DLDP-BF: A Differentiated Local Differential Privacy Bloom Filter for Membership Queries

Lei Mo, Xiujun Wang, Longkun Guo

Keywords: Bloom filter, membership query, Local Differential Privacy

TL;DR: We propose DLDP-BF, a differentiated local differential privacy Bloom filter that adapts hash functions, ensuring LDP guarantees, formal utility–privacy tradeoff, and superior performance over existing methods.

Abstract:

In privacy-preserving data processing, Bloom filters are widely used for their efficiency and scalability.
However, existing methods adopt a fixed number of hash functions for all elements, disregarding their varying importance or frequency within the dataset.
This uniform treatment leads to a suboptimal trade-off between privacy and utility, as high-priority elements, such as frequent or critical data, require more precise encoding and finely tuned privacy protection, while less significant elements can tolerate greater uncertainty without severely affecting system performance.
To address this issue, we propose a Differentiated Local Differential Privacy Bloom Filter for Membership Queries (DLDP-BF).
This method dynamically allocates hash functions based on the relative importance of elements, enabling configuration of differentiated Bloom filters.
DLDP-BF allocates more resources to high-priority elements, improving their encoding precision and reducing perturbations, thereby ensuring query accuracy for critical data.
Furthermore, we design a novel local differential privacy (LDP) budget allocation algorithm based on differentiated Bloom filters that adaptively adjusts noise intensity based on element importance.
This algorithm ensures strict privacy protection while minimizing the impact on data utility.
We construct a mathematical model linking the importance of elements and privacy budget allocation, and theoretically demonstrate that our method maintains privacy while also balancing data utility.
Experimental results show that DLDP-BF significantly improves data utility while preserving privacy. Specifically, it achieves an average reduction in RMSE of 37.1\% and an average improvement in accuracy of 9.05\%.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 2117