kACTUS 2: Privacy Preserving in Classification Tasks Using k-Anonymity

Abstract

Abstract. k -anonymity is the method used for masking sensitive data which successfully solves the problem of re-linking of data with an external source and makes it difficult to re-identify the individual. Thus kanonymity works on a set of quasi-identifiers (public sensitive attributes), whose possible availability and linking is anticipated from external dataset, and demands that the released dataset will contain at least k records for every possible quasi-identifier value. Another aspect of k is its capability of maintaining the truthfulness of the released data (unlike other existing methods). This is achieved by generalization, a primary technique in k -anonymity. Generalization consists of generalizing attribute values and substituting them with semantically consistent but less precise values. When the substituted value doesn't preserve semantic validity the technique is called suppression which is a private case of generalization. We present a hybrid approach called compensation which is based on suppression and swapping for achieving privacy. Since swapping decreases the truthfulness of attribute values there is a tradeoff between level of swapping (information truthfulness) and suppression (information loss) incorporated in our algorithm.We use k -anonymity to explore the issue of anonymity preservation. Since we do not use generalization, we do not need a priori knowledge of attribute semantics. We investigate data anonymization in the context of classification and use tree properties to satisfy k-anonymization. Our work improves previous approaches by increasing classification accuracy.

The Privacy Issue of Counterfactual Explanations: Explanation Linkage Attacks

“…Pureness is therefore more suitable than the classiication metric to measure how often the anonymous counterfactual explanation gives us correct advice. We choose the datasets described in Table 5, as they are all tabular datasets that contain various personal attributes through which individuals could be identiied, and are often used in research about privacy-preserving data mining [26,50,51]. 13 All these datasets contain private information such as inancial and health data that people generally do not want to be made public.…”

A New Weight and Sensitivity Based Variable Maximum Distance to Average Vector Algorithm for Wearable Sensor Data Privacy Protection

“…Masoumzadeh and Joshi [10] proposed LBS (k, T)-anonymity; and Quoc and Dang [11] proposed eM 2 algorithm based on a Member Migration (MM) technique to satisfy the requirement of k-anonymity and reduce the information loss. Kisilevich et al proposed KACTUS and KACTUS2 in two studies [12], [13] respectively. The two studies lead to the query model based on k-anonymity, expanding and improving the basic anonymous models as it reduces the risk of privacy leakage and improves the effectiveness of data.…”

An Un-realized Algorithm for Effective Privacy Preservation Using Classification and Regression Trees

“…Lindell used ID3 algorithm for achieving SMC. Similarly, Kisilevich proposed a KATCUS algorithm based on ID3 decision tree approach [10]. Fong proposed an un-realization technique for sanitization of the dataset and an ID3 algorithm and its modifications are used for privacy preserving decision tree learning [11][12][13].…”