Candidate: Mahmoud Khalafalla
Title: Comprehensive study of physical unclonable functions on FPGAs: correlation driven Implementation, deep learning modeling attacks, and countermeasures
Date: May 14, 2020
Time: 9:00 AM
Place: REMOTE PARTICIPATION
Supervisor(s): Gebotys, Catherine
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 PhD dissertation contributes to 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 type 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.