Energy-Efficient Neural Network Hardware Design and Circuit Techniques to Enhance Hardware Security

Abstract

Artificial intelligence (AI) algorithms and hardware are being developed at a rapid pace for emerging applications such as self-driving cars, speech/image/video recognition, deep learning, etc. Today’s AI tasks are mostly performed at remote datacenters, while in the future, more AI workloads are expected to run on edge devices. To fulfill this goal, innovative design techniques are needed to improve energy-efficiency, form factor, and as well as the security of AI chips. In this dissertation, two topics are focused on to address these challenges: building energy-efficient AI chips based on various neural network architectures, and designing “chip fingerprint” circuits as well as counterfeit chip sensors to improve hardware security. First of all, in order to deploy AI tasks on edge devices, we come up with various energy and area efficient computing platforms. One is a novel time-domain computing scheme for fully connected multi-layer perceptron (MLP) neural network and the other is an efficient binarized architecture for long short-term memory (LSTM) neural network. Secondly, to enhance the hardware security and ensure secure data communication between edge devices, we need to make sure the authenticity of the chip. Physical Unclonable Function (PUF) is a circuit primitive that can serve as a chip “fingerprint” by generating a unique ID for each chip. Another source of security concerns comes from the counterfeit ICs, and recycled and remarked ICs account for more than 80% of the counterfeit electronics. To effectively detect those counterfeit chips that have been physically compromised, we came up with a passive IC tamper sensor. This proposed sensor is demonstrated to be able to efficiently and reliably detect suspicious activities such as high temperature cycling, ambient humidity rise, and increased dust particles in the chip cavity.