Source-linked AI summary

Fooling End-to-end Speaker Verification by Adversarial Examples

Felix Kreuk, Yossi Adi, Moustapha Cisse, Joseph Keshet

arXiv:1801.03339v2cs.LGcs.CL

TL;DR

The paper asks whether end-to-end neural speaker-verification systems are vulnerable to adversarial examples. It generates human-indistinguishable perturbations and evaluates white-box and black-box attacks, finding reduced accuracy and sharply increased false-positive rates.

  • Problem

    Speaker verification is deployed for biometric authentication, but the vulnerability of end-to-end neural systems to adversarial examples requires investigation.

  • Method

    The paper generates adversarial examples for an end-to-end neural speaker-verification model by maximizing loss under a constrained perturbation.

  • Results

    The attacks reduce verification accuracy and dramatically increase false-positive rates across white-box and black-box settings.

  • Takeaways & Limitations

    Adversarial examples can make utterances sound like speaker A while enabling the system to claim they were produced by speaker B.

  • Takeaways & Limitations

    Future work remains to evaluate traditional speaker-verification systems and apply adversarial training to neural-based systems.

Abstract

from arXiv · show

Automatic speaker verification systems are increasingly used as the primary means to authenticate costumers. Recently, it has been proposed to train speaker verification systems using end-to-end deep neural models. In this paper, we show that such systems are vulnerable to adversarial example attack. Adversarial examples are generated by adding a peculiar noise to original speaker examples, in such a way that they are almost indistinguishable from the original examples by a human listener. Yet, the generated waveforms, which sound as speaker A can be used to fool such a system by claiming as if the waveforms were uttered by speaker B. We present white-box attacks on an end-to-end deep network that was either trained on YOHO or NTIMIT. We also present two black-box attacks: where the adversarial examples were generated with a system that was trained on YOHO, but the attack is on a system that was trained on NTIMIT; and when the adversarial examples were generated with a system that was trained on Mel-spectrum feature set, but the attack is on a system that was trained on MFCC. Results suggest that the accuracy of the attacked system was decreased and the false-positive rate was dramatically increased.

1. INTRODUCTION

Speaker verification authenticates claimed speakers and is deployed in banking and e-commerce, making its vulnerability a practical security threat. This paper investigates adversarial examples against end-to-end neural speaker verification, distinguishing them from spoofing attacks.

  • Speaker verification checks whether a spoken utterance was produced by the claimed speaker.
  • The technology is used for biometric authentication by banks and e-commerce services online and by phone.
  • End-to-end neural models learn directly from utterances, capturing long-range context while reducing system complexity.
  • Adversarial examples add small, carefully chosen perturbations that can change a neural network’s prediction while remaining indistinguishable to human listeners.
  • The paper generates attacks against an end-to-end neural speaker-verification model using the architecture proposed in.
  • Unlike spoofing with a falsifying voice, the approach creates utterances that sound like speaker A but can be accepted as speaker B.

2. NOTATIONS AND DEFINITIONS

The paper formalizes speaker verification as a neural function that compares an unknown utterance with enrollment utterances and estimates whether they belong to the claimed speaker.

  • An acoustic speech signal is represented as a variable-length sequence of feature vectors x = (x_1, ..., x_T).
  • The verification system receives an unknown utterance and n enrollment utterances associated with speaker k.
  • Its output is a probability p ∈ [0, 1] estimating whether the unknown utterance was produced by speaker k.
  • The neural function g_θ is trained by minimizing negative log-likelihood over utterance, enrollment-set, and binary-label tuples.

3. END-TO-END DEEP NETWORK MODEL

The end-to-end model embeds the unknown and enrollment utterances with recurrent LSTMs, averages enrollment embeddings, and converts cosine similarity into a verification probability.

  • The model architecture was initially proposed in and serves as a baseline for end-to-end speaker-verification work.
  • Each unknown or enrollment utterance is processed by a recurrent LSTM into an embedding vector of size D.
  • The n enrollment embeddings are averaged into a single speaker representation.
  • The model compares the unknown-utterance embedding with the averaged enrollment embedding using cosine similarity.
  • A final affine layer maps cosine similarity to a probability estimate and acts as an automatically learned detection threshold.

4. GENERATING ADVERSARIAL EXAMPLES

Adversarial examples are generated by perturbing the test utterance within a norm constraint to increase verification loss while preserving perceptual similarity.

  • An adversarial example is a perturbed version of an input that remains human-indistinguishable but causes an incorrect network label.
  • The attack optimizes the perturbed utterance to maximize loss under a p-norm distance bound of ϵ from the original.
  • Here, ϵ controls adversarial strength and p specifies the norm used to constrain similarity to the original.
  • A first-order Taylor approximation yields the fast gradient sign method, which adds a small scaled gradient-sign perturbation to the input.
  • To model a realistic attack, noise is added only to the test utterance while the enrollment set remains unchanged.

5. EXPERIMENTS

Experiments evaluate adversarial attacks on speaker-verification models across white-box, cross-dataset, and cross-feature settings, while testing whether the examples remain perceptually indistinguishable. The attacks reduced accuracy and substantially degraded false-positive-rate performance.

  • Experimental setup: The experimental pipeline perturbed acoustic-feature vectors, reconstructed adversarial waveforms, and tested them against the original and differently trained models.Models were trained on YOHO or NTIMIT using Mel-spectrum or MFCC features.
  • Experimental setup: Adversarial examples generated with fast gradient sign perturbations using ϵ ∈(0.2, 0.3) misclassified examples with high probability while remaining similar to the originals.The attack perturbed test utterance vectors before waveform reconstruction.
  • ABX testing: Human listeners correctly classified 54% of original-versus-adversarial waveform pairs in the ABX test, supporting their perceptual similarity.The test used 50 audio pairs and eight listeners per pair.
  • White-box attack: White-box attacks were evaluated using system accuracy and false-positive rate, with Table 1 reporting accuracy and Table 2 reporting FPR.The white-box setting assumes access to model internals and gradients.
  • Cross-dataset: In the cross-dataset attack, model A’s accuracy fell from 81.55% on the clean test set to 58.93% on the adversarial set, a difference of 22.62%.The FPR increased from 12% to 46%.
  • Cross-features: In the cross-feature attack, the MFCC model’s accuracy fell from 81% on the clean test set to 62.25% on the adversarial set, a difference of 18.75%.The FPR increased from 16% to 46%.

6. CONCLUSION

End-to-end neural speaker-verification models require reliability evaluation beyond traditional accuracy metrics. The authors identify robustness evaluation and adversarial training as future directions.

  • Improved accuracy over traditional speech-verification components makes broader reliability evaluation critical.
  • Future work includes evaluating traditional speaker-verification systems against adversarial examples and applying adversarial training to neural-based systems.
Loading 1801.03339v2…