Diffusion-based Adversarial Purification Protecting Face Authentication Systems from Multi-task Gray-box Attacks

Deep learning technologies have achieved superior performance in security applications, e.g., the face authentication. However, increasing security concerns are raised by adversarial attacks, which uncover critical vulnerabilities of deep learning models. This thesis revisits these vulnerabilities of deep learning technologies and reveals the limitations of existing adversarial attacks and defenses. Furthermore, this thesis proposes two adversarial attacks as measurement tools in such attack scenarios that existing methods do not apply. Finally, this thesis proposes defenses against adversarial attacks to enhance the security of face authentication systems.