Towards scalable compliant data processing using data masking and summarization
The significant growth of personal data collection driven by data-intensive applications has raised widespread concerns about privacy and data protection. To address these concerns, many governments have enacted new data protection laws and imposed restrictions on cross-border data movement. This has led organizations to rethink their data pipelines to ensure compliance, a challenge we refer to as compliant data processing. Existing approaches to compliant data processing typically rely on data masking and summarization to prevent exposure to sensitive data while supporting processing/analytics. However, these methods face challenges with continuous data (e.g., streams), complex access requirements, or high-dimensional data, leaving many real-world applications unsupported. In this thesis, we present three novel approaches to address gaps in compliant data processing: for streaming, relational, and high-dimensional data, respectively. First, we focus on data streams and introduce Condor, a framework that facilitates the definition of synopsis-based streaming jobs under unknown query workloads. Condor combines windowing and summarization techniques, supporting approximate query answering without exposing raw data. Second, we address the problem of enabling partial access control through data masking in relational databases. We propose Mascara, a middleware for specifying and enforcing data disclosure policies via data masking. During query execution, Mascara modifies the query plan to apply the required masking operations and employs a utility estimation framework based on relative entropy to select the disclosure-compliant query that minimizes information loss. Finally, we focus on high-dimensional data processing and introduce Pacha Sketches, a novel data summarization technique for answering multidimensional count queries without direct access to raw data. Pacha Sketches distinguish between categorical and numerical dimensions to effectively organize the multidimensional space, use filtering techniques to prune uninhabited regions, and employ frequency-based sketches to track the counts of remaining regions. Overall, this thesis advances the state of the art in compliant data processing by integrating principles and techniques of data masking and summarization across streaming, relational, and high-dimensional data domains. Our evaluation demonstrates that all proposed approaches scale effectively across diverse datasets and workloads while maintaining high performance and accuracy, rendering the compliance overhead negligible.
