Graduation Year
2020
Document Type
Dissertation
Degree
Ph.D.
Degree Name
Doctor of Philosophy (Ph.D.)
Degree Granting Department
Computer Science and Engineering
Major Professor
Srinivas Katkoori, Ph.D.
Committee Member
Sriram Chellapan, Ph.D.
Committee Member
Robert Karam, Ph.D.
Committee Member
Morris Chang, Ph.D.
Committee Member
Ramachandran Kandethody, Ph.D.
Keywords
Constrained Platforms, Hardware Security, Internet of Medical Things, Machine Learning
Abstract
Artificial intelligence and ubiquitous sensor systems have seen tremendous advances in recent times, resulting in groundbreaking impact across domains such as healthcare, entertainment, and transportation through a collective ecosystem called the Internet of Things. The advent of 5G and improved wireless networks will further accelerate the research and development of tools in deep learning, sensor systems, and computing platforms by providing improved network latency and bandwidth. While tremendous progress has been made in the Internet of Things, current work has largely focused on building robust applications that leverage the data collected through ubiquitous sensor nodes to provide actionable rules and patterns. Such frameworks do not inherently take into account the issues that come with scale such as privacy, the security of the data, and the ability to provide a completely immersive experience. This is particularly significant since, due to the somewhat limited scope of computing resources, the IoT edge nodes themselves do not process the observed data. Instead, they transmit the collected data to more powerful servers for processing. This information transmission can place strain on the network while introducing security concerns such as eavesdropping and man-in-the-middle attacks.
In this dissertation, we address these concerns by using machine learning as a disruptive tool by developing privacy-aware, lightweight algorithms while evaluating the feasibility of hardware security primitives such as physical unclonable functions (PUFs) for IoT node security. To be specific, we develop unsupervised algorithms for continuous activity monitoring from ubiquitous sensors without any labeled data, which forms the first step to a decentralized learning paradigm. The proposed framework is inherently privacy-preserving by limiting the amount of data transmitted through the network while providing real-time feedback. Second, we analyze the properties of different deep learning approaches with respect to power consumption, memory footprint, and latency and provide design-time optimizations to enable implementation on compute-constrained platforms. Finally, we evaluate the feasibility of using PUF-based authentication for IoT edge nodes by exploring their susceptibility to machine learning attacks. We show that strong PUF architectures are susceptible to a non-invasive machine learning-based cloning attack. We also propose a probabilistic, discriminator model to bolster the security of the PUF-based authentication protocol by identifying possible instances of cloning attacks and bolstering the PUF-based authentication. Combined, these approaches offer a way forward for the development of an IoT framework for continuous activity monitoring that can scale to millions of nodes while ensuring the privacy and security of the observed data. We show that the proposed activity monitoring algorithm can effectively recognize and segment activities from streaming data without any labeled data on constrained platforms with close-to-real-time latency. Our design-time improvements for deep learning algorithms can yield up to 11x reduction in power consumption compared to another FPGA platform with 96x more memory capacity while maintaining the state-of-the-art classification accuracy. Through extensive experiments, we show that strong PUF architectures can be successfully cloned, including those encrypted using two different encryption protocols in DES and AES and with varying degrees of obfuscation. The proposed discriminator can distinguish cloned PUF devices and authentic PUFs with an average accuracy of 96.01% and can be used for rapidly authenticating millions of IoT nodes remotely from the cloud server.
Scholar Commons Citation
Laguduva Ramnath, Vishalini, "Machine Learning for the Internet of Things: Applications, Implementation, and Security" (2020). USF Tampa Graduate Theses and Dissertations.
https://digitalcommons.usf.edu/etd/8240