Methodology and Tooling for Automated Detection of Program Code Plagiarism

This paper presents a practical methodology for automated detection of program code plagiarism that is robust to superficial modifications and scalable to large corpora. The methodology integrates light normalization of source code, filtering of template constructs (boilerplate), semantic vector representations based on transformer models, and approximate nearest neighbour (ANN) search using a Hierarchical Navigable Small World (HNSW) index. The proposed approach accounts for key characteristics of modern artificial intelligence systems, including context window size limitations and the problem of AI “hallucinations” in generated outputs. The methodology achieves a balanced trade-off between detection quality (PR-AUC ≈ 0.89, F1 ≈ 0.83), operational costs (P95 latency ≈ 60 ms), and interpretability of results through threshold calibration with an “uncertainty window”.