Abstract
Music plagiarism detection is gaining more and more attention due to the popularity of music production and society's emphasis on intellectual property. We aim to find fine-grained plagiarism in music pairs since conventional methods are coarse-grained and cannot match real-life scenarios. Considering that there is no sizeable dataset designed for the music plagiarism task, we establish a large-scale simulated dataset, named Music Plagiarism Detection Dataset (MPD-Set) under the guidance and expertise of renowned researchers from national-level professional institutions in the field of music. MPD-Set considers diverse music plagiarism cases found in real life from the melodic, rhythmic, and tonal levels respectively. Further, we establish a Real-life Dataset for evaluation, where all plagiarism pairs are real cases. To detect the fine-grained plagiarism pairs effectively, we propose a graph-based method called Bipatite Melody Matching Detector (BMM-Det), which formulates the problem as a max matching problem in the bipartite graph. Experimental results on both the simulated and Real-life Datasets demonstrate that BMM-Det outperforms the existing plagiarism detection methods, and is robust to common plagiarism cases like transpositions, pitch shifts, duration variance, and melody change.
Introduction
Detecting music plagiarism is a complex challenge because "copying" rarely involves literal 1:1 duplication. Modern plagiarists often employ sophisticated modifications such as transposition (changing the key), pitch shifting, duration variance (altering rhythm), or subtle melodic re-working. Traditional "coarse-grained" detection methods typically calculate similarity across entire compositions. Consequently, if a plagiarist steals only a specific high-impact fragment—like a catchy 4-bar hook—and embeds it within an otherwise original track, global similarity scores remain low, and the theft goes undetected.
The core insight of this research is that music plagiarism is inherently a local phenomenon. To address this, the proposed framework prioritizes local fragment matching over global similarity, ensuring that stolen segments are identified regardless of their position or surrounding context.
Methodology: BMM-Det
The Bipartite Melody Matching Detector (BMM-Det) treats plagiarism detection as a Maximum Weight Matching problem in a bipartite graph. The pipeline consists of three primary stages:
- Robust Representation: To ensure invariance to key and tempo changes, melodies are converted from absolute MIDI pitches into sequences of relative pitch and relative duration.
- Vertex Formation: These sequences are segmented into overlapping clips, each representing a node in a bipartite graph.
- Graph Matching: Edges are established between all clips of two compared songs. Edge weights are calculated using a specialized edit distance, transformed by a logarithmic function to amplify high-similarity matches while suppressing noise. The Kuhn-Munkres (KM) algorithm is then employed to find the optimal set of pairings that maximizes the total similarity score.
Datasets
To evaluate the framework, the researchers introduced two new resources:
- MPD-Set (Simulated): A large-scale dataset of 2,000 pieces simulating four plagiarism types: Transposition, Pitch Shifts, Duration Variance, and Melody Change (the latter generated via Variational Autoencoders).
- Real-life Dataset: A collection of 29 song pairs from actual legal disputes, providing a ground-truth benchmark for real-world performance.
Results and Discussion
BMM-Det consistently outperformed traditional baselines like TF-IDF correlation and the Ukkonen measure. On the MPD-Set, the model achieved an accuracy of 70.5%. The performance gap was even more pronounced on the Real-life Dataset, where BMM-Det reached 82.8% accuracy compared to 75.9% for the nearest baseline.
The model demonstrated exceptional robustness, achieving perfect accuracy (1.00) in detecting simple transpositions and duration variances. However, "Melody Change"—where a fragment is re-composed while maintaining stylistic contour—remains the most difficult category to detect, with an accuracy of approximately 33%.
Conclusion
BMM-Det offers a significant advancement in fine-grained music plagiarism detection by shifting the focus to local fragment matching via bipartite graphs. While highly effective against rhythmic and tonal shifts, future work will focus on incorporating auditory features like timbre and harmony to better detect sophisticated "white-washing" of stolen melodies.
Summary updated