UCLA

In the paradigm of Internet-of-Things (IoT), smart devices will proliferate our living and working spaces. The recent decade has already witnessed an explosive growth of smartphones and wearable devices. A plethora of newer and even more powerful systems are emerging. IoT will enable more fluid human-computer interaction and immersive experiences in smart homes. IoT will facilitate rich sensing and actuating in intelligent warehousing and manufacturing. IoT will also empower fast and accurate perception and decision making in autonomous vehicles. The paradigm has elevated the role of the devices that constitute the edge of the network. Because of the sensitive nature and the sheer volume of the data generated by those devices, edge computing becomes a more effective and efficient option. While it brings better privacy protection and latency reduction in applications, edge computing is associated with various constraints. For the sizable list of devices that are operating on batteries, their sustainable operation usually calls for extremely efficient and judicious use of energy. Further, the inherent vulnerability accompanying the deployment in unsafe environments requires extra layers of security.

In this dissertation, we study the energy and security problems of edge computing in the context of machine learning. We present various learning-based techniques for improving energy efficiency. In contrast to the traditional resource allocation mechanisms that typically adopt handcrafted rules and heuristics, we adopt a framework where we use machine learning learn to create online resource allocation strategies from optimal offline solutions. We demonstrate the effectiveness of the framework in applications and scenarios including DVFS, computation offloading and sensor networks. In the video decoding case, our machine learning enabled strategies have approximated optimal solutions with an average of 2\% error and achieved 40\% in energy savings. In an increasing number of edge computing applications, machine learning algorithms themselves constitute the core and the major workload. Many of those applications have high energy consumption and are vulnerable to security issues such as intellectual property theft. To solve the problems, we derive techniques directly from the machine learning processes. We present computer vision-oriented adaptive subsampling strategies for image sensors, model pruning and customization methods for deep neural networks, and deep neural network watermarking for intellectual property protection. These techniques improve energy efficiency and security of machine learning at very little or even zero cost of the performance of the models.