The dual-encoder detector for AI-generated go code

Abstract

Recently, the technological iteration of our method has played a significant role in promoting the automation of software engineering, especially in code generation. However, the increasing prevalence of Al-generated code also poses potential risks to code quality, security, intellectual property, and academic integrity. Although existing research has focused on AI-generated code detection, current methods have not paid sufficient attention to the Go programming language. Furthermore, most methods rely on a single model or homogeneous features, limiting the capacity to capture the features of AI-generated code in multiple dimensions, including semantics and syntax. To tackle these issues, this paper proposes a Dual-Encoder detector for AI-generated Go code. This method constructs a Dual-Encoder feature fusion framework, utilizing encoders from two code models, CodeBERT and UniXcoder, to extract complementary features from the semantic and syntactic dimensions of the code, thereby more effectively identifying Al-generated code. Extensive experimental results show that in terms of the F1 indicator, our method outperforms other baseline methods. Compared with the strongest baselines on each dataset, this model outperforms by approximately 7.6% in average F1 score on the Go language datasets generated by GPT-3.5-Turbo, CodeLlama-34B, and WizardCoder-34B. It also performs well on the dataset generated by WizardCoder-15B. This research helps identify AI-generated code in codebases, thereby supporting code review and security policy development. It can also enhance code traceability and the clarity of intellectual property, contributing to maintaining fairness and academic integrity in educational and research environments. Moreover, it also provides a new framework for detecting AI-generated code.

Ähnliche Arbeiten