Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Detecting music deepfakes is easy but actually hard

Darius Afchar, Gabriel Meseguer-Brocal, Romain Hennequin

TL;DR

This work addresses the rising threat of music deepfakes by introducing a general-purpose detector that distinguishes real audio from autoencoder reconstructions, leveraging waveform-based inputs and a lightweight CNN. Using the FMA dataset and multiple decoders, the approach achieves very high detection accuracy (around 99%), suggesting artefact-based detection is feasible. However, the authors rigorously critique deployment risks, showing detectors can be brittle under common audio manipulations, fail to generalise to unseen decoders, suffer calibration issues, and require interpretable recourse strategies. The paper argues that robust, ethically sound deployment demands attention to robustness, generalisation, calibration, interpretability, and governance, outlining concrete future directions and cautions about overreliance on detection scores alone.

Abstract

In the face of a new era of generative models, the detection of artificially generated content has become a matter of utmost importance. The ability to create credible minute-long music deepfakes in a few seconds on user-friendly platforms poses a real threat of fraud on streaming services and unfair competition to human artists. This paper demonstrates the possibility (and surprising ease) of training classifiers on datasets comprising real audio and fake reconstructions, achieving a convincing accuracy of 99.8%. To our knowledge, this marks the first publication of a music deepfake detector, a tool that will help in the regulation of music forgery. Nevertheless, informed by decades of literature on forgery detection in other fields, we stress that a good test score is not the end of the story. We step back from the straightforward ML framework and expose many facets that could be problematic with such a deployed detector: calibration, robustness to audio manipulation, generalisation to unseen models, interpretability and possibility for recourse. This second part acts as a position for future research steps in the field and a caveat to a flourishing market of fake content checkers.

Detecting music deepfakes is easy but actually hard

TL;DR

This work addresses the rising threat of music deepfakes by introducing a general-purpose detector that distinguishes real audio from autoencoder reconstructions, leveraging waveform-based inputs and a lightweight CNN. Using the FMA dataset and multiple decoders, the approach achieves very high detection accuracy (around 99%), suggesting artefact-based detection is feasible. However, the authors rigorously critique deployment risks, showing detectors can be brittle under common audio manipulations, fail to generalise to unseen decoders, suffer calibration issues, and require interpretable recourse strategies. The paper argues that robust, ethically sound deployment demands attention to robustness, generalisation, calibration, interpretability, and governance, outlining concrete future directions and cautions about overreliance on detection scores alone.

Abstract

In the face of a new era of generative models, the detection of artificially generated content has become a matter of utmost importance. The ability to create credible minute-long music deepfakes in a few seconds on user-friendly platforms poses a real threat of fraud on streaming services and unfair competition to human artists. This paper demonstrates the possibility (and surprising ease) of training classifiers on datasets comprising real audio and fake reconstructions, achieving a convincing accuracy of 99.8%. To our knowledge, this marks the first publication of a music deepfake detector, a tool that will help in the regulation of music forgery. Nevertheless, informed by decades of literature on forgery detection in other fields, we stress that a good test score is not the end of the story. We step back from the straightforward ML framework and expose many facets that could be problematic with such a deployed detector: calibration, robustness to audio manipulation, generalisation to unseen models, interpretability and possibility for recourse. This second part acts as a position for future research steps in the field and a caveat to a flourishing market of fake content checkers.
Paper Structure (12 sections, 4 figures, 3 tables)

This paper contains 12 sections, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Detecting fake music seems easy
  3. Proposed framework
  4. Considered dataset and deepfake generators
  5. First results
  6. Caveats on deepfake detectors
  7. Robustness to manipulations
  8. Encoder generalisation
  9. Calibration and specifications
  10. Recourse and interpretation
  11. Conclusion
  12. Ethics statement

Figures (4)

  • Figure 1: Model calibration curve. Note that predictions in $[0.1, 0.9]$ are in minority, which results in higher confidence intervals at 95% in that range.
  • Figure 2: Mixing calibration curve. We generate an equal number of real/fake mixes (4955) per mixing factor. The confidence intervals are too small to be visible here.
  • Figure 3: Feature attribution example. Mixed real/fake spectrogram on the top, corresponding per-patch predictions on the bottom (red for real, blue for fake).
  • Figure :