Introduction
Automatic Speaker Verification (ASV) systems are increasingly vulnerable to sophisticated spoofing attacks, including speech synthesis and voice conversion. Traditional countermeasure (CM) models typically rely on softmax-based classification. However, softmax is inherently "greedy," often producing overconfident predictions even for out-of-distribution (OOD) samples. This overconfidence is a significant security risk, as a model may confidently misclassify a novel, unseen attack as "bonafide."
To address this, we present FADEL (Fake Audio Detection with Evidential Learning), which incorporates Evidential Deep Learning (EDL) to quantify predictive uncertainty. By moving beyond point estimates to a distribution over class probabilities, FADEL allows models to "know what they don't know."
Methodology: Evidential Deep Learning
The core of FADEL is the replacement of the standard softmax layer with an evidential framework based on Subjective Logic. The authors integrated FADEL into two state-of-the-art backbone models: Res-TSSDNet and AASIST.
As shown in Figure 1, the process involves:
- Evidence Generation: The final layer of the backbone model uses a Softplus activation to produce non-negative evidence values ($e_k$) for each class.
- Dirichlet Distribution: These evidence values parameterize a Dirichlet distribution, $D(\mathbf{p}|\boldsymbol{\alpha})$, where $\alpha_k = e_k + 1$. This distribution represents the model's belief over possible class probability assignments.
- Uncertainty Quantification: Total uncertainty ($u$) is inversely proportional to the total evidence collected. High evidence leads to low uncertainty and vice versa.
- Learning Objective: The model is trained using an evidential loss function ($L_{EL}$) that balances classification accuracy with the minimization of evidence for incorrect classes, effectively regularizing the model against overconfidence.
Experimental Setup
FADEL was evaluated using the ASVspoof2019 and ASVspoof2021 Logical Access (LA) datasets. The 2019 evaluation set contains 13 unseen attack types, serving as a test for OOD robustness. The 2021 dataset provides a cross-dataset challenge, testing generalization across different recording environments and newer attack methodologies.
Results and Discussion
FADEL consistently improves performance over standard softmax baselines. When integrated with the AASIST backbone, FADEL reduced the Equal Error Rate (EER) by 18% and the minimum tandem Detection Cost Function (min t-DCF) by 27% on the ASVspoof2019 LA dataset. Ablation studies confirmed that the choice of activation function is critical, with Softplus providing the optimal balance for minimizing the min t-DCF.
Generalization and Uncertainty
A key advantage of FADEL is its superior generalization. In cross-dataset tests (trained on 2019, tested on 2021), AASIST-FADEL achieved an EER of 5.60%, outperforming specialized generalization techniques like ASAM (6.10%).
Crucially, the model's internal uncertainty serves as a reliable proxy for its error rate. As illustrated in Figure 3, there is a strong positive correlation between the average uncertainty and the EER across various spoofing algorithms (A07-A19). This validates that the evidential framework provides a physically meaningful metric of "untrustworthiness" for specific inputs.
Conclusion
FADEL addresses the overconfidence problem in fake audio detection by leveraging Evidential Deep Learning. By quantifying uncertainty, the system becomes more robust to unseen attacks and provides a transparent metric for its own reliability. Future work will explore the application of this framework to multi-modal biometric security and other forms of deepfake detection.
Summary updated