A privacy preserving technique for distance-based classification with worst case privacy guarantees

There has been relatively little work on privacy preserving techniques for distance based mining. The most widely used ones are additive perturbation methods and orthogonal transform based methods. These methods concentrate on privacy protection in the average case and provide no worst case privacy guarantee. However, the lack of privacy guarantee makes it difficult to use these techniques in practice, and causes possible privacy breach under certain attacking methods. This paper proposes a novel privacy protection method for distance based mining algorithms that gives worst case privacy guarantees and protects the data against correlation-based and transform-based attacks. This method has the following three novel aspects. First, this method uses a framework to provide theoretical bound of privacy breach in the worst case. This framework provides easy to check conditions that one can determine whether a method provides worst case guarantee. A quick examination shows that special types of noise such as Laplace noise provide worst case guarantee, while most existing methods such as adding normal or uniform noise, as well as random projection method do not provide worst case guarantee. Second, the proposed method combines the favorable features of additive perturbation and orthogonal transform methods. It uses principal component analysis to decorrelate the data and thus guards against attacks based on data correlations. It then adds Laplace noise to guard against attacks that can recover the PCA transform. Third, the proposed method improves accuracy of one of the popular distance-based classification algorithms: K-nearest neighbor classification, by taking into account the degree of distance distortion introduced by sanitization. Extensive experiments demonstrate the effectiveness of the proposed method.     

A generic and distributed privacy preserving classification method with a worst-case privacy guarantee

It is often necessary for organizations to perform data mining tasks collaboratively without giving up their own data. This necessity has led to the development of privacy preserving distributed data mining. Several protocols exist which deal with data mining methods in a distributed scenario but most of these methods handle a single data mining task. Therefore, if the participating parties are interested in more than one classification methods they will have to go through a series of distributed protocols every time, thus increasing the overhead substantially. A second significant drawback with existing methods is that they are often quite expensive due to the use of encryption operations. In this paper a method has been proposed that addresses both these issues and provides a generic approach to efficient privacy preserving classification analysis in a distributed setting with a worst-case privacy guarantee. The experimental results demonstrate the effectiveness of this method.