Privacy-Preserving Boosting in the Local Setting

• Published in: IEEE transactions on information forensics and security Vol. 16; pp. 4451 - 4465
• Main Authors: Wang, Sen, Chang, J. Morris
• Format: Journal Article
• Language: English
• Published: New York IEEE 2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Boosting; Classifiers; Data mining; Data models; Decision trees; Differential privacy; Ensemble learning; local differential privacy; Machine learning; Machine learning algorithms; Object recognition; Pattern recognition; Prediction algorithms; Privacy
• ISSN: 1556-6013; 1556-6021
• DOI: 10.1109/TIFS.2021.3097822

In machine learning, boosting is one of the most popular methods that is designed to combine multiple base learners into a superior one. The well-known Boosted Decision Tree classifier has been widely adopted in data mining and pattern recognition. With the emerging challenge in privacy, the personal images, browsing history, and financial reports, which are held by individuals and entities are more likely to contain sensitive information. The privacy concern is intensified when the data leaves the hand of owners and is used for further mining. Such privacy issues demand that the machine learning algorithms should be privacy-aware. Recently, Local Differential Privacy has been proposed as an effective privacy protection approach, which allows data owners to perturb the data before any release. In this paper, we propose a distributed privacy-preserving boosting algorithm that can be applied to various types of classifiers. By adopting LDP as a building block, the proposed boosting algorithm leverages the aggregation of the perturbed data shares to build the base learner, which ensures that privacy is well preserved for the participated data owners. Our experiments demonstrate that the proposed algorithm effectively boosts various classifiers and the boosted classifiers maintain a high utility.