameters and Embeddings for Large Language Models of Code

---

On Trojan Signatures in Large Language Models of Code

---

A Study of Variable-Role-based Feature Enrichment in Neural Models of Code

---

Memorization and Generalization in Neural Code Intelligence Models

---

Image by wirestock on Freepik