Emerging Opportunities and Challenges in Hardware Security

Thumbnail Image


Publication or External Link





Recent years have seen the rapid development of many emerging technologies in various aspects of computer engineering, such as new devices, new fabrication techniques of integrated circuits (IC), new computation frameworks, etc.

In this dissertation, we study the security challenges to these emerging technologies as well as the security opportunities they bring. Specifically, we investigate the security opportunities in double patterning lithography, the security challenges in physical unclonable functions, and security issues in machine learning.

Double patterning lithography (DPL) is an emerging fabrication technique for ICs. We study the security opportunities that DPL brings at the layout level. DPL is used to set up two independent mask development lines which do not need to share any information. Under this setup, we consider the attack model where the untrusted employee(s) who has access to only one mask may try to infer the entire circuit design or insert additional malicious circuitry into the design. As a countermeasure, we customize DPL to decompose the layout into two sub-layouts in such a way that each sub-layout individually exposes minimum information about the other and hence protects the entire layout from any untrusted personnel.

Physical unclonable functions (PUF) are a type of circuits for which each copy (of the same circuit structure) has a unique and unpredictable functionality.

The unpredictable behavior is caused by the manufacturing variations of electronic devices. However, for many state-of-the-art PUF designs, we show that the device variations can be estimated using an optimization-theoretic formulation and hence the PUF's input-output behavior becomes predictable. Simulations show a substantial reduction in attack complexity compared to previously proposed machine learning based attacks.

Neural network (NN) is an emerging computation framework for machine learning (ML). It is increasingly popular for system developers to use pre-trained NN models instead of training their own because training is painstaking and sometimes requires private data. We call these pre-trained neural models neural intellectual properties (IP). Neural IPs raise multiple security concerns. On the one hand, as the IP user does not know about the training process, it is crucial to ensure the integrity of the neural IP. To this end, we investigate possible hidden malicious functionality, i.e. neural Trojans, that can be embedded into neural IPs and propose effective mitigation techniques. On the other hand, the neural IP owner may want to protect the NN model from reverse engineering attacks. However, it has been shown that hardware side-channels can be exploited to decipher the structure of neural networks. We propose both a novel attack approach based on cache timing side-channel and a defensive memory access mechanism.

NNs also raise challenges to conventional hardware security techniques. Specifically, we focus on its challenge to logic locking, a strong key-based protection of hardware IP against untrusted foundries by injecting incorrect behavior into the digital functionality when the key is incorrect. We formally prove a trade-off between the amount of injected error and the complexity of Boolean satisfiability (SAT)-based attacks to find the correct key. Due to the inherent error resiliency of NNs, state-of-the-art logic locking schemes fail to inject enough error to derail NN-based applications while maintaining exponential SAT complexity. To fix this issue, we propose a novel secure and effective logic locking scheme, called Strong Anti-SAT (SAS), to lock the hardware and make sure that the NN modes undergo significant accuracy loss when any wrong key is applied.