Date of Graduation


Document Type


Degree Name

Doctor of Philosophy in Computer Science (PhD)

Degree Level



Computer Science & Computer Engineering


Qinghua Li

Committee Member

Xintao Wu

Second Committee Member

Jingxian Wu

Third Committee Member

Brajendra Nath Panda


Cloud-Assisted Systems, machine learning, mathematical optimization, data privacy, user data, Geometric Programming (GP), threat model, chosen plaintext attack (CPA), gradient projection method (GPM), logistic regression, neural network models, fair exchange, heterogeneous architectures


Nowadays industries are collecting a massive and exponentially growing amount of data that can be utilized to extract useful insights for improving various aspects of our life. Data analytics (e.g., via the use of machine learning) has been extensively applied to make important decisions in various real world applications. However, it is challenging for resource-limited clients to analyze their data in an efficient way when its scale is large. Additionally, the data resources are increasingly distributed among different owners. Nonetheless, users' data may contain private information that needs to be protected.

Cloud computing has become more and more popular in both academia and industry communities. By pooling infrastructure and servers together, it can offer virtually unlimited resources easily accessible via the Internet. Various services could be provided by cloud platforms including machine learning and data analytics.

The goal of this dissertation is to develop privacy-preserving cloud-assisted data analytics solutions to address the aforementioned challenges, leveraging the powerful and easy-to-access cloud. In particular, we propose the following systems.

To address the problem of limited computation power at user and the need of privacy protection in data analytics, we consider geometric programming (GP) in data analytics, and design a secure, efficient, and verifiable outsourcing protocol for GP. Our protocol consists of a transform scheme that converts GP to DGP, a transform scheme with computationally indistinguishability, and an efficient scheme to solve the transformed DGP at the cloud side with result verification. Evaluation results show that the proposed secure outsourcing protocol can achieve significant time savings for users.

To address the problem of limited data at individual users, we propose two distributed learning systems such that users can collaboratively train machine learning models without losing privacy. The first one is a differentially private framework to train logistic regression models with distributed data sources. We employ the relevance between input data features and the model output to significantly improve the learning accuracy. Moreover, we adopt an evaluation data set at the cloud side to suppress low-quality data sources and propose a differentially private mechanism to protect user's data quality privacy. Experimental results show that the proposed framework can achieve high utility with low quality data, and strong privacy guarantee.

The second one is an efficient privacy-preserving federated learning system that enables multiple edge users to collaboratively train their models without revealing dataset. To reduce the communication overhead, we select well-aligned and large-enough magnitude gradients for uploading which leads to quick convergence. To minimize the noise added and improve model utility, each user only adds a small amount of noise to his selected gradients, encrypts the noise gradients before uploading, and the cloud server will only get the aggregate gradients that contain enough noise to achieve differential privacy. Evaluation results show that the proposed system can achieve high accuracy, low communication overhead, and strong privacy guarantee.

In future work, we plan to design a privacy-preserving data analytics with fair exchange, which ensures the payment fairness. We will also consider designing distributed learning systems with heterogeneous architectures.