Region-Aware Optimization of Hierarchical Bloom Filters for Privacy-Preserving Face Membership Verification

Face recognition systems typically store biometric templates that remain closely related to the original facial data. This creates privacy concerns because, unlike passwords, compromised biometric information cannot be replaced. Hierarchical Bloom Filters (HBFs) provide a privacy-preserving alternative by converting facial information into binary templates and storing only keyed hash-based representations. Matching is performed through membership testing, producing a similarity score without requiring storage of the original image or raw biometric template.

This thesis investigates how the internal structure of an HBF can be improved for face-based membership testing, with particular emphasis on whether facial regions provide a more effective organization than a uniform hierarchical layout. A series of controlled experiments examines the effects of Bloom filter parameters, hierarchy levels, template information, and the distribution of identity-related information across facial regions. Based on these findings, a compact region-based HBF design is proposed in which informative facial locations are selected according to their measured utility.

Two implementations of the proposed design are evaluated using different facial feature representations while maintaining the same HBF backend and storage constraints. Across five face datasets, the proposed region-based designs improve membership performance over the original HBF structure while substantially reducing template and database storage. The learned-feature variant also demonstrates improved robustness to variations in scale, pose, illumination, and expression.