Linear and non-linear machine learning attacks on physical unclonable functions

TL;DR

This paper investigates linear and deep learning attacks on optical physical unclonable functions (PUFs), using simulations and real data to evaluate attack effectiveness and improve modeling techniques for near-perfect prediction of PUF outputs.

Contribution

It introduces a comprehensive simulation framework for optical PUFs, compares linear and deep learning attack methods, and develops improved models capable of nearly perfect output prediction.

Findings

Deep learning models outperform linear models in attack success

Simulated datasets reveal factors influencing security levels

Real PUF data confirms models' high attack accuracy

Abstract

In this thesis, several linear and non-linear machine learning attacks on optical physical unclonable functions (PUFs) are presented. To this end, a simulation of such a PUF is implemented to generate a variety of datasets that differ in several factors in order to find the best simulation setup and to study the behavior of the machine learning attacks under different circumstances. All datasets are evaluated in terms of individual samples and their correlations with each other. In the following, both linear and deep learning approaches are used to attack these PUF simulations and comprehensively investigate the impact of different factors on the datasets in terms of their security level against attackers. In addition, the differences between the two attack methods in terms of their performance are highlighted using several independent metrics. Several improvements to these models and new attacks will be introduced and investigated sequentially, with the goal of progressively improving modeling performance. This will lead to the development of an attack capable of almost perfectly predicting the outputs of the simulated PUF. In addition, data from a real optical PUF is examined and both compared to that of the simulation and used to see how the machine learning models presented would perform in the real world. The results show that all models meet the defined criterion for a successful machine learning attack.