Self-Reliance for the Internet of Things: Blockchains and Deep Learning on Low-Power IoT Devices

Abstract: The rise of the Internet of Things (IoT) has transformed common embedded devices from isolated objects to interconnected devices, allowing multiple applications for smart cities, smart logistics, and digital health, to name but a few. These Internet-enabled embedded devices have sensors and actuators interacting in the real world. The IoT interactions produce an enormous amount of data typically stored on cloud services due to the resource limitations of IoT devices. These limitations have made IoT applications highly dependent on cloud services. However, cloud services face several challenges, especially in terms of communication, energy, scalability, and transparency regarding their information storage. In this thesis, we study how to enable the next generation of IoT systems with transaction automation and machine learning capabilities with a reduced reliance on cloud communication. To achieve this, we look into architectures and algorithms for data provenance, automation, and machine learning that are conventionally running on powerful high-end devices. We redesign and tailor these architectures and algorithms to low-power IoT, balancing the computational, energy, and memory requirements. The thesis is divided into three parts: Part I presents an overview of the thesis and states four research questions addressed in later chapters. Part II investigates and demonstrates the feasibility of data provenance and transaction automation with blockchains and smart contracts on IoT devices. Part III investigates and demonstrates the feasibility of deep learning on low-power IoT devices. We provide experimental results for all high-level proposed architectures and methods. Our results show that algorithms of high-end cloud nodes can be tailored to IoT devices, and we quantify the main trade-offs in terms of memory, computation, and energy consumption.