Improving Privacy-utility Tradeoffs in Privacy-preserving Data Release with Context Information

Abstract

As data collection, storage, and usage become more pervasive, there's a growing need to incorporate privacy into the data consumption pipeline. Furthermore, as data acquisition becomes increasingly personalized and tailored to specific applications, contextual knowledge about the underlying data often becomes available, such as prior distributions, data correlations, and more. Existing privacy notions, like Differential Privacy (DP) and its variants, focus on the design of privacy-preserving mechanisms without explicitly accounting for this contextual knowledge. As a result, the privacy protection mechanisms based on these notions might lead to a sub-optimal utility-privacy tradeoff. In this dissertation, we demonstrate that such contextual knowledge can be effectively leveraged to achieve higher utility while still providing rigorous privacy guarantees. We introduce Local Information Privacy (LIP), a context-aware version of Local Differential Privacy (LDP), with privacy guarantees bounded between epsilon and 2epsilon-LDP. By utilizing context, LIP significantly outperforms 2epsilon-LDP in terms of utility. We explore various LIP variants and analyze how they relate to existing privacy notions. Building on LIP and its variants, we develop privacy-preserving mechanisms, starting with discrete-valued or continuously-valued single data points. We then consider advanced mechanisms that account for uncertain data priors or incorporate encoding or hashing techniques. Furthermore, we study mechanisms for sequential data release or query answering, ensuring either sequential information privacy or perfect privacy. Through experiments with both synthetic and real data, our results show that our mechanisms achieve a much better utility-privacy tradeoff than LDP-based mechanisms.