Machine learning attack on hardware implementation of one-way function

• Main Author: Lauw, Andri Renardi
• Other Authors: Chang Chip Hong
• Format: Final Year Project
• Language: English
• Published: 2017
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/71815
• Institution: Nanyang Technological University

Physical Unclonable Functions (PUF) with exponentially growing number of challenges is an ideal candidate to become a device authentication for Internet of Things (IoT) devices. However, recent research shows that one of the most popular type of PUF, Arbiter PUF, is prone to mathematical modelling attack. Machine Learning (ML) based modelling attacks are the currently most relevant and effective attack form for PUF. This project attempts to attack Arbiter PUF with several machine learning algorithms and compare their performance to obtain the most effective learning algorithm against this type of PUF. It turns out that Support Vector Machine (SVM) is the most effective learning algorithm among all the six algorithms tested (Naïve Bayes, Decision Tree, Logistic Regression, Gradient Boosting, Neural Network, and SVM) to attack the PUF in terms of effectiveness and efficiency, which has prediction accuracy of 99% with relatively small amount of training data. Furthermore, a countermeasure against machine learning attack using hash function is proposed and tested. The challenge is masked with SHA-256 hashing algorithm before fed into the PUF. The result shows that fortification of the system using a hash function greatly reduces the effectiveness of machine learning attack to PUF. It reduces the prediction accuracy from 99% to nearly 50%, which is only slightly better than random chance. The hashing algorithm breaks the linear relation between the challenges with their corresponding responses due to its non-linear nature, hence effective to shield the system against machine learning attack which leverage the linear model of the system.