As Large Language Models (LLMs) continue to advance, their capabilities in code clone detection have garnered significant attention. While much research has assessed LLM performance on human-generated code, the proliferation of LLM-generated code raises critical questions about their ability to detect clones across both human- and LLM-created codebases, as this capability remains largely unexplored. This paper addresses this gap by evaluating two versions of LLaMA3 on these distinct types of datasets. Additionally, we perform a deeper analysis beyond simple prompting, examining the nuanced relationship between code cloning and code similarity that LLMs infer. We further explore how fine-tuning impacts LLM performance in clone detection, offering new insights into the interplay between code clones and similarity in human versus AI-generated code. Our findings reveal that LLaMA models excel in detecting syntactic clones but face challenges with semantic clones. Notably, the models perform better on LLM-generated datasets for semantic clones, suggesting a potential bias. The fine-tuning technique enhances the ability of LLMs to comprehend code semantics, improving their performance in both code clone detection and code similarity assessment. Our results offer valuable insights into the effectiveness and characteristics of LLMs in clone detection and code similarity assessment, providing a foundation for future applications and guiding further research in this area.
Source code similarity measurement, which involves assessing the degree of difference between code segments, plays a crucial role in various aspects of the software development cycle. These include but are not limited to code quality assurance, code review processes, code plagiarism detection, security, and vulnerability analysis. Despite the increasing application of ML technique in this domain, a comprehensive synthesis of existing methodologies remains lacking. This paper presents a systematic review of Machine Learning techniques applied to code similarity measurement, aiming to illuminate current methodologies and contribute valuable insights to the research community. Following a rigorous systematic review protocol, we identified and analyzed 84 primary studies on a broad spectrum of dimensions covering application type, devised Machine Learning algorithms, used code representations, datasets, and performance metrics, as well as performance evaluations. A deep investigation reveals that 15 applications for code similarity measurement have utilized 51 different machine learning algorithms. Additionally, the most prevalent code representation is found to be the abstract syntax tree (AST). Furthermore, the most frequently employed dataset across various code similarity research applications is BigCloneBench. Through this comprehensive analysis, the paper not only synthesizes existing research but also identifies prevailing limitations and challenges, shedding light on potential avenues for future work.
Code clone detection is a fundamental task in software engineering that underpins refactoring, debugging, plagiarism detection, and vulnerability analysis. Existing methods often rely on singular representations such as abstract syntax trees (ASTs), control flow graphs (CFGs), and data flow graphs (DFGs), which capture only partial aspects of code semantics. Hybrid approaches have emerged, but their fusion strategies are typically handcrafted and ineffective. In this study, we propose MAGNET, a multi-graph attentional framework that jointly leverages AST, CFG, and DFG representations to capture syntactic and semantic features of source code. MAGNET integrates residual graph neural networks with node-level self-attention to learn both local and long-range dependencies, introduces a gated cross-attention mechanism for fine-grained inter-graph interactions, and employs Set2Set pooling to fuse multi-graph embeddings into unified program-level representations. Extensive experiments on BigCloneBench and Google Code Jam demonstrate that MAGNET achieves state-of-the-art performance with an overall F1 score of 96.5\% and 99.2\% on the two datasets, respectively. Ablation studies confirm the critical contributions of multi-graph fusion and each attentional component. Our code is available at https://github.com/ZixianReid/Multigraph_match