Artificial intelligence techniques driven by deep learning have experienced significant advancements in the past decade. The usage of deep learning methods has increased dramatically in practical application domains such as autonomous driving, healthcare, and robotics, where the utmost hardware resource efficiency, as well as strict hardware safety and reliability requirements, are often imposed. The increasing computational cost of deep learning models has been traditionally tackled through model compression and domain-specific accelerator design. As the cost of conventional fault tolerance methods is often prohibitive in consumer electronics, the question of functional safety and reliability for deep learning hardware is still in its infancy. This dissertation outlines a novel approach to deliver dramatic boosts in hardware safety, reliability, and resource efficiency through a synergistic co-design paradigm. We first observe and make use of the unique algorithmic characteristics of deep neural networks, including plasticity in the design process, resiliency to small numerical perturbations, and their inherent redundancy, as well as the unique micro-architectural properties of deep learning accelerators such as regularity. The advocated approach is accomplished by reshaping deep neural networks, enhancing deep neural network accelerators strategically, prioritizing the overall functional correctness, and minimizing the associated costs through the statistical nature of deep neural networks. To illustrate, our analysis demonstrates that deep neural networks equipped with the proposed techniques can maintain accuracy gracefully, even at extreme rates of hardware errors. As a result, the described methodology can embed strong safety and reliability characteristics in mission-critical deep learning applications at a negligible cost. The proposed approach further offers a promising avenue for handling the micro-architectural challenges of deep neural network accelerators and boosting resource efficiency through the synergistic co-design of deep neural networks and hardware micro-architectures.

The increasing ubiquity of communication-capable devices such as smartphones and smart vehicles has paved the way for a plethora of useful Participatory Cyber-Physical Systems (CPSs) in the defense, medical, and commercial sectors. Such systems promise to collect and analyze vast amounts of timestamped, location-contextualized user data, allowing various agencies to act on the results. However, given that the utility of these systems depends absolutely on the number of willing participants and the quality of their data, CPSs must be accessible to attract as large a user base as possible, while also possessing countermeasures against adversarial data intended to mislead the decision-making of the systems.

We begin by defining the structure of these CPS systems and the capabilities of the adversary, showing the particular danger of fake, software-simulated Sybil nodes that provide the adversary immense deceptive power at no meaningful cost or risk. We claim that no currently existing Sybil countermeasure successfully addresses both the requirements posed by CPS systems and the deceptive techniques available to the adversary---namely the introduction of a small number of physical malicious nodes who would help disguise the Sybils---and argue that a new approach is needed.

To address the Sybil threat, we develop a holistic 3-stage Sybil detection framework that may be employed on-demand to detect Sybil nodes around a target location. The first stage comprises of an inter-node communication scheme that uses local radio broadcasts as proof-of-physical-presence, and is designed specifically to address the CPS requirements and the adversarial deception techniques. The second stage consists of a detection layer that examines each node's observed pairwise RSSI readings, searching for those inconsistent with their expected distributions given their claimed locations. Due to the time-constrained nature of the communication scheme however, many pairwise connections randomly fail and therefore do not have any RSSI readings attached to them. This binary connection success/failure information is utilized by the third stage, wherein we develop a probabilistic framework to assess the likelihood of a node's observed combination of connection outcomes.

To evaluate our approach, we have built an extensive simulation environment that allows for complex adversarial obfuscation strategies. Using this environment, we demonstrate both the efficacy and robustness of our detection methodology.

The cheap versatility of modern embedded systems is being harnessed to attack sprawling tasks with teams of cooperating machines - infrastructure monitoring, automated patient health monitoring, and the management of cyber-physical systems are some prime examples. Parallel advances in networking technology, particularly MIMO signal processing capabilities,

facilitate wireless tracking and localization for coordination of physical activities among independent nodes. As these technologies enter wider markets, they inevitably become targets of malicious adversaries, and the ability to detect and mitigate attacks becomes necessary. In many applications, the ability to securely determine the location or proximity of one entity relative to another is important for maintaining system security. This work concentrates on a specific and relatively new RF-based localization scheme called Received Signal Strength Ratio (RSSR). Derived from a more primitive signal-strength-based localization technique, RSSR takes the ratio of signal strength measurements at 2 (or more) receivers to determine the distance of a transmitter from the device. This construct has attracted interest for its potential utility in securing ad-hoc networks and body-area networks. However, researchers proposing RSSR as the backbone of proximity-based authentication systems have not thoroughly considered certain realistic attacker capabilities. In this work, we present a threat model that characterizes the security of an RSSR-based proximity authentication system in more detail than previous research; describe a generic attack on the security of such systems; and discuss a set of mitigation strategies that ultimately restore the effectiveness of RSSR as a secure distance and proximity verification scheme.