Mobile Privacy in Machine Learning

A broad spectrum of mobile privacy problems from different modalities exists in various applications, including malware attack (e.g., information leakage through malicious apps), mobile authorization/identification (e.g., unauthorization or fraudulent activities detection), private data communication (e.g., local mobile photo labeling), and private model updating (e.g., local mobile AI model updating). The capability to model problems from these heterogeneous scenarios is potentially transformative for investigating privacy mechanisms and establishing a protection system. In this thesis, we introduce the different solutions for mobile privacy applications in the various scenarios which generally interested in using machine learning for mobile privacy and protecting mobile privacy while using machine learning on private data. In particular, our works involve five mobile privacy problems in different applications: malware detection, mobile authorization/identification, private knowledge transfer, and private model updating.