pith. sign in

arxiv: 2606.30780 · v1 · pith:PAXO3NHEnew · submitted 2026-06-29 · 📡 eess.AS · cs.AI

Detecting Audio Deepfakes on the Edge:Lightweight SSL-Based Detection in a Browser Plugin

Octavian Pascu , Dan Oneata , Horia Cucu , Nicolas M. Muller This is my paper

Pith reviewed 2026-07-01 01:44 UTC · model grok-4.3

classification 📡 eess.AS cs.AI
keywords audio deepfakesdeepfake detectionon-device AIself-supervised learningbrowser pluginedge computingprivacy preservation
0
0 comments X

The pith

A truncated self-supervised backbone with logistic classifier detects audio deepfakes on-device faster and more accurately than existing solutions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that audio deepfake detection can be performed locally in a browser using a lightweight model derived from self-supervised learning. By truncating a self-supervised backbone and pairing it with a simple logistic classifier, the approach achieves better accuracy and speed than the AASIST baseline while keeping data private. This matters because journalists and fact-checkers can verify audio authenticity without sending recordings to external servers. The model is integrated into a browser plugin for easy use.

Core claim

A truncated self-supervised backbone with a simple logistic classifier is both very fast and often more accurate than existing solutions, outperforming the AASIST baseline by 10% in accuracy and improving inference speed by 40%. This enables an on-device browser plugin for private deepfake detection.

What carries the argument

Truncated self-supervised backbone paired with a logistic classifier, which reduces computation while maintaining detection performance for on-device use.

If this is right

  • Journalists can check audio sources locally without privacy risks from cloud processing.
  • The plugin allows secure detection directly in the browser for fact-checkers.
  • Improved speed supports real-time or near-real-time verification of audio clips.
  • Lightweight design makes deepfake detection accessible on standard consumer devices.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar truncation techniques might apply to video or image deepfake detection for edge devices.
  • The approach could extend to other privacy-sensitive audio analysis tasks beyond deepfakes.
  • Performance on diverse recording conditions needs validation beyond the tested sets.

Load-bearing premise

The accuracy and speed improvements will generalize to deepfake generation methods and audio conditions not seen in the evaluation datasets.

What would settle it

Evaluating the model on audio deepfakes created by new synthesis methods or recorded in different environments and observing if the 10% accuracy gain over AASIST holds.

Figures

Figures reproduced from arXiv: 2606.30780 by Dan Oneata, Horia Cucu, Nicolas M. Muller, Octavian Pascu.

Figure 1
Figure 1. Figure 1: Audio Deepfake Detection Chrome Extension: (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Tradeoff between Out-Of-Domain performance, infer [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Out-of-domain EER vs model parameters for tested [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
read the original abstract

Audio deepfakes are a growing challenge for the general public, as well as for journalists and fact-checkers. The latter need reliable tools to verify the authenticity of their sources, while at the same time keeping their information private. Commercial deepfake detection solutions rely on cloud-based processing, which raises privacy concerns. To solve this problem, we propose an on-device audio deepfake detection model. We show that a truncated self-supervised backbone with a simple logistic classifier is both very fast and often more accurate than existing solutions. Our solution outperforms the baseline AASIST by 10% and improves inference speed by 40%. We integrate this model into a browser plug-in, which allows journalists and fact-checkers to detect deepfakes easily and securely. Code for the plugin is available at https://github.com/OctavianPascu97/Audio-Deepfakes-Browser-Plugin.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper proposes an on-device audio deepfake detector consisting of a truncated self-supervised learning backbone paired with a logistic classifier. This model is claimed to outperform the AASIST baseline by 10% accuracy while running 40% faster, and is packaged as a browser plugin for private, local verification by journalists and fact-checkers.

Significance. If the accuracy and speed claims prove robust under proper evaluation, the work would offer a practical, privacy-preserving alternative to cloud-based detectors and demonstrate that lightweight SSL features can suffice for this task. The browser-plugin deployment itself is a concrete engineering contribution.

major comments (3)
  1. [Abstract] Abstract: The central empirical claims (10% accuracy gain and 40% inference speedup versus AASIST) are stated without any description of the evaluation dataset(s), train/test splits, number of deepfake generators, acoustic conditions, or statistical significance testing. This absence makes the reported margins unverifiable and prevents assessment of generalization.
  2. No section supplies the identity of the SSL backbone, the precise truncation strategy (layer indices or feature dimensions), training protocol, or any ablation that isolates the contribution of truncation versus the logistic head. Without these details the reproducibility of the speed/accuracy trade-off cannot be checked.
  3. The manuscript contains no cross-generator or out-of-distribution experiments. The weakest assumption—that the reported margin will hold on unseen synthesis pipelines and microphone characteristics—is therefore untested, directly undermining the practical claim for a browser plugin that will encounter arbitrary recording conditions.
minor comments (1)
  1. The GitHub link for the plugin is provided, but the manuscript does not indicate whether the model weights or training scripts are also released.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting areas where additional clarity and evaluation would strengthen the manuscript. We address each major comment below and will revise accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central empirical claims (10% accuracy gain and 40% inference speedup versus AASIST) are stated without any description of the evaluation dataset(s), train/test splits, number of deepfake generators, acoustic conditions, or statistical significance testing. This absence makes the reported margins unverifiable and prevents assessment of generalization.

    Authors: We agree the abstract is overly concise and omits key evaluation context. In the revised manuscript we will expand the abstract to specify the datasets, train/test splits, number of generators, acoustic conditions, and note that statistical significance testing was performed, while preserving length limits. revision: yes

  2. Referee: No section supplies the identity of the SSL backbone, the precise truncation strategy (layer indices or feature dimensions), training protocol, or any ablation that isolates the contribution of truncation versus the logistic head. Without these details the reproducibility of the speed/accuracy trade-off cannot be checked.

    Authors: We acknowledge these details are missing from the current manuscript. The revision will add a dedicated methods subsection specifying the SSL backbone, exact truncation (layers and dimensions), training protocol, and ablation experiments isolating truncation versus the logistic head. revision: yes

  3. Referee: The manuscript contains no cross-generator or out-of-distribution experiments. The weakest assumption—that the reported margin will hold on unseen synthesis pipelines and microphone characteristics—is therefore untested, directly undermining the practical claim for a browser plugin that will encounter arbitrary recording conditions.

    Authors: We agree this is a limitation of the presented evaluation. The revision will add explicit discussion of generalization limits and, to the extent feasible with existing data, preliminary cross-generator results or analysis of acoustic variability to better support the browser-plugin use case. revision: partial

Circularity Check

0 steps flagged

No circularity: purely empirical model comparison with no derivation chain

full rationale

The manuscript contains no mathematical derivation, equations, or claimed first-principles results. All central claims consist of empirical measurements (accuracy and latency on specific datasets versus the external AASIST baseline). No self-definitional loops, fitted inputs renamed as predictions, or load-bearing self-citations appear in the provided text. The work is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no equations, parameters, or assumptions; ledger is empty by necessity.

pith-pipeline@v0.9.1-grok · 5693 in / 970 out tokens · 24589 ms · 2026-07-01T01:44:49.172641+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 8 canonical work pages

  1. [1]

    Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks,

    J.-w. Jung, H.-S. Heo, H. Tak, H.-j. Shim, J. S. Chung, B.-J. Lee, H.-J. Yu, and N. Evans, “Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks,” inICASSP 2022-2022 IEEE international conference on acoustics, speech and signal processing (ICASSP), pp. 6367–6371, IEEE, 2022

    2022

  2. [2]

    Pushing the limits of raw waveform speaker recognition,

    J.-w. Jung, Y . J. Kim, H.-S. Heo, B.-J. Lee, Y . Kwon, and J. S. Chung, “Pushing the limits of raw waveform speaker recognition,”arXiv preprint arXiv:2203.08488, 2022

    work page arXiv 2022

  3. [3]

    Towards gen- eralisable and calibrated audio deepfake detection with self-supervised representations,

    O. Pascu, A. Stan, D. Oneata, E. Oneata, and H. Cucu, “Towards gen- eralisable and calibrated audio deepfake detection with self-supervised representations,” inInterspeech, vol. 2024, pp. 4828–4832, 2024

    2024

  4. [4]

    End-to-end spectro-temporal graph attention networks for speaker ver- ification anti-spoofing and speech deepfake detection,

    H. Tak, J.-w. Jung, J. Patino, M. Kamble, M. Todisco, and N. Evans, “End-to-end spectro-temporal graph attention networks for speaker ver- ification anti-spoofing and speech deepfake detection,”arXiv preprint arXiv:2107.12710, 2021

    work page arXiv 2021

  5. [5]

    End-to-end anti-spoofing with rawnet2,

    H. Tak, J. Patino, M. Todisco, A. Nautsch, N. Evans, and A. Larcher, “End-to-end anti-spoofing with rawnet2,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 6369–6373, IEEE, 2021

    2021

  6. [6]

    Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,

    H. Tak, M. Todisco, X. Wang, J.-w. Jung, J. Yamagishi, and N. Evans, “Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation,”arXiv preprint arXiv:2202.12233, 2022

    work page arXiv 2022

  7. [7]

    Deepfake total

    “Deepfake total.” [Online; accessed 2025-04-07]

    2025

  8. [8]

    Sensity ai: Best deepfake detection software in 2025,

    “Sensity ai: Best deepfake detection software in 2025,” 3 2024. [Online; accessed 2025-04-07]

    2025

  9. [9]

    Harder or different? understanding generalization of audio deepfake detection,

    N. M. M ¨uller, N. Evans, H. Tak, P. Sperl, and K. B ¨ottinger, “Harder or different? understanding generalization of audio deepfake detection,” arXiv preprint arXiv:2406.03512, 2024

    work page arXiv 2024

  10. [10]

    Deepfake voice detector

    “Deepfake voice detector.” [Online; accessed 2025-04-02]

    2025

  11. [11]

    Layer-wise analysis of a self-supervised speech representation model,

    A. Pasad, J.-C. Chou, and K. Livescu, “Layer-wise analysis of a self-supervised speech representation model,” in2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pp. 914–921, IEEE, 2021

    2021

  12. [12]

    Audio deepfake detection with self-supervised wavlm and multi-fusion attentive classi- fier,

    Y . Guo, H. Huang, X. Chen, H. Zhao, and Y . Wang, “Audio deepfake detection with self-supervised wavlm and multi-fusion attentive classi- fier,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 12702–12706, IEEE, 2024

    2024

  13. [13]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” Advances in neural information processing systems, vol. 33, pp. 12449– 12460, 2020

    2020

  14. [14]

    Xls-r: Self-supervised cross-lingual speech representation learning at scale.arXiv preprint arXiv:2111.09296, 2021

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. V on Platen, Y . Saraf, J. Pino,et al., “Xls-r: Self-supervised cross-lingual speech representation learning at scale,”arXiv preprint arXiv:2111.09296, 2021

    work page arXiv 2021

  15. [15]

    ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech,

    X. Wang, J. Yamagishi, M. Todisco, H. Delgado, A. Nautsch, N. Evans, M. Sahidullah, V . Vestman, T. Kinnunen, K. A. Lee,et al., “ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech,”Computer Speech & Language, vol. 64, p. 101114, 2020

    2019

  16. [16]

    ASVspoof 2021: accelerating progress in spoofed and deepfake speech detection,

    J. Yamagishi, X. Wang, M. Todisco, M. Sahidullah, J. Patino, A. Nautsch, X. Liu, K. A. Lee, T. Kinnunen, N. Evans,et al., “Asvspoof 2021: accelerating progress in spoofed and deepfake speech detection,” arXiv preprint arXiv:2109.00537, 2021

    work page arXiv 2021

  17. [17]

    FoR: A dataset for synthetic speech detection,

    R. Reimao and V . Tzerpos, “FoR: A dataset for synthetic speech detection,” in2019 International Conference on Speech Technology and Human-Computer Dialogue (SpeD), pp. 1–10, IEEE, 2019

    2019

  18. [18]

    MLAAD: The multi-language audio anti-spoofing dataset,

    N. M. M ¨uller, P. Kawa, W. H. Choong, E. Casanova, E. G¨olge, T. M¨uller, P. Syga, P. Sperl, and K. B¨ottinger, “MLAAD: The multi-language audio anti-spoofing dataset,” in2024 International Joint Conference on Neural Networks (IJCNN), pp. 1–7, IEEE, 2024

    2024

  19. [19]

    Does audio deepfake detection generalize?

    N. M. M ¨uller, P. Czempin, F. Dieckmann, A. Froghyar, and K. B ¨ottinger, “Does audio deepfake detection generalize?,”arXiv preprint arXiv:2203.16263, 2022

    work page arXiv 2022

  20. [20]

    TIMIT- TTS: A text-to-speech dataset for multimodal synthetic media detection,

    D. Salvi, B. Hosler, P. Bestagini, M. C. Stamm, and S. Tubaro, “TIMIT- TTS: A text-to-speech dataset for multimodal synthetic media detection,” IEEE access, vol. 11, pp. 50851–50866, 2023

    2023

  21. [21]

    WaveFake: A data set to facilitate audio deepfake detection,

    J. Frank and L. Sch ¨onherr, “WaveFake: A data set to facilitate audio deepfake detection,”arXiv preprint arXiv:2111.02813, 2021

    work page arXiv 2021