MICF: An effective sanitization algorithm for hiding sensitive patterns on data mining

Abstract

Data mining mechanisms have widely been applied in various businesses and manufacturing companies across many industry sectors. Sharing data or sharing mined rules has become a trend among business partnerships, as it is perceived to be a mutually benefit way of increasing productivity for all parties involved. Nevertheless, this has also increased the risk of unexpected information leaks when releasing data. To conceal restrictive itemsets (patterns) contained in the source database, a sanitization process transforms the source database into a released database that the counterpart cannot extract sensitive rules from. The transformed result also conceals non-restrictive information as an unwanted event, called a side effect or the “misses cost”. The problem of finding an optimal sanitization method, which conceals all restrictive itemsets but minimizes the misses cost, is NP-hard. To address this challenging problem, this study proposes the maximum item conflict first (MICF) algorithm. Experimental results demonstrate that the proposed method is effective, has a low sanitization rate, and can generally achieve a significantly lower misses cost than those achieved by the MinFIA, MaxFIA, IGA and Algo2b methods in several real and artificial datasets.

Introduction

Data mining is an interdisciplinary field bringing together techniques to extract information from large databases [1]. Over the past few years, data mining mechanisms have widely been applied in various business organizations (for example, retail, insurance, finance, banking, and communication) and manufacturing companies such as Texas Instruments (fault diagnosis), Ford (harshness, noise, and vibration analysis), Motorola (CDMA Base Station Placement), Boeing (Post-flight Diagnostics), and Kodak (data visualization) [2].

Large companies use powerful data acquisition systems (such as minicomputers, microprocessors, and senor networks) to collect, analyze, and transfer data. The role of knowledge discovery in databases (KDD) and data mining methodologies, therefore, has become extremely important for extracting useful knowledge from huge amounts of raw data [2], [3].

For design procedure capture, which is a specific engineering knowledge that can be captured from the design events monitored during design process [4], data mining techniques can be used in different stages of production. Manufacturing is typically a controlled process such as wafer fabrication. Using mining tools to analyze the collected data can result in efficient strategies to improve the production process, find out the unusual steps during the manufacturing process [5], and improve the reliability of system identification [6].

Data mining is an evolutionary step along the path of problem solving through data analysis [1]. Releasing collected data or mined rules for sharing has become a crucial trend among business partnerships as it results in increased productivity for all companies involved. Nevertheless, the released data have also increased the risk of incurring sensitive information leaks [7]. Organizations should evaluate and decrease the risk of disclosing information. Therefore, developing efficient approaches to maintaining an organization’s competitive edge in business by restricting information leakages has become an important issue. Consider the following two scenarios:

• Some paper manufacturers have their own databases that record their patterns of stock and sale. For their mutual benefit, multiple companies decide to share their databases to cooperatively generate information and trends found in the shared large database. However, each company prefers to keep their own strategic patterns confidential from the others. Thus, a company can both uncover more interesting trends from the combined shared database than that available only from their own database, and can prevent sensitive information from being discovered by other companies [8].

• The captured design procedure knowledge helps companies to understand how experienced designers carry out their designs and guide other designers to design better. Moreover, the knowledge can be used for training novice designers so that they can quickly learn how to execute prominent designs [4]. In order to preserve strategic or sensitive intelligence and still share such knowledge among allied companies, a data sanitization process or privacy-preserving techniques must be applied.

An intelligent system can be developed to achieve this. Moreover, the privacy-preserving data mining will be one of the key techniques in such a system. The mechanisms usually transform the source database into a new one from which sensitive information cannot be extracted. The procedure of transforming the source database into a new database that hides some sensitive rules is called the sanitization process [9].

In the association rule mining, all rules are derived from the frequent itemsets (patterns). Accordingly, one essential and efficient way to protect some restrictive patterns is to decrease their support values, which can be done by deleting or modifying items in several transactions. Such approaches usually follow two restrictions: (1) reduce the impact on the source database; and (2) find an appropriate balance between privacy and knowledge discovery. An itemset which must be concealed is called a restrictive itemset. The impact on a source database of deleting items from transactions can be measured by the sanitization rate. The sanitization rate is defined as the ratio of deleted items to the total support values of restrictive itemsets in the source dataset. The sanitization process can also conceal some non-restrictive itemsets, which is a side effect called the “misses cost”. An optimal sanitization process, which both conceals all restrictive patterns and minimize the misses cost, is an NP-hard problem [7].

The item grouping algorithm (IGA) has been proposed to enforce privacy in mining frequent itemsets [10]. IGA groups restrictive itemsets and assigns a victim item to each group. This approach can reduce the impact on the database if the sanitized transaction contains more than one restrictive itemset.

IGA has a low misses cost, but can be improved further by reducing the number of deleted items. Moreover, it must deal with the overlap between groups. To deal with this problem, this study proposes a novel algorithm called maximum item conflict first (MICF). The degree of conflict of an item in a sensitive transaction is defined as the number of restrictive patterns affected when an item is deleted. MICF selects an item with the maximum degree of conflict for deletion. Therefore, MICF simultaneously decreases the support values for the maximum number of restrictive patterns and reduces the number of items to be removed from the source database.

This study focuses on the task of deleting items from transactions to conceal frequent itemsets in association rule mining and on the issue of no hiding failure. All association rules derivable from these hidden frequent itemsets are thus also hidden. That is, no extra artificial itemset can be generated from the sanitized dataset.

The rest of this paper is organized as follows. Section 2 presents an overview of the current methods for solving the problem of privacy-preserving association rule mining. Section 3 explains the proposed maximum item conflict first (MICF) algorithm for hiding all restrictive itemsets. The time complexity analysis is represented in Section 4. Section 5 provides experimental results and evaluates the performance of the proposed algorithm. Finally, we conclude in Section 6 with a summary of our work.