Local Differentially Private Fuzzy Counting in Stream Data Using Probabilistic Data Structures

Privacy-preserving estimation of counts items in streaming data finds applications several real-world scenarios including word auto-correction and traffic management applications. Recent works RAPPOR [1] Apple's count-mean sketch (CMS) algorithm [2] propose privacy preserving mechanisms for count large volumes using probabilistic structures like counting Bloom filter CMS. However, these existing methods fall short providing a sound solution real-time Since the size structure is not adaptive to volume data, utility (accuracy estimate) can suffer over time due increased false positive rates. Further, lookup operation needs be highly efficient answer estimate queries real-time. More importantly, local Differential used approaches provide guarantees come at cost (impacting accuracy estimation). In this work, we novel (local) Differentially private mechanism that provides high problem with similar or even lower budgets while providing: a) fuzzy report related (for instance account typing errors variations), b) improved querying efficiency reduce response counts. Our algorithm uses combination two Cuckoo filter. We formal proofs present extensive experimental evaluation our real synthetic English words datasets both exact scenarios. substantially outperforms prior work terms time, significantly higher estimation) under guarantees, communication overhead.