Shapley-Value-Based Feature Attribution for Risk-Utility Tradeoff in Data Privacy

• Activity Number: 167 - Data Mining and Econometrics
• Type: Contributed
• Date/Time: Tuesday, August 10, 2021 : 10:00 AM to 11:50 AM
• Sponsor: Business and Economic Statistics Section
• Author(s): Francis Bilson Darku* and Xinxue Qu and Hong Guo
• Companies: University of Notre Dame and University of Notre Dame and University of Notre Dame
• Keywords: Data privacy; rist-utility tradeoff; Shapley-value; feature attribution

Abstract

With abundant data, companies can offer better products with the help of machine-learning and AI tools that attracts more customers, thus generating more data. While enjoying the benefits of data analytics, consumers are exposed to severe data privacy concerns. Data privacy protection has therefore become more important. To protect users’ privacy, several methods proposed in the literature typically mask or drop confidential variable(s) in a dataset to reduce disclosure risk. Considering potential intruders may use other variables present in a dataset to infer the confidential values, masking or removing the confidential variable(s) only may not be the most effective solution. To address the gap in the literature, this paper proposes a unified framework of feature attribution for data privacy protection. The proposed framework suggests a comprehensive feature attribution and selection procedures to guide data masking. It uses a Shapley-value-based feature attribution approach to fairly allocate the risk and utility of a dataset to the individual variables contained in that dataset, thereby identifying variables that pose higher risks but contribute less to utility.