![[2602.18910] SLDP: Semi-Local Differential Privacy for Density-Adaptive Analytics](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.18910] SLDP: Semi-Local Differential Privacy for Density-Adaptive Analytics
Summary
The paper introduces Semi-Local Differential Privacy (SLDP), a framework that enhances privacy-preserving analytics by decoupling privacy costs from iterative refinement, allowing for high-resolution data analysis.
Why It Matters
SLDP addresses significant challenges in local differential privacy by enabling more effective and efficient data analysis without compromising user privacy. This advancement is crucial for fields requiring high utility in data analytics while ensuring compliance with privacy regulations.
Key Takeaways
- SLDP provides a novel approach to local differential privacy by assigning privacy regions based on user density.
- The framework allows for high-resolution analytics without incurring additional privacy costs during refinement.
- Experimental results demonstrate SLDP's effectiveness on both synthetic and real-world datasets.
Computer Science > Machine Learning arXiv:2602.18910 (cs) [Submitted on 21 Feb 2026] Title:SLDP: Semi-Local Differential Privacy for Density-Adaptive Analytics Authors:Alexey Kroshnin, Alexandra Suvorikova View a PDF of the paper titled SLDP: Semi-Local Differential Privacy for Density-Adaptive Analytics, by Alexey Kroshnin and Alexandra Suvorikova View PDF HTML (experimental) Abstract:Density-adaptive domain discretization is essential for high-utility privacy-preserving analytics but remains challenging under Local Differential Privacy (LDP) due to the privacy-budget costs associated with iterative refinement. We propose a novel framework, Semi-Local Differential Privacy (SLDP), that assigns a privacy region to each user based on local density and defines adjacency by the potential movement of a point within its privacy region. We present an interactive $(\varepsilon, \delta)$-SLDP protocol, orchestrated by an honest-but-curious server over a public channel, to estimate these regions privately. Crucially, our framework decouples the privacy cost from the number of refinement iterations, allowing for high-resolution grids without additional privacy budget cost. We experimentally demonstrate the framework's effectiveness on estimation tasks across synthetic and real-world datasets. Subjects: Machine Learning (cs.LG) MSC classes: 68P27 Cite as: arXiv:2602.18910 [cs.LG] (or arXiv:2602.18910v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2602.18910 Focus to le...
