Source-linked AI summary
Towards AI-Generated Music Plagiarism Detection as a Version Identification Problem
Fotis Koutsikos, Ioannis Prokopiou, Spyridon Kantarelis, Vassilis Lyberatos, Pantelis Vikatos, Athanasios Aidinis, Themos Stafylakis, Athanasios Voulodimos, Giorgos Stamou
TL;DR
Detecting plagiarism in AI-generated music is challenging because re-synthesis can preserve musical content while obscuring surface similarity. The paper frames the task as version identification, introduces COPYCAT, and reports that supervised coordinate-wise modeling raises overall F0.5 from 0.612 to 0.803.
Problem
Whether pretrained music version identification representations transfer from human-to-human covers to human-to-AI plagiarism remains underexplored, especially when re-synthesis obscures surface similarity.
Source: §1 · p. 1Method
The paper introduces COPYCAT, a benchmark of controlled human and AI-derived pairs, and evaluates MVI models with supervised classification over coordinate-wise embedding shifts.
Source: §1 · p. 1, §2.2 · p. 2Results
F0.5 rose from 0.612 to 0.803 with the Hybrid Top-512 classifier using coordinate-wise shifts.
Source: §4 · p. 4Takeaways & Limitations
The paper concludes that AI-derived plagiarism signal is redistributed across embedding coordinates rather than lost, beyond what scalar distance thresholds capture.
Source: §4 · p. 4The system's output is a technical similarity judgment, not a legal determination of substantial similarity.
Source: §2.1 · p. 2
Abstract
from arXiv · showhide
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 $F_{0.5}$ from $0.612$ to $0.803$.
1. INTRODUCTION
Music plagiarism detection faces legal ambiguity and limited verified cases, while existing computational work has largely focused on human-to-human relationships. This work frames AI-generated plagiarism as version identification and introduces COPYCAT to study transfer to human-to-AI cases.
1. INTRODUCTION
- Legal assessment considers melody, harmony, rhythm, and overall impression, but no agreed standard defines substantial similarity.Scarcity of legally verified plagiarism cases also limits automated-method development and evaluation.Source: §1 · p. 1
- Computational approaches have primarily addressed human-to-human musical relationships, leaving transfer of MVI representations to human-to-AI plagiarism underexplored.MVI recognizes corresponding recordings despite changes in performance, instrumentation, tempo, and production.Source: §1 · p. 1
- Generative re-synthesis can preserve musical content while introducing opaque acoustic distortions that render surface-level similarity measures ineffective.This makes re-synthesis a particularly acute case for testing transfer from human covers.Source: §1 · p. 1
- COPYCAT provides 350,654 controlled pairwise comparisons built from the SMP dataset and expanded with three generative architectures and signal-level obfuscations.The framework evaluates CLEWS and WEALY, finding that supervised classification over coordinate-wise embedding shifts recovers plagiarism signal that scalar distances do not fully exploit.Source: §1 · p. 1
2. METHODOLOGY
The methodology defines plagiarism detection as a pairwise version-identification judgment, then constructs controlled examples spanning human reinterpretation, signal obfuscation, and AI re-synthesis. It compares pretrained MVI branches using scalar thresholds, score fusion, and supervised coordinate-wise shifts.
2.1. Problem Formulation
- An original-candidate pair is evaluated as a possible version or derivative using embedding distance and a decision threshold, with the resulting judgment treated as technical similarity rather than a legal determination.The formulation applies regardless of whether the candidate was produced by a human or a generative model.Source: §2.1 · p. 2
2.2. Benchmark Design
- The benchmark uses fixed-length excerpts from alleged similar passages as anchors, then pairs them with positive examples transformed through human reinterpretation, DSP, AI re-synthesis, or combined AI and DSP.Transformation categories support evaluation across increasing concealment severity, while hard distractors can provide more challenging negatives.Source: §2.2 · p. 2
2.3. Modeling
- Modeling compares acoustic CLEWS and semantic WEALY through tuned distance thresholds, vocal-aware score fusion, and supervised classification on absolute coordinate-wise embedding shifts.The shift approach tests engineered features, the full shift vector, and a hybrid representation with selected volatile dimensions.Source: §2.3 · p. 2
3. EXPERIMENTS
COPYCAT combines real-world plagiarism allegations with human covers, AI re-synthesis, and DSP transformations to evaluate pretrained music-version-identification models. Scalar thresholds struggle with AI-derived cases, while coordinate-wise classification recovers dispersed signals, with the Hybrid Top-512 model reaching overall F0.5 = 0.803.
3.1. Experimental Setup
- The benchmark uses 1,219 twenty-second anchors from 70 plagiarism allegations, expanded into 350,654 comparisons through three generative architectures, DSP transformations, and mined negative pairs.The generative set includes full re-synthesis and stem-conditioned variants; source audio is commercially licensed, so reproducibility artifacts are released instead of raw audio.Source: §3.1 · pp. 2–3
- Frozen CLEWS and WEALY embeddings feed calibrated distance thresholds and XGBoost classifiers evaluated with stratified cross-validation grouped by original-track filename.The hybrid retains volatile CLEWS dimensions selected through a convergence sweep; classifier results are averaged over 10 seeds with 95% confidence intervals.Source: §3.1 · p. 3
3.2. Results
- F0.5 rises from 0.612 for the CLEWS threshold to 0.803 for the final Hybrid Top-512 model.The hybrid combines 24 engineered features with the 512 most volatile CLEWS dimensions.Source: §3.2 · p. 3
- The 24 engineered features yield only marginal improvement over the CLEWS threshold (F0.5 = 0.617 vs 0.612), while the full 1024-dimensional CLEWS shift reaches F0.5 = 0.811.Source: §3.2 · p. 3
- WEALY distances rank above CLEWS distances among engineered features, and their cumulative contribution adds +0.19 F0.5 over the engineered-only baseline.This reveals semantic-branch value that scalar score fusion did not expose.Source: §3.2 · p. 3
- Origin-centered UMAP shows source-specific displacement: AudioLDM 2 and MusicGen are most distant, MGE-LDM intermediate, and human covers closest to anchors.DSP variants form peripheral shells around base derivatives, with radial patterns growing with concealment severity.Source: §3.2 · p. 3, Figure 2 · p. 4
- Hybrid precision ranges from 97.3% on human plagiarism to 72.0% on AI+DSP, while recall rises from 30.6% to 86.5% and from 10.3% to 50.0%, respectively.Table 1 compares category-wise performance across WEALY, CLEWS threshold baselines, and the XGBoost Hybrid Top-512 classifier.Source: §3.2 · p. 3, Table 1 · p. 4
- MusicGen and AudioLDM 2 distances remain far above the calibrated threshold with flat DSP trendlines, while pitch-shifted originals remain within the detectable region.The plot overlays τ = 986.4 as a visual reference; re-synthesis has already displaced embeddings near maximal acoustic distance, leaving coordinate-wise structure rather than magnitude as the useful signal.Source: §3.2 · p. 4, Figure 3 · p. 4
4. CONCLUSIONS
The paper evaluates pretrained music-version-identification models on COPYCAT, a benchmark of human plagiarism, DSP obfuscation, and AI re-synthesis. Coordinate-wise shifts recover plagiarism signals that scalar thresholds miss, but generative re-synthesis with adversarial DSP remains difficult and the framework is positioned for expert-review triage.
4. CONCLUSIONS
- The Hybrid Top-512 classifier raises F0.5 from 0.612 to 0.803 by recovering plagiarism signals dispersed across embedding coordinates.The largest gains occur on AI+DSP cases.Source: §4 · p. 4
- Human plagiarism and simple DSP remain within reach, but generative re-synthesis combined with adversarial DSP pushes recall to its limits.The paper characterizes AI-mediated derivation as a harder regime than human-to-human plagiarism.Source: §4 · p. 4
- The framework is best used to surface candidate cases for expert review, not as a substitute for legal judgment.Future work includes broader generative architectures and DSP perturbations, alternative MVI backbones, encoder fine-tuning, lyric-preserving datasets, and source attribution.Source: §4 · p. 4