Privacy and Security Preserving Distributed Machine Learning and Data Storage in Large-Scale Internet of Things

Abstract

With the rapid growth of the Internet of Things (IoT) devices, managing and processing the enormous volume of distributed IoT data has become a significant challenge. To improve efficiency while maintaining trust and security in distributed IoT data storage, this thesis first presents a framework named MapChain-D, which includes two mapped blockchains located at different groups of devices for data storage and indexing.

After collecting and storing the IoT data, ML-based methods can be used to process the data and explore more insightful patterns of the data collected. Distributed Machine Learning (DML) frameworks have been proposed to train a shared model collaboratively by multiple participants to improve privacy and efficiency. However, trust issues have become a concern in DML because participants might maliciously modify the DML models. To mitigate model integrity attacks, this thesis then presents a model verification technique named VHFL, which verifies the model integrity in a hierarchical DML network.

Recent research has also highlighted potential privacy concerns from the trained model parameters shared in DML. Existing DML frameworks do not allow clients to set their personalised privacy levels based on their preferences for model sharing. To overcome this challenge, we proposed two personalised model sharing techniques for DML, named FlexSplit and FlexMask, aiming to provide clients with flexible options to participate in DML. Specifically, FlexSplit allows each client to select a layer to be shared with the edge server. In contrast, FlexMask enables each client to determine a personalised mask that selects different neurons in a layer to hide a specific attribute.

In a DML network, different clients might also be interested in various training tasks. The final part of this thesis presents a personalised neuron selection technique, named FlexNS, for each DML client to leverage a pre-trained model to train a personalised model using private data.