Source-linked AI summary
STC Antispoofing Systems for the ASVspoof2019 Challenge
Galina Lavrentyeva, Sergey Novoselov, Andzhukaev Tseren, Marina Volkova, Artem Gorlanov, Alexandr Kozlov
TL;DR
The paper addresses generalized spoofing detection for logical-access synthesis and voice-conversion attacks and physical-access replay attacks. It develops enhanced Light CNN systems using varied acoustic features and angular-margin softmax, achieving high performance on unknown attacks in both conditions. However, modeled-data results may not transfer reliably to real cases.
Problem
Speaker-verification systems remain vulnerable to increasingly capable spoofing attacks, motivating reliable detection methods for logical-access and physical-access conditions.
Method
STC uses enhanced Light CNN antispoofing systems with different acoustic features and trains them using angular-margin softmax.
Results
High performance on unknown spoofing attacks demonstrates the stability of the offered approach in both evaluation conditions.
Takeaways & Limitations
Angular-margin softmax improves system quality and stabilizes training for both logical-access and physical-access scenarios.
Takeaways & Limitations
Modeled training and evaluation data raise concerns about mismatch with real-case spoofing attacks, and logical-access performance degrades for some unknown attack types.
Abstract
from arXiv · showhide
This paper describes the Speech Technology Center (STC) antispoofing systems submitted to the ASVspoof 2019 challenge. The ASVspoof2019 is the extended version of the previous challenges and includes 2 evaluation conditions: logical access use-case scenario with speech synthesis and voice conversion attack types and physical access use-case scenario with replay attacks. During the challenge we developed anti-spoofing solutions for both scenarios. The proposed systems are implemented using deep learning approach and are based on different types of acoustic features. We enhanced Light CNN architecture previously considered by the authors for replay attacks detection and which performed high spoofing detection quality during the ASVspoof2017 challenge. In particular here we investigate the efficiency of angular margin based softmax activation for training robust deep Light CNN classifier to solve the mentioned-above tasks. Submitted systems achieved EER of 1.86% in logical access scenario and 0.54% in physical access scenario on the evaluation part of the Challenge corpora. High performance obtained for the unknown types of spoofing attacks demonstrates the stability of the offered approach in both evaluation conditions.
1. Introduction
ASVspoof2019 addresses generalized spoofing detection in logical-access and physical-access scenarios because deployed speaker-verification systems remain vulnerable to increasingly capable attacks. STC proposes enhanced Light CNN systems with varied acoustic features and investigates angular-margin softmax and batch normalization.
- Motivation: Deployed automatic speaker-verification systems remain vulnerable to spoofing attacks, creating a need for reliable countermeasures.
- Challenge scope: ASVspoof2019 evaluates generalized countermeasures under logical access with synthesis and voice conversion, and physical access with replay attacks.
- Contribution: STC proposed systems for both scenarios using an enhanced Light CNN architecture and different acoustic features.
- Contribution: The paper explores angular-margin softmax and batch normalization for improving antispoofing-system quality.
- Scope boundary: Training and evaluation data were modeled using acoustic replay simulation, raising concerns about applicability to real-case scenarios.
2. LCNN system modifications
The modified systems use enhanced Light CNNs as final score estimators over raw acoustic representations, with batch normalization and angular-margin softmax added to improve training and discrimination. The explored inputs include spectral and cepstral features.
- LCNN architecture: All proposed systems use an enhanced Light CNN architecture with Max-Feature-Map activation for antispoofing.
- Acoustic features: The systems use raw acoustic inputs including log power spectra derived through CQT, FFT, and DCT, plus LFCC and CQCC features.
- LCNN architecture: LCNN directly estimates final scores from low-level acoustic features rather than serving as a feature extractor before GMM scoring.
- Training modifications: Batch normalization follows MaxPooling layers to increase training stability and convergence speed.
- Angular margin: Angular-margin softmax trains discriminative features on a hypersphere by imposing an angular margin between classes.
- Angular margin: The angular-margin parameter m controls the size of the margin between classes.
3. Experimental setup
Experiments train systems on the ASVspoof2019 training split, validate and adjust fusion weights on development data, and evaluate unseen attack algorithms and replay configurations. Single systems and equal-weight score fusions cover both access scenarios.
- Data partitioning: Training used only the ASVspoof2019 training split, while development data supported validation and weight adjustment for fusion.
- Data partitioning: The evaluation set contains unseen genuine trials and spoofing attacks generated with unknown algorithms and replay configurations.
- System configurations: Single systems were prepared for each scenario using the selected acoustic features and LCNN architecture.
- Logical access: Logical-access systems include LFCC-LCNN, LFCC-CMVN-LCNN, CQT-LCNN, and FFT-LCNN configurations.
- Physical access: Physical-access systems include LFCC-LCNN and CQT-LCNN configurations.
- Score fusion: Submitted primary systems fuse subsystem scores with equal weights after normalization by each system’s genuine-score standard deviation.
4. Results and Discussion
The experiments evaluate baseline modifications and proposed deep-learning systems across logical-access and physical-access scenarios using EER and min-tDCF. Angular-margin softmax and batch normalization improved LCNN performance and training stability, while evaluation results showed strong performance on unknown attacks, with limitations for some logical attacks and modeled-data mismatch.
- EER and min-tDCF were used as the primary metrics for evaluating single and fusion systems.
- Development and evaluation results confirmed the efficiency of deep-learning approaches for the ASVspoof2019 spoofing-detection tasks.
- SAD reduced LFCC- and CQCC-based system quality in both logical-access and physical-access scenarios, so it was excluded from challenge systems.The authors attribute this to discriminative spoofing or genuine-speech information in nonspeech and boundary regions.
- Angular-margin softmax improved system quality and stabilized training for both logical-access and physical-access systems.The training-process comparison is shown for PA CQT-LCNN and LA FFT-LCNN systems.
- Batch normalization and angular-margin softmax improved the original LCNN across different low-level acoustic features in both scenarios.
- High evaluation-set performance on unknown spoofing attacks demonstrated stability of the proposed approach in both evaluation conditions.
- The final LA system degraded on several unknown attack types, with A17 voice conversion using waveform filtering being the most difficult.The affected attack types were A10–A15 and A17–A18.
- PA detection performance depended on replay quality, with the highest-quality attacks recorded close to the talker using high-quality loudspeakers.Figure 3 analyzes replay-device quality, talker and ASV-system distances, and reverberation characteristics.
5. Conclusion
STC’s LCNN-based systems performed well in both ASVspoof2019 evaluation conditions, including unknown attack types, with lower EER in physical access than logical access.
- 1.86% EER was achieved in the logical access scenario for unknown attack types.
- 0.54% EER was achieved in the physical access scenario for unknown attack types.
- The submitted LCNN-based systems performed well across both logical access and physical access cases.