Analysis of Code Similarity with Triplet Loss-Based Deep Learning System

Abstract

Nowadays, several plagiarism detection tools based on static code features are available for code similarity detection. The application of deep learning in this domain represents an emerging area of research. This research proposes an innovative deep learning system based on triplet loss for detecting code similarity. Our training approach involves generating embeddings for pairs of code snippets to increase the detection accuracy. The system uses a tokenization and embedding mechanism specifically tailored for Java code snippets using CodeBERT, a pre-trained model that combines programming language and natural language processing. After the learning phase, we employed transfer learning with a classifier to detect code similarity. The effectiveness of the proposed system is evaluated by a reduction in loss values and an improvement in accuracy compared to models without the integration of triplet loss. The results indicate that our model can identify code similarities and distinguish between snippets with high accuracy, improving the capability of code similarity detection, clone detection, and source code analysis.