A Light Convolutional GRU-RNN Deep Feature Extractor for ASV Spoofing Detection

Alejandro Gomez-Alanis, Antonio M. Peinado, Jose A. Gonzalez, Angel M. GomezInterspeech 2019

TL;DR

This paper introduces a novel deep learning architecture, the Light Convolutional GRU-RNN (LC-GRNN), designed to detect various types of ASV (Automatic Speaker Verification) spoofing attacks with a single system. By combining light convolutional layers for frame-level feature extraction and Gated Recurrent Units (GRUs) for temporal dependencies, the LC-GRNN achieves state-of-the-art performance on ASVspoof challenge datasets, outperforming previous methods.

What Problem Does This Paper Address?

Automatic Speaker Verification (ASV) systems are designed to authenticate a user's identity based on their voice. However, these systems are vulnerable to spoofing attacks, where malicious actors attempt to impersonate legitimate users. The ASVspoof challenge aims to develop robust systems capable of detecting these attacks. The primary challenge is to create a single, unified system that can effectively detect diverse spoofing methods, including text-to-speech (TTS), voice conversion (VC), and replay attacks, without requiring separate models for each attack type. Furthermore, it's crucial for these systems to generalize well to unseen attacks not present in the training data.

Methods

The core contribution of this paper is the Light Convolutional GRU-RNN (LC-GRNN) architecture. This hybrid model integrates two key components:

  • Light Convolutional Neural Networks (LCNNs): These are used for frame-level feature extraction. The "light" aspect comes from employing Max-Feature-Map (MFM) activations, which help reduce feature dimensionality and enhance discriminative power. These convolutional layers are designed to capture subtle artifacts characteristic of spoofed speech.
  • Gated Recurrent Units (GRUs): These are employed to learn long-term temporal dependencies within the extracted frame-level features. By processing sequences of features, GRUs can model how these features evolve over time, which is crucial for understanding the overall characteristics of an utterance.

The LC-GRNN architecture processes speech signals in context windows. At each time step, LCNNs extract features, and these are then fed into GRU layers. The output of the final GRU layer is then passed through fully connected layers to produce a spoofing identity vector, which is ultimately used by a classifier (SVM, LDA, or PLDA) to make the final genuine/spoofed decision. The paper highlights that this architecture is novel for spoofing detection, adapting techniques previously used in areas like video representation and optical character recognition.

Datasets

The proposed system was evaluated on the following datasets:

  • ASVspoof 2015: Primarily used for logical access (LA) attacks (TTS and VC).
  • ASVspoof 2017: Used for physical access (PA) attacks (replay).
  • ASVspoof 2019: The most recent challenge, used for both logical access (LA) and physical access (PA) attacks. The paper notes that for training, a subset of known attacks was used, while the system was evaluated on both known and unknown attacks to assess generalization capabilities.

Key Insights and Ideas

  • Unified Spoofing Detection: The LC-GRNN demonstrates that a single architecture can effectively handle diverse spoofing attack types (TTS, VC, replay), simplifying the development of anti-spoofing systems.
  • Frame-Level Discriminative Features: The integration of LCNNs with MFM activations proves highly effective in extracting subtle, discriminative features at the frame level, which are crucial for identifying spoofing artifacts.
  • Temporal Context Matters: GRUs are essential for capturing the temporal dynamics of speech, allowing the system to understand how frame-level features combine to form an utterance and distinguish genuine speech from spoofed versions.
  • Generalization to Unseen Attacks: The system shows significant improvement over baseline methods, indicating its ability to generalize to spoofing attacks not encountered during training.
  • Classifier Choice: While the LC-GRNN is the feature extractor, the choice of backend classifier (PLDA, LDA, SVM) impacts performance, with PLDA often yielding the best results, especially when combined with scoring normalization.

Key Results

The LC-GRNN system achieved state-of-the-art results across multiple ASVspoof challenges:

  • ASVspoof 2015: The proposed system with a PLDA classifier outperformed baseline systems and other deep feature-based methods, achieving a low Equal Error Rate (EER). For the S10 attack, PLDA without scoring normalization was particularly effective.
  • ASVspoof 2017: The LC-GRNN with PLDA and scoring normalization achieved the best performance, outperforming other state-of-the-art single systems with a significantly lower EER.
  • ASVspoof 2019: The system was submitted as a single system to the challenge and demonstrated superior performance compared to baseline systems (CQCC + GMM and LFCC + GMM) on both logical access and physical access tasks. For instance, on the ASVspoof 2019 PA evaluation set, the primary system achieved a relative improvement of 74.69% and 79.80% over the baselines.

Limitations and Future Work

  • Limitations: The paper mentions that scoring normalization is beneficial when unseen attacks are similar to seen ones, but less so when the genuine class is easily mistaken for an attack from a new technique (e.g., the MaryTTS-based S10 attack in ASVspoof 2015). This suggests that while the system is robust, certain types of novel attacks might still pose challenges.
  • Future Work: The authors do not explicitly detail future work directions in the provided text. However, the success of the LC-GRNN suggests potential avenues for further research could include exploring different LCNN configurations, integrating other types of recurrent units, or investigating more advanced backend classifiers. Further exploration into how scoring normalization interacts with different attack types could also be a valuable research direction.

Summary updated

BibTeX
@inproceedings{gomezalanis2019light,
  author = {Alejandro Gomez-Alanis and Antonio M. Peinado and Jose A. Gonzalez and Angel M. Gomez},
  title = {{A Light Convolutional GRU-RNN Deep Feature Extractor for ASV Spoofing Detection}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure