A privacy-preserving technique for Euclidean distance-based mining algorithms using Fourier-related transforms

Abstract

Privacy preserving data mining has become increasingly popular because it allows sharing of privacy-sensitive data for analysis purposes. However, existing techniques such as random perturbation do not fare well for simple yet widely used and efficient Euclidean distance-based mining algorithms. Although original data distributions can be pretty accurately reconstructed from the perturbed data, distances between individual data points are not preserved, leading to poor accuracy for the distance-based mining methods. Besides, they do not generally focus on data reduction. Other studies on secure multi-party computation often concentrate on techniques useful to very specific mining algorithms and scenarios such that they require modification of the mining algorithms and are often difficult to generalize to other mining algorithms or scenarios. This paper proposes a novel generalized approach using the well-known energy compaction power of Fourier-related transforms to hide sensitive data values and to approximately preserve Euclidean distances in centralized and distributed scenarios to a great degree of accuracy. Three algorithms to select the most important transform coefficients are presented, one for a centralized database case, the second one for a horizontally partitioned, and the third one for a vertically partitioned database case. Experimental re-

Fourier Magnitude-Based Privacy-Preserving Clustering on Time-Series Data

“…SMC-based solutions [3] provide secure mining algorithms for the distributed environment, and they are orthogonal to our PPC problem. On the other hand, distortion-based solutions [1], [6]- [8] are generally used for the centralized environment that consists of multiple data providers and one or more third parties. A simple distortion-based solution is using random data perturbation [1], [8], but it may incur bad clustering accuracy [6].…”

“…In this paper we address the problem of privacy-preserving clustering (PPC in short) on sensitive time-series data [5], [6]. Typical examples are as follows: (1) drivers do not wish to disclose their exact speed recorded in the vehicle monitoring system, but they still allow clustering of driving patterns [8]; (2) patients with heart disease do not want to disclose their private electrocardiogram (ECG) data, but they still allow clustering of patient ECG data.…”

“…PPC solutions can be classified into (1) secure multiparty computation(SMC)-based solutions [3] and (2) distortion-based solutions [1], [6], [7]. In this paper we focus on the distortion-based approach, in which the data providers distort original time-series and publish the distorted time-series to third parties.…”

“…In this paper we focus on the distortion-based approach, in which the data providers distort original time-series and publish the distorted time-series to third parties. A simple distortion-based solution is random data perturbation [1], [8], but it does not preserve distance orders and shows bad clustering accuracy [6]. Other distortion methods [6], [7] were proposed to overcome this problem, but they may cause privacy breach if their distorting techniques or parameters are publicly revealed.…”

“…To preserve as many Euclidean distance orders as possible, we present magnitude selection strategies. The first strategy, called sequential selection, is simply choosing the first few magnitudes as in the coefficient-based method [6]. We then propose two greedy strategies that select magnitudes based on the given sample time-series.…”