arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2204.04792cs.CR
收藏

差分隐私下位置轨迹的鲁棒指纹

Robust Fingerprint of Location Trajectories Under Differential Privacy

Yuzhou Jiang, Emre Yilmaz, Erman Ayday

AI总结:

针对位置数据共享中的隐私与未经授权分发问题,提出结合差分隐私、效用后处理与采样的鲁棒指纹方案,可抵御多种攻击并提升数据效用。

摘要:

直接发布这些数据会引发隐私和责任方面的担忧(例如,由于此类数据集被未经授权分发),因为位置数据包含用户的敏感信息,例如规律的移动模式和偏好地点。为解决这一问题,我们提出了一种新型指纹方案,该方案能够同时识别位置数据集的未经授权再分发,并为共享数据提供差分隐私保证。鉴于差分隐私机制导致的数据效用下降,我们引入了一种以效用为中心的后处理方案,以恢复位置轨迹中各点之间的时空相关性。我们进一步将该后处理方案作为一种采样方法集成到指纹方案中。所提出的指纹方案缓解了差分隐私机制引入的噪声对共享数据集效用造成的下降(即通过保留数据的公开已知统计信息来添加指纹)。同时,得益于差分隐私的一个基本属性——对后处理的免疫性,该方案在整个过程中不会违反差分隐私。我们提出的指纹方案对针对指纹方案的已知且经过充分研究的攻击具有鲁棒性,包括随机翻转攻击、基于相关性的翻转攻击以及多方合谋,这使攻击者难以推断指纹编码并逃避指控。通过在两个真实位置数据集和两个合成数据集上的实验,我们表明该方案实现了较高的指纹鲁棒性,并优于现有方法。此外,所提出的指纹方案提高了差分隐私数据集的数据效用,这对数据分析者有利。

英文摘要:

Directly releasing those data raises privacy and liability (e.g., due to unauthorized distribution of such datasets) concerns since location data contain users' sensitive information, e.g., regular moving patterns and favorite spots. To address this, we propose a novel fingerprinting scheme that simultaneously identifies unauthorized redistribution of location datasets and provides differential privacy guarantees for the shared data. Observing data utility degradation due to differentially-private mechanisms, we introduce a utility-focused post-processing scheme to regain spatio-temporal correlations between points in a location trajectory. We further integrate this post-processing scheme into our fingerprinting scheme as a sampling method. The proposed fingerprinting scheme alleviates the degradation in the utility of the shared dataset due to the noise introduced by differentially-private mechanisms (i.e., adds the fingerprint by preserving the publicly known statistics of the data). Meanwhile, it does not violate differential privacy throughout the entire process due to immunity to post-processing, a fundamental property of differential privacy. Our proposed fingerprinting scheme is robust against known and well-studied attacks against a fingerprinting scheme including random flipping attacks, correlation-based flipping attacks, and collusions among multiple parties, which makes it hard for the attackers to infer the fingerprint codes and avoid accusation. Via experiments on two real-life location datasets and two synthetic ones, we show that our scheme achieves high fingerprinting robustness and outperforms existing approaches. Besides, the proposed fingerprinting scheme increases data utility for differentially-private datasets, which is beneficial for data analyzers.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑