On Design and Machine Learning Resiliency of Memristor- and eFlash-Memory-Based Strong Physical Unclonable Functions
The emergence of the Internet of Things (IoT) has enabled an unprecedented expansion of interconnected networks and devices over which a huge amount of personal and/or sensitive data is carried. As a result, privacy and security issues are among the most significant challenges in designing IoT devices. These challenges can hardly be addressed using conventional cryptographic approaches because they rely on storing secret keys in memories, which not only are vulnerable to physical and side-channel attacks but also consume huge area and vast amounts of power.
Hardware-based security approaches such as physical unclonable functions (PUFs) have attracted considerable attention as replacements for conventional methods. PUFs are well suited to a wide spectrum of security applications including key generation and authentication because they generate secure keys on the fly (rather than explicitly storing any security-critical information). This is achieved by utilizing electronic devices that entail inherent sources of randomness, which in turn help create unique keys for different physical entities.
Recently, a variety of emerging nano-scale non-volatile memories are being explored for use in the design of PUFs including memristors and embedded flash (eFlash) memories. The highly non-linear current-voltage characteristics and the inherent process variations of these memory devices make them promising candidates for designing PUFs. Additionally, the ultra-low power consumption and low computation time of these devices enable their use in applications with stringent requirements on energy efficiency and throughput.
This dissertation presents memristor- and eFash-memory-based PUF designs that show promising security characteristics such as near-to-ideal uniformity, diffuseness, robustness, and reliability. The robustness is verified by demonstrating the high output randomness with the test suits of the National Institute of Standards and Technology and by studying various machine learning attacks.
The specific contributions of this dissertation is that investigates several unexplored areas in crossbar-memory-design PUFs, e.g., finding optimal design for maximizing robustness characteristics, studying the impact of the capacity of machine learning models on robustness, and the impact of environmental change and thermal noise on reliability.