Privacy-Preserving and Robust Data Analytics under Distributed Settings

Abstract

In the face of exponential data growth and stringent privacy regulations, safeguarding sensitive information within data processing and analytics workflows, especially in distributed systems, has become paramount. High-profile data breaches and privacy regulations like the General Data Protection Regulation (GDPR) underscore this urgency. Privacy-enhancing data release and learning paradigms such as Local Differential Privacy (LDP) and Federated Learning (FL) offer promising solutions, but they often grapple with maintaining utility while ensuring robust privacy guarantees, particularly when dealing with complex data types and adversarial attacks. This dissertation tackles the critical need for privacy-preserving data analytics in distributed systems, focusing on LDP and FL settings. Key contributions include a novel LDP-based framework for key-value data collection that boosts utility by exploiting correlations between keys and values, featuring two mechanisms (PCKV-UE, PCKV-GRR) that optimize perturbation and sampling. We also develop a method for multi-dimensional data aggregation that reduces noise through attribute correlations, enhancing query accuracy without sacrificing privacy. Additionally, we introduce a mechanism for privacy-preserving range queries and frequency estimation under local d-privacy with improved utility. We then propose a new privacy notion called Input-Discriminative LDP (ID-LDP), which tailors privacy protections to individual input sensitivity, further enhancing utility. For FL, we present DP-BREM and DP-BREM+, ensuring both differential privacy and Byzantine robustness via client momentum and secure aggregation. Extensive experiments on synthetic and real-world data validate the superior privacy, utility, and robustness of our methods compared to state-of-the-art approaches.