Spoofing-robust automatic speaker verification: Architecture, explainability and joint optimisation

This thesis explores Automatic Speaker Verification (ASV) systems and their vulnerabilities to spoofing attacks, highlighting the necessity for robust spoofing countermeasures (CMs). It introduces the application of Partially Connected Differentiable Architecture Search (PC-DARTS) for optimising network architectures for voice anti-spoofing, demonstrating competitive performance and better generalisation against unseen attacks. Further, it employs SHapley Additive exPlanations (SHAP) to analyse and visualise the impact of individual input features on detection performance, providing insights into system behaviour and attack characteristics. Lastly, it proposes an integrated spoofing-aware speaker verification system, emphasising the benefits and challenges of joint optimisation of ASV and CMs for enhanced detection capabilities and system robustness against spoofing attacks.