Adversarial Optimization for Dictionary Attacks on Speaker Verification

Mirko Marras, Paweł Korus, Nasir Memon, Gianni FenuInterspeech 2019

TL;DR

This paper investigates the vulnerability of speaker verification systems to "dictionary attacks," where a single "master voice" is optimized to impersonate a large number of users by chance. The authors demonstrate that such attacks are feasible and can achieve significant impersonation rates, even against modern, robust systems, posing a real threat to high-stakes applications.

What Problem Does This Paper Address?

Speaker verification systems, which authenticate individuals based on their voice, are increasingly used for security and convenience. While traditional attacks focus on impersonating a specific individual (e.g., using synthesized speech), this paper explores a different threat: "dictionary attacks." These attacks, inspired by similar work in fingerprinting, aim to create a single "master voice" that can fool the system into accepting it as many different users, without prior knowledge of those users. The core problem is to understand how susceptible current speaker verification systems are to this type of broad-spectrum attack and to develop methods for generating such master voices.

Methods

The authors propose a two-pronged approach:

  1. Menagerie Analysis: They first analyze existing speech data to identify "natural wolves" – utterances that, by chance, already exhibit a high rate of false acceptance against other users. This helps understand the baseline vulnerability.
  2. Adversarial Optimization: To generate more potent master voices, they employ an adversarial optimization technique. This involves:
  3. Spectrogram Representation: Converting audio into spectrograms, a visual representation of sound frequencies over time.
  4. Siamese Network: Using a Siamese network architecture to compare spectrograms and compute similarity scores.
  5. Iterative Gradient Descent: Applying an iterative process where the system learns to slightly perturb the input spectrogram of a potential master voice. The goal is to maximize the probability of the system misclassifying this perturbed voice as belonging to a large number of enrolled users.
  6. Gradient Filtering: Crucially, they filter gradients to focus on users with whom the current master voice has a non-negligible chance of matching, avoiding futile optimization directions.
  7. Spectrogram-to-Waveform Inversion: Finally, the optimized spectrogram is converted back into an audio waveform using the Griffin-Lim algorithm.

Datasets

  • VoxCeleb: The study leverages the VoxCeleb dataset, a large corpus of unconstrained speech from celebrities.
  • Avc2-Train: Used for exploration, training, and optimization.
  • Avc2-Test: A separate, disjoint population used for evaluating the performance of generated master voices.
  • Both Avc2-Train and Avc2-Test consist of 1,000 speakers each, equally divided between males and females, with 50 utterances per speaker.

Key Insights and Ideas

  • "Wolves" Exist Naturally: Even without adversarial optimization, some individuals' voices naturally exhibit a higher tendency to be misclassified as other users. This suggests an inherent vulnerability in speaker verification systems.
  • Gender Bias Matters: The gender of the speakers significantly impacts impersonation rates. The study found that gender-biased verification models can exacerbate the difference in vulnerability between genders.
  • Master Voices Generalize: Adversarially generated master voices are not only effective against the specific population they were trained on but also generalize well to unseen populations.
  • Content vs. Speaker: While speaker characteristics are important, the study suggests that speech content and background noise also play a role, and adversarial optimization can overcome these limitations.
  • Feasible Threat Model: Dictionary attacks are presented as a realistic and significant threat, especially for high-stakes applications where robust security is paramount.

Key Results

  • Significant Impersonation Rates: Even in conservative settings, a master voice can match approximately 20% of females and 10% of males within a single presentation attempt.
  • Improvement Over Seed Voices: Adversarial optimization significantly improves impersonation rates compared to naturally occurring "seed voices" (SVs), with an average improvement of 20 percentage points for the Equal Error Rate (EER) threshold and 10 percentage points for the False Acceptance Rate (FAR) of 1%.
  • Intra-Sex Matching Dominance: The most significant improvements in impersonation rates were observed when the seed, training, and testing genders were the same (e.g., male-male-male).
  • Robustness: Master voices are robust to spectrogram computation and inversion processes.

Limitations and Future Work

  • Limitations:
    • The study acknowledges the impact of decision thresholds on impersonation rates. While a 1% FAR might seem high, achieving acceptable True Positive Rates (TPRs) for stricter thresholds remains a challenge for state-of-the-art models.
  • The current work focuses on specific verification policies (Any10, Avg10) and thresholds.
  • Future Work:
    • Investigate master voice transferability across different speaker verification systems.
  • Explore the use of generative models for master voice generation.

Summary updated

BibTeX
@inproceedings{marras2019adversarial,
  author = {Mirko Marras and Paweł Korus and Nasir Memon and Gianni Fenu},
  title = {{Adversarial Optimization for Dictionary Attacks on Speaker Verification}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure