Comprehensive study of physical unclonable functions on FPGAs: correlation driven Implementation, deep learning modeling attacks, and countermeasures
Loading...
Date
2020-06-09
Authors
KHALAFALLA, MAHMOUD
Advisor
Gebotys, Catherine
Journal Title
Journal ISSN
Volume Title
Publisher
University of Waterloo
Abstract
For more than a decade and a half, Physical Unclonable Functions (PUFs) have been
presented as a promising hardware security primitive. The idea of exploiting variabilities
in hardware fabrication to generate a unique fingerprint for every silicon chip introduced a
more secure and cheaper alternative. Other solutions using non-volatile memory to store
cryptographic keys, require additional processing steps to generate keys externally, and
secure environments to exchange generated keys, which introduce many points of attack
that can be used to extract the secret keys.
PUFs were addressed in the literature from different perspectives. Many publications
focused on proposing new PUF architectures and evaluation metrics to improve security
properties like response uniqueness per chip, response reproducibility of the same PUF
input, and response unpredictability using previous input/response pairs. Other research
proposed attack schemes to clone the response of PUFs, using conventional machine learning
(ML) algorithms, side-channel attacks using power and electromagnetic traces, and fault
injection using laser beams and electromagnetic pulses. However, most attack schemes to
be successful, imposed some restrictions on the targeted PUF architectures, which make
it simpler and easier to attack. Furthermore, they did not propose solid and provable
enhancements on these architectures to countermeasure the attacks. This leaves many
open questions concerning how to implement perfect secure PUFs especially on FPGAs,
how to extend previous modeling attack schemes to be successful against more complex
PUF architectures (and understand why modeling attacks work) and how to detect and
countermeasure these attacks to guarantee that secret data are safe from the attackers.
This Ph.D. dissertation contributes to the state of the art research on physical unclonable
functions in several ways. First, the thesis provides a comprehensive analysis of the implementation of secure PUFs on FPGAs using manual placement and manual routing
techniques guided by new performance metrics to overcome FPGAs restrictions with minimum
hardware and area overhead. Then the impact of deep learning (DL) algorithms is
studied as a promising modeling attack scheme against complex PUF architectures, which
were reported immune to conventional (ML) techniques. Furthermore, it is shown that
DL modeling attacks successfully overcome the restrictions imposed by previous research
even with the lack of accurate mathematical models of these PUF architectures. Finally,
this comprehensive analysis is completed by understanding why deep learning attacks are
successful and how to build new PUF architectures and extra circuitry to thwart these types
of attacks. This research is important for deploying cheap and efficient hardware security
primitives in different fields, including IoT applications, embedded systems, automotive
and military equipment. Additionally, it puts more focus on the development of strong intrinsic PUFs which are widely proposed and deployed in many security protocols used
for authentication, key establishment, and Oblivious transfer protocols.
Description
Keywords
Hardware security, Security primitives, Physical unclonable functions, Deep learning modelling attacks, Machine learning, FPGAs, Hardware security architecture