In recent years, the collection of user data has significantly increased. Hence, the secure and efficient utilization of such data has become a crucial concern. Two significant methodologies, namely private set intersection (PSI) and privacy-preserving machine learning (PPML), have emerged as effective means to mitigate privacy breaches. PSI enables multiple parties with mutual distrust to securely compute the intersection of their respective sets without unveiling any information beyond the intersection. Simultaneously, PPML protects sensitive data throughout the machine learning lifecycle, employing strategies such as homomorphic encryption, differential privacy and local differential privacy. This study introduces an advanced PPML algorithm that leverages the Scalable Unified Privacy Mechanism (SUPM), an innovative PPML architecture rooted in local differential privacy. Our algorithm harnesses the odds ratio for efficient dimension reduction and judicious allocation of the privacy budget, thereby achieving superior performance when compared to the current algorithm. Furthermore, we evaluate the performance of the proposed algorithm on two datasetsBreast Cancer Wisconsin (Diagnostic) and Ionosphereto demonstrate its superiority in classification tasks over existing approaches. The evaluation is conducted using logistic regression and support vector machine, employing metrics including accuracy, precision, recall, and the F1 score.

Keyword: privacy-preserving machine learning, local differential privacy, privacy budget