Practical secure data analytics systems with privacy assurance

The striking progress of data analytics has catalyzed a wide spectrum of human endeavors, including biomedicine, multimedia, and transportation. Enterprises offer data analytics services to provide ready-made intelligence for end-user applications and improve their businesses. Being increasingly indispensable parts of applications in this era, these services are deployed via two primary stages in practice: 1) a training stage to learn a model over aggregated individuals' data; and 2) an inference stage to predict using an already trained model. Such practice seems appealing, yet current services operating on clear data raise acute privacy concerns, posing hurdles to their practical deployments. Individual data often carry private information and should always remain confidential. Models are deemed as intellectual properties and usually embed traces of (sensitive) training data. This thesis focuses on designing secure and practical data analytic systems that empower enterprises providing versatile analytical services without revealing either the individuals' data or their proprietary machine learning models. We present four systems resorting to advanced privacy-enhancing cryptographic techniques for secure computations ranging from data mining to neural network inference. Harnessing insights from systems, cryptography, data analytics, as well as subtly tailored secure computation protocols, our systems foster practical secure analytic services while providing stringent and provable cryptographic security guarantees. Together, our systems demonstrate promising performance that are often more lightweight and efficient compared with prior art in real-world applications like medical diagnosis and image classification.