綜合而言，本論文將可信生物特徵辨識重新定義為一個統一框架，其中穩健性、安全性和隱私保護與辨識精度並重，共同構成系統設計的核心目標。所提出的解決方案表明，可靠性源於整體性設計——即在複雜環境下穩定運行、抵禦對抗性攻擊，並有效保障用戶身份資訊。這些研究成果為面向實際部署、值得用戶信賴的生物特徵辨識系統的發展奠定了堅實基礎。;Biometric technologies are central to digital identity， yet their adoption is impeded by environmental variability， security threats， and privacy risks—requiring systems that are accurate， robust， secure， and privacy-preserving. This dissertation addresses these pillars across iris， fingerprint， and voice modalities.
For iris recognition， a robust segmentation framework using an Interleaved Residual U-Net (IRUNet) accurately localizes boundaries under non-cooperative conditions involving occlusions， reflections， and noise. To address data scarcity， a synthesis-driven self-supervised learning approach integrates conditional GANs with parametric mask generation， enabling scalable augmentation and improved generalization. Experiments confirm state-of-the-art segmentation and system reliability.
For fingerprint security， advanced learning strategies are explored to strengthen presentation attack detection， with a focus on generalization to unseen spoof materials and evolving acquisition conditions. Findings underscore the necessity of adaptive and generalizable defenses to preserve system integrity against emerging threats.
For voice privacy， anonymization and speaker de-identification techniques are investigated， balancing identity protection with intelligibility， naturalness， and utility. The analysis emphasizes the value of user-controllable mechanisms for real-world privacy-aware applications.
Collectively， this work frames trustworthy biometrics as a unified paradigm in which robustness， security， and privacy are co-essential with accuracy. Reliability， as shown， arises from holistic design—enabling adverse-condition operation， adversarial resilience， and strong identity safeguarding. These contributions advance deployable biometric systems that merit and sustain user trust.
