Nonlinear Model Combination Approach to Decentralised and Privacy-Preserving Classification

For distributed data resources, the current machine learning approaches to address the issues such as data privacy, data transfer restrictions, communication, and computation costs are complicated. To address such an issue, this paper proposes a decentralised privacy-preserving learning method based on the nonlinear model combination approach. It allows distributed sites to build classification models at global and local levels without sharing or disclosing distributed data resources or data centralisation. The proposed method restricts the exchanged information between sites with only the local models learned from the local data and therefore shares minimal information. The experiment results show better or comparable prediction accuracy to the centralised learning approach and other existing distributed machine learning approaches. Compared with other distributed learning approaches, such as federated learning which requires sharing gradient information and iterative learning, the proposed method provides an effective alternative with much less information sharing and reduced computation cost.