Secure deep neural networks in adversarial settings

Artificial intelligence (AI) has been introduced in almost every aspect of our lives to assist in smart decision making. As a key technology, it is fundamentally changing the way people work and live. The thesis studies the security problems of AI models. The research exposes the attack surface of modern AI applications and proposes defences to mitigate the threats. It is a critical step in building trustworthy and resilient AI applications that can truly benefit the nation.