Evolution strategy based machine learning attack on physical unclonable function
Hadi Sutikno, Eric Darian
Date of Issue2018
School of Electrical and Electronic Engineering
Recently, Physical Unclonable Function (PUF) has been rising as an alternative solution for device authentication. Yet, it still has some problems such as unreliability and vulnerable to Machine Learning Attack. Previous research proves that PUF is prone to Machine Learning Attack. Therefore, this project aims to improve the security of PUF-based authentication protocol by developing an algorithm which can predict the reliability of Challenge-Response Pairs (CRPs) based on Machine Learning Attack Model Parameters.In this project, we are mainly using one of Evolution Strategies algorithm namely Covariance Matrix Adaptation to attack a Matlab-simulated Arbiter PUF. Based on the result, Evolution Strategies based Machine Learning algorithm is more effective than Support Vector Machine (SVM) algorithm in attacking and modelling the Arbiter PUF. Using the PUF model parameter obtained from the CMA-ES Attack, an algorithm which has the ability to predict the reliability of CRPs is developed. This algorithm shows that it is possible to predict whether CRPs are reliable or not based on the parameters generated from Machine Learning Attack. The result of this project shows that we are able to quickly and accurately predict the reliability of CRPs for an Arbiter PUF. We achieved about 98.6% accuracy in predicting whether future CRPs are reliable or not. Therefore, it opens up the opportunity to improve the security of PUF-based Authentication Protocol by using mixture of reliable and unreliable CRPs. This will make anyone who tries to attack PUF by stealing the CRPs in the authentication process not getting the effective training data.
Final Year Project (FYP)
Nanyang Technological University