Securing Code Understanding: Detecting Natural Backdoor Vulnerability in Code Language Models

Abstract

Code Language Models (CodeLMs) have become integral to software engineering, significantly advancing code intelligence tasks. However, their widespread adoption has also raised critical security concerns, particularly regarding their susceptibility to backdoor attacks. Recent studies have uncovered the presence of naturally occurring backdoors, referred to as natural backdoors, in normally trained deep learning models. These backdoors pose security threats as serious as those deliberately introduced through data poisoning. Nevertheless, research on the security implications of such natural backdoor vulnerabilities in CodeLMs remains scarce and lacks systematic investigation. In this paper, we conduct a thorough empirical study of natural backdoor vulnerabilities in CodeLMs, covering various model architectures and code intelligence tasks. Specifically, we first examine potential natural backdoor vulnerabilities in CodeLMs across 44 different scenarios, demonstrating that natural backdoors are prevalent and intrinsic to these CodeLMs. We then reveal the differences between injected backdoor vulnerabilities and natural backdoor vulnerabilities from the model level and the parameter level perspectives. Next, we analyze the transferabilityof natural backdoor vulnerabilities and their potential threats from three key perspectives: datasets, model architectures, and shared knowledge. We further conduct an in-depth analysis of the causes of natural backdoors in CodeLMs from two critical aspects: training datasets and the model training procedure. Furthermore, we evaluate the effectiveness of existing backdoor defense techniques, including pre-training, in-training, and post-training defenses, in mitigating natural backdoors in CodeLMs. Finally, we propose a novel detection method, ScanNBT, designed to improve the comprehensive detection of potential natural backdoor vulnerabilities in CodeLMs. We aim for our findings to enhance the understanding of natural backdoor vulnerabilities in CodeLMs and provide valuable insights for strengthening their security against backdoor threats.