A decision cost-based learning approach for spoofing-aware speaker verification

Automatic Speaker Verification (ASV) systems play a significant role in biometric identity verification, offering both user-friendly and effective authentication solutions. These systems are widely utilized in various applications, including security systems, access control, and mobile device authentication, with the primary goal of identifying or verifying speaker identity. However, ASV systems remain vulnerable to spoofing attacks, thereby posing significant security risks in the absence of proper countermeasures. Specifically, attacks such as replay attacks, text-to-speech synthesis, and voice conversion threaten the reliability of speaker verification systems. To mitigate these threats, various countermeasure (CM) systems have been developed to detect and prevent spoofing attempts. In recent years, there has been a growing body of research on developing spoofing- aware speaker verification systems through the integration of ASV and CM systems. These next-generation systems enhance speaker verification by not only verifying user identity but also improving resilience against spoofing attacks. In particular, multi- stage fusion strategies, parallel architectures operating on deep feature representations, and metric-based optimization methods have been employed. In this study, state-of-the-art models such as ECAPA-TDNN, WavLM, and AASIST are integrated, and both score-level and deep embedding-level fusion techniques are employed. Additionally, a system-independent evaluation metric, the architecture- agnostic detection cost function (a-DCF), is optimized as the loss function. Comprehensive experiments conducted on the ASVspoof 2019 and ASVspoof 5 datasets demonstrate the effectiveness of the proposed methods. The results show significant reductions in error rates and enable more accurate differentiation between target, nontarget and spoof speech samples. This thesis contributes to advancing biometric security by combining cutting-edge models and optimization strategies. It represents a significant step toward enhancing the security of real-world applications. Accordingly, this work offers valuable contributions to both academic research and practical implementations in the field of speaker verification.