A Multi-condition Training Strategy for Countermeasures Against Spoofing Attacks to Speaker Recognizers

João Monteiro, Jahangir Alam, Tiago H. FalkOdyssey 2020 · DOI

TL;DR

This paper tackles the challenge of training robust detectors for spoofing attacks on speaker recognition systems, particularly focusing on end-to-end deep learning models. They propose a data augmentation strategy that significantly increases training data size and diversity, leading to improved detection performance compared to traditional methods.

What Problem Does This Paper Address?

Automatic speaker recognition (ASR) systems, which identify individuals by their voice, are increasingly being deployed in real-world applications. However, these systems are vulnerable to "spoofing attacks," where malicious actors attempt to fool the ASR system using various techniques like replaying recorded voices or generating synthetic speech. While deep learning models have shown great promise in ASR, they can be susceptible to subtle input variations (adversarial attacks) and simpler spoofing methods. A key challenge in developing effective countermeasures is the limited availability of diverse spoofing attack data for training robust detectors. This paper aims to address this by developing an effective training strategy for end-to-end spoofing detectors that can generalize well to unseen attacks.

Methods

The core of the paper's contribution lies in its data augmentation strategy and the use of an end-to-end Time Delay Neural Network (TDNN) architecture.

  • End-to-End TDNN: The authors employ a TDNN architecture, similar to the one used for x-vectors in speaker verification. This architecture processes raw audio (or spectral features) directly and outputs a score indicating the likelihood of a spoofing attack. The "end-to-end" aspect means the system is a single component, simplifying inference compared to multi-stage pipelines.
  • Data Augmentation: To overcome the scarcity of spoofing data, they implement a multi-condition training strategy. This involves:
  • Offline Augmentation: Applying speed perturbation (factors of 0.9 and 1.1), low-pass filtering (cutoff at 3.8 kHz), and high-pass filtering (cutoff at 3.8 kHz) to the existing training data. This increases the data size by a factor of five.
  • Online Augmentation: During mini-batch construction, they implement strategies to ensure data diversity and robustness to varying audio durations. This includes repeating short segments and randomly trimming longer ones to a consistent length.
  • Balancing Mini-batches: To address the imbalance between genuine and spoofing examples, they employ an oversampling strategy for genuine samples.

Datasets

The primary dataset used is the ASVspoof 2019 challenge dataset. This dataset includes both "logical access" (LA) attacks (synthetic speech) and "physical access" (PA) attacks (replayed speech). The authors use the training and development partitions for training and hyperparameter tuning, and the evaluation partition for final performance assessment. They do not introduce a new benchmark dataset but focus on improving performance on an existing one.

Key Insights and Ideas

  • Data Augmentation is Crucial: The paper strongly emphasizes that effective data augmentation is critical for training robust end-to-end spoofing detectors, especially given the limited availability of spoofing data.
  • End-to-End Simplicity: The adoption of an end-to-end TDNN architecture simplifies the detection pipeline, allowing for a single forward pass for inference, which is a significant advantage over traditional multi-stage systems.
  • Preserving Attack Artifacts: A key challenge in augmentation is to preserve the artifacts introduced by spoofing attacks while introducing new, non-attack-related variations. The chosen augmentation techniques (speed perturbation, filtering) are shown to achieve this for replay attacks.
  • Generalization vs. Overfitting: The paper highlights a trade-off. While extensive augmentation can improve generalization, it can sometimes lead to a slight degradation in performance on the development set if the augmented data starts to resemble attack artifacts too closely. However, this can be beneficial for overall robustness.

Key Results

The proposed end-to-end TDNN models, trained with their multi-condition data augmentation strategy, significantly outperform traditional GMM-based classifiers and even their own i-vector/PLDA baselines on both physical access (PA) and logical access (LA) tasks.

  • PA Task (Development Set): The best performing model (TDNN+Filt.) achieved an EER of 0.74% and a min-tDCF of 0.0244, a substantial improvement over baselines.
  • LA Task (Development Set): The best performing model (TDNN+SP) achieved an EER of 0.04% and a min-tDCF of 0.0012, also outperforming baselines.
  • Evaluation Set Performance: Similar trends were observed on the evaluation set, with the TDNN+SP+Filt. model achieving an EER of 1.77% and min-tDCF of 0.0597 for PA, and an EER of 7.12% and min-tDCF of 0.1674 for LA.

Limitations and Future Work

  • Development vs. Evaluation Mismatch: The authors note a performance mismatch between the development and evaluation sets for LA attacks. They hypothesize this is due to different attack generation strategies used for these partitions, and that while augmentation helps, some artifacts might overlap.
  • Potential Overlap of Augmentation and Attack Artifacts: For LA attacks, the augmentation artifacts (e.g., from speed perturbation) can sometimes overlap with synthetic attack artifacts, leading to a slight degradation.
  • Future Work: The authors suggest that a fully online implementation of their augmentation strategy could further enhance data diversity and make the training data size virtually unbounded. They also mention that their approach could be extended to other types of spoofing attacks.

Summary updated

BibTeX
@inproceedings{monteiro2020multi,
  author = {João Monteiro and Jahangir Alam and Tiago H. Falk},
  title = {{A Multi-condition Training Strategy for Countermeasures Against Spoofing Attacks to Speaker Recognizers}},
  booktitle = {Odyssey},
  year = {2020},
  doi = {10.21437/odyssey.2020-42},
}
Figure