UC Santa Barbara

Privacy considerations have become increasingly important for cloud-based information services. There are significant research challenges in top-k document search over outsourced large-scale datasets. It is because letting a cloud server access ranking features and perform advanced scoring computation may unsafely reveal privacy-sensitive information. With a practical restriction towards fast query response time, the heavy-weight cryptographic tools are often too expensive to deploy, and thus a server-hosted search system needs to seek optimized tradeoffs among privacy, efficiency, and relevance.

In this dissertation, a series of efficiency optimized document retrieval solutions is proposed with additive ranking or neural ranking when privacy protection is considered. We firstly introduce an efficiency-enhancing design that obfuscates the access pattern of the inverted index data during query processing in a trusted execution environment (TEE). Then this dissertation presents our work on ORAM-based top-k document retrieval with additive ranking in a TEE, and discusses techniques to accelerate matching with window navigation based index pruning and path caching. Finally, we discuss a privacy-aware neural ranking method with analytic and experimental studies. This dissertation includes evaluation results with TREC datasets on the efficiency and relevance of our proposed schemes against multiple baselines.