The rapid advancement of voice-based biometric systems, particularly Automatic Speaker Verification (ASV), has revolutionised authentication and access control across various sectors, including finance, healthcare, and security. However, this progress is accompanied by escalating challenges such as sophisticated spoofing attacks, adversarial manipulations, and ethical concerns surrounding privacy, fairness, and sustainability. This thesis presents a comprehensive exploration of these challenges, with a focus on the pioneering work of Massimiliano Todisco in enhancing the robustness, privacy, and ethical integrity of ASV systems.
Central to this research are the development of advanced spoofing countermeasures, notably the introduction of Constant Q Cepstral Coefficients (CQCCs), which significantly improve the detection of synthetic and replayed audio attacks. Building on this foundation, integrated frameworks for Presentation Attack Detection (PAD) and ASV fusion are proposed, leveraging Gaussian back-end strategies to enhance system resilience without compromising efficiency. Additionally, the thesis delves into adversarial robustness, introducing models like Malafide and Malacopula that expose vulnerabilities in existing ASV systems, thereby driving the creation of more resilient defenses.
Privacy and fairness are meticulously addressed with adversarial learning and cryptographic methods. These approaches ensure that biometric data is protected against unauthorised access and demographic biases are mitigated, promoting equitable performance across diverse user groups.
Looking forward, the thesis envisions the integration of multimodal biometrics, integration of Spiking Neural Networks (SNNs) and sensing technologies to bolster liveness detection and thwart advanced spoofing techniques. Furthermore, the research emphasises sustainable AI practices by optimising ASV systems for energy efficiency, aligning with Green AI principles. It also explores the ethical applications of deepfake technology in healthcare and education, advocating for responsible use through robust watermarking and provenance-tracking methods.
Through the establishment of benchmarks via initiatives like ASVspoof and VoicePrivacy challenges, Todisco’s work has significantly influenced both academic research and industry practices, fostering a collaborative and standardised approach to biometric security. This thesis underscores the importance of a holistic, interdisciplinary strategy that harmonises technical innovations with ethical imperatives, ensuring that biometric systems are not only secure and efficient but also fair and privacy-preserving.
