Robust privacy-preserving KNN for smart healthcare with participant dropout resilience

Abstract

The $k$-nearest neighbor (KNN) algorithm, as a simple and effective machine learning method, has been widely used in smart healthcare for disease diagnosis and drug recommendation. However, with the continuous generation and use of personal health data, KNN algorithms face data privacy challenges in smart healthcare systems. To address these challenges, numerous privacy-preserving KNN schemes have been put forward, mostly using secure multi-party computation (SMPC) or differential privacy techniques. Nevertheless, these approaches often concentrate on two-party models and lead to substantial computational overhead or compromise the accuracy of model training/prediction. In this article, we present a three-party privacy-preserving KNN scheme with a privileged party. We employ vector space secret sharing (VSSS) and additive secret sharing to devise a suite of lightweight sub-protocols for implementing the crucial operations in KNN: distance evaluation and ascending sorting. Additionally, the scheme ensures robustness by leveraging the access control structure of VSSS. Concretely, this solution allows two auxiliary parties to collude, and the model prediction task remains achievable even if one of the auxiliary parties becomes unavailable. We also analyze the communication and computational overhead of proposed algorithms. Furthermore, we conduct extensive experiments to evaluate the performance of our scheme on common benchmark datasets Breast Cancer Wisconsin (Diagnostic) and MNIST. The results demonstrate our scheme’s performance outperforms most of the compared schemes.