Predictive model training systems actually lack positive samples such as biomedical data from healthy people due to data privacy preservation issues that mHealth users face. This paper therefore presents a novel approach that preserves data privacy with the objective of incentivizing mHealth users to confidently release their biomedical data. It focuses on logistic regression, a machine learning method, and its limitations regarding data privacy preservation.

The proposed privacy scheme has been evaluated using a real-world dataset. It has good performance and ensures high privacy-preserving capability. The error rates of the models that it computes are constant, while those computed by the local approach increase as the user number increases. This approach is more secure against collusion than the native approach, a summation protocol that works as a ring in which an initiator starts the summation process. The approach uses a homomorphic approach, which is “robust against collusion attacks and ... efficient for computation over multiple iterations” and anonymized data. Test results show that the proposed “computation protocol leaks no information beyond the intermediate and final aggregated regression parameters” and information leakage during the interactions is bounded. This provided data privacy preservation scheme would surely increase user confidence in mHealth applications and thus increase the provision of positive samples for predictive model training.

The alternating direction method of multipliers (ADMM) and its limitations are discussed. Functions are formulated, though not deeply described. However, parameter descriptions are missing in the functions’ where clause. Also missing is a comparison of the distributed approaches using vertically and horizontally partitioned data with each other, as well as a comparison of private predictive model training algorithms with each other.

Aside from these weaknesses, the paper is an excellent reference for researchers working on data privacy. The authors clearly describe their concept, evaluate their approach, and point out its novelty.