Efficient Machine Learning in Hierarchical Distributed Systems to Protect Privacy

Abstract

Data expansion in both quantities and scales has transformed distributed machine learning into a key instrument that completes big data tasks such as prediction, classification, etc. However, it is impossible to mix raw data from all data holders for learning purposes owing to actual physical restrictions and a potential data breach. In order to overcome this challenge, the distributed confidentiality learning techniques are used to know all distributed data without exposing any actual information. The present approaches, however, have limitations on the sophisticated distributed system. First, traditional techniques to privacy-saving learning are based on huge cryptographic basic training sets that significantly degrade the learning speed due to overhead processing. However, the complicated architecture of the system in the real system becomes a barrier. In this study, we present an effective machine-learning technology for privacy protection using hierarchical distributed networks.

We adapt and enhance the algorithm for collaborative learning. In addition to reducing the overhead for the learning process, the suggested approach ensures total security at each level of the hierarchy. We also provide an asyncronous approach to enhance the learning efficiency on the basis of an evaluation of collaborative convergence in multiple research groups of the distributed hierarchical system. Finally, comprehensive testing of real data is carried out in order to evaluate the confidentiality, effectiveness and efficiency of our suggested solutions.