UC San Diego

Industrial Internet of Things (IIoT) is an adaptation of traditional IoT for industrial environments enabling full automation, remote monitoring, and smart maintenance. Predictive analytics aims to utilize the collected system data and provide meaningful insight into business decisions using Machine Learning (ML). However, there are numerous challenges that should be addressed to fully benefit from IIoT predictive analytics. These include performance, securtiy, and efficiency of ML algorithms. This dissertation provides solutions to each of these challenges.

Dynamic IIoT settings and conditions can significantly impact individual ML prediction performance. As a solution to this problem, ensemble learning systematically combines multiple ML methods to increase prediction performance and robustness. However, in order to deploy ensemble learning solutions in IIoT systems, additional training overhead and learning ability under limited supervision should be addressed. To address the first challenge, we propose a diversity-induced optimally-weighted ensemble learner. Our solution provides 39.2% faster retraining compared to only accuracy included ensemble with 3.4% loss in accuracy. To solve the second problem, we devise a novel few shot ensemble learning framework. It results in up to 16.4% improvement over the best algorithm by only using 0.3% of the training data.

IIoT security is challenging due to its increased inter-connectivity, small scale devices, and large attack surface. Among various cyberattacks, adversarial attacks craft perturbed examples to affect ML prediction performance. These attacks can cause serious outcomes on IIoT systems, yet they are not carefully addressed in the IIoT domain. To fill this research gap, we develop (1) a stacking ensemble learning-based framework that stays resilient against various adversarial attacks, and (2) diversity promoting ensemble adversarial training approach as a defense mechanism. Our stacking ensemble improves resiliency against adversarial attacks by up to 60% compared to the most resilient single ML method. Furthermore, our defense improves the resiliency by up to 97% compared to state-of-the-art training settings.

IIoT systems need efficient and robust learning solutions due to their resource constrained devices and the potential for noise and variability. Hyper-dimensional computing (HDC) is a brain-inspired learning solution which is shown to be efficient, accurate, and robust. We first apply HDC for predictive analytics solution and test it against various adversarial attacks. We observe that HDC has up to 67.5% higher resiliency compared to the state-of-the-art deep learning methods while being up to 25.1× faster to train. Although we showed that HDC is more resilient than DL methods, there is still a need to further investigate its resiliency against adversarial attacks to use HDC for predictive analytics safely. For that purpose, we design a novel HDC adversarial attack. Our approach improves attack success rate by up to 36%, and F1 score by up to 61% compared to the most effective state-of-the-art single adversarial attack.