| Summary: | With the rapid development of sensing and communication technologies, the Internet of Things (IoT) is becoming a global data generation infrastructure. To utilize the massive data generated by IoT for achieving better system intelligence, machine learning and inference on the IoT data at the edge and core (i.e., cloud) of the IoT are needed. However, the pervasive data collection and processing engender various privacy concerns. While various privacy preservation mechanisms have been proposed in the context of cloud computing, they may be ill-suited for IoT due to the resource constraints at the IoT edge. This thesis primarily studies data obfuscation as a lightweight method to preserve data privacy for cloud-based collaborative machine learning and inference in IoT. Specifically, it presents three approaches: the first is for cascadable, collusion-resilient, and privacy-preserving cloud-based inference and the other two are for privacy-preserving collaborative training of a deep neural network based on distributed IoT data. All approaches protect the privacy contained in the data contributed by distributed IoT devices as the participants against a semi-honest coordinator in the centralized cloud system. They deliver different privacy protection properties. The first approach protects the raw forms and certain privacy attributes of the inference data from the participants by applying participant-specific obfuscation neural networks; the second approach protects the raw forms of the training data contributed by the participants by applying multiplicative random projection; the third approach protects the differential privacy of the contributed training data via additive perturbation. These three approaches are computationally lightweight and can be executed by resource-limited edge devices including smartphones and even mote-class sensor nodes. Extensive performance evaluations performed on multiple datasets and real implementations on IoT hardware platforms show the effectiveness and efficiency of these approaches in protecting data privacy while maintaining the learning and inference performance.
|
|---|