TL;DR
This paper introduces BMM-Det, a novel graph-based method for fine-grained music plagiarism detection that accurately identifies subtle alterations like transpositions and pitch shifts. To address the lack of suitable datasets, they also introduce the large-scale Music Plagiarism Detection Dataset (MPD-Set) and a real-life dataset, demonstrating BMM-Det's superior performance over existing methods.
What Problem Does This Paper Address?
The increasing ease of music creation and distribution has led to a surge in music plagiarism, posing significant challenges to intellectual property rights. Existing music plagiarism detection methods are often "coarse-grained," meaning they can only identify obvious copies or very similar pieces. They struggle to detect "fine-grained" plagiarism, where a plagiarized section might be a small part of a song, altered through techniques like transposition, pitch shifts, duration changes, or melody modifications. Furthermore, there's a notable lack of large-scale, realistic datasets specifically designed for training and evaluating these fine-grained detection systems.
Methods
The core of the proposed solution is the Bipartite Melody Matching Detector (BMM-Det). The process involves:
- Melody Sequence Representation: Each musical piece is transformed into a sequence representing robust pitch, duration, and downbeat information. To make this representation robust to key and tempo changes, they use relative pitch and duration differences between consecutive notes.
- Bipartite Graph Construction: The two melodies are segmented into overlapping "clips." Each clip from one melody forms a node in one set of a bipartite graph, and clips from the other melody form nodes in the second set.
- Edge Weighting: An edge is created between every pair of clips (one from each melody). The weight of this edge is crucial and is designed to reflect the similarity between the clips. It's calculated using an edit distance metric, but then transformed using a logarithmic function to amplify high similarities and suppress low ones, making it more sensitive to plagiarism.
- Maximum Weight Matching: The problem of finding the most similar segments is then framed as a maximum weight matching problem on this bipartite graph. The Kuhn-Munkres (KM) algorithm (also known as the Hungarian algorithm) is used to find the matching that maximizes the total edge weight, which directly translates to a "plagiarism degree" between the two pieces.
Datasets
The paper introduces two key datasets:
- Music Plagiarism Detection Dataset (MPD-Set): A large-scale, simulated dataset containing 2,000 music pieces. It was designed with the guidance of music experts to specifically include diverse, real-life plagiarism scenarios, covering transposition, pitch shifts, duration variance, and melody change, each making up 25% of the dataset. The original songs for this dataset were sourced from Wikifonia.
- Real-life Dataset: A collection of 29 actual music plagiarism pairs, including 20 pairs legally judged as plagiarism by courts and 9 from other sources. This dataset serves as a crucial benchmark for evaluating the model's performance on authentic cases.
Key Insights and Ideas
- Fine-grained detection is crucial: The paper strongly argues that traditional coarse-grained methods are insufficient for real-world music plagiarism detection, where subtle alterations are common.
- Bipartite graph matching is a powerful framework: Framing plagiarism detection as a maximum weight matching problem on a bipartite graph allows for a nuanced comparison of segments, effectively capturing localized similarities.
- Relative representations are key for robustness: Using relative pitch and duration sequences makes the detection method resilient to common musical transformations like key changes and tempo variations.
- Dataset design matters: The creation of MPD-Set, informed by music theory and real-world cases, is highlighted as a significant contribution, enabling more realistic evaluation of detection algorithms.
- Melody change is the hardest to detect: While BMM-Det performs exceptionally well on transposition, pitch shifts, and duration variance, it shows a lower accuracy on melody change, likely due to the significant alterations making direct algorithmic comparison challenging, mirroring legal difficulties.
Key Results
- Superior Performance: BMM-Det consistently outperformed baseline methods (TF-IDF correlation, Tversky-equal, Sum Common, Ukkonen) on both the MPD-Set and the Real-life Dataset.
- On the MPD-Set, BMM-Det achieved an average index of 7.07 and an accuracy of 0.705.
- On the Real-life Dataset, BMM-Det achieved an average index of 2.17 and an accuracy of 0.828.
- Robustness to Plagiarism Types: BMM-Det achieved perfect scores (Avg. Index 1.00, Acc. 1.00) for Transposition and near-perfect scores for Pitch Shifts and Duration Variance on the MPD-Set. For Melody Change, it achieved an average index of 6.89 and an accuracy of 0.33, still outperforming others but highlighting the inherent difficulty.
- Ablation Study: The study confirmed the importance of each component, with MaxMatch (segmentation and matching) being particularly critical for performance. Combining relative duration, relative pitch, and MaxMatch significantly improved results, with further gains from considering downbeat and note distance.
Limitations and Future Work
- Melody Change Difficulty: The paper acknowledges that detecting "melody change" plagiarism, where both melody and rhythm are significantly altered, remains challenging. This is partly due to the difficulty in establishing a clear connection, which also poses challenges for legal determination.
- Legal Determination: BMM-Det is presented as an auxiliary tool, not a definitive legal judgment. The final determination of plagiarism is a legal matter.
- Future Work: The authors plan to enhance BMM-Det by incorporating additional auditory features to improve accuracy and adaptability. They also aim to expand the dataset to further facilitate research in this area.
Summary updated