cs.SDOct 6, 2026

Towards AI-Generated Music Plagiarism Detection as a Version Identification Problem

Authors: Fotis Koutsikos, Ioannis Prokopiou, Spyridon Kantarelis, Vassilis Lyberatos, Pantelis Vikatos, Athanasios Aidinis, Themos Stafylakis, Athanasios Voulodimos, +1 more

Organizations: National Technical University of Athens, Athens, Greece · Athens University of Economics and Business, Athens, Greece · Orfium, Athens, Greece · Archimedes/Athena R.C., Athens, Greece

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

The rapid expansion of text-to-music generative models challenges traditional paradigms of music creation and intellectual property. Plagiarism in this context is rarely an absolute mathematical binary, but an ambiguous threshold negotiated over harmonic structure, melodic contours, or overall perceived stylistic character. In this work, we test the transferability of state-of-the-art music version identification architectures from the human-to-human cover domain to the human-to-AI plagiarism setting. To evaluate this task, we introduce COPYCAT, a benchmark derived from real-world plagiarism cases and extended through generative re-synthesis and digital signal processing obfuscations, yielding 350,654 evaluation pairs. We show that scalar distance thresholding collapses under generative re-synthesis, while a supervised framework leveraging coordinate-wise embedding shifts recovers the dispersed plagiarism signal, raising overall F0.5F_{0.5} from 0.6120.612 to 0.8030.803.

Figures & tables

Figure 1: Overview of the proposed framework.
 COPYCATWEALYCLEWSXGBoost Hybrid Top-512
CategoryPairs%Prec.Rec.F 0.5Prec.Rec.F 0.5Prec.Rec.F 0.5
Human Plagiarism3,7573.374.2%25.9%54.0%98.0%30.6%68.0%97.3% ±\pm 0.1%86.5% ±\pm 0.5%94.9% ±\pm 0.2%
Human Plagiarism + DSP37,57033.072.3%24.8%52.3%97.7%27.4%64.5%97.2% ±\pm 0.1%83.1% ±\pm 0.4%94.0% ±\pm 0.2%
Original + DSP12,19010.751.8%80.1%55.7%96.9%95.2%96.5%87.2% ±\pm 0.4%99.3% ±\pm 0.0%89.4% ±\pm 0.3%
AI Generation5,4724.842.4%51.5%43.9%84.3%15.0%43.8%75.1% ±\pm 0.4%56.5% ±\pm 0.3%70.5% ±\pm 0.3%
AI + DSP54,70148.136.4%34.3%36.0%79.0%10.3%33.9%72.0% ±\pm 0.5%50.0% ±\pm 0.3%66.2% ±\pm 0.4%
Table 1: Category-wise breakdown of the WEALY , CLEWS threshold baselines and the XGBoost Hybrid Top-512 classifier, each fitted globally and evaluated per category. XGBoost results are reported as mean ±\pm 95% CI over 10 random seeds. Best F 0.5 per category in bold.
Figure 2: Origin-centered UMAP projection of CLEWS embeddings by transformation category (colors as in Fig. 3 ).
Figure 3: CLEWS pairwise distances. Dashed line: calibrated threshold τ=986.4\tau=986.4 .

Explore similar work

CardsList

  1. Tracing Inputs, Verifying Outputs: Validating Attribution in Music Generation

    Oct 7, 2026Taejun Kim, Wonil Kim, Jongmin Jung +10Music Generation Evaluation

    Neutune, Seoul, South Korea · KAIST, Daejeon, South Korea

  2. On the Human and Computer Alignment of Attribute-Based Music Matches

    Sep 1, 2026Roser Batlle-Roca, Woosung Choi, Joan Serrà +5Music Generation EvaluationMusic Information Retrieval

    Music Technology Group, Universitat Pompeu Fabra · Sony AI · Joint Research Centre, European Commission +1

  3. Finding the noise: Zero-shot AI Music Detection

    Jul 28, 2026Darius Afchar, Romain HennequinAudio Deepfake DetectionAI-Generated Music Detection

    Deezer Research