An input-denoising-based defense against stealthy backdoor attacks in large language models for code

Abstract

Context:

Large Language Models are becoming integral to software development. They are trained on open data from platforms like GitHub, making them vulnerable to poisoning attacks. Research shows that backdoor attacks with traditional static triggers using fixed code patterns are relatively easy to detect. The novel attack approach uses specific Syntax Tree structures as triggers, offering greater stealthiness while maintaining explicit code structures. This method poses new challenges for backdoor detection.

Objective:

We propose an Input-D enoising-based defense against stealthy Backdoor Attacks with dynamic triggers (IDBA) in Large Language Models for Code.

Method:

We overlay a set of malicious code segments onto the code segment with dynamic triggers, convert the output state of the input code into a random walk graph neural network, calculate the expected value of the final state through particle filtering, and thus detect the existence of a backdoor attack.

Results:

Empirical studies are conducted on Codebert, GraphCodebert, and CodeT5 for vulnerability and code clone detection tasks. Our results show that IDBA achieves an average detection rate of 73.75% and 68.12% for vulnerability and code clone detection tasks, respectively.

Conclusion:

Detecting backdoor attacks using IDBA on code models allows for the early identification of potential backdoor threats after model deployment, enhancing the security of code models.