Optimal performance of simple low-cost optical physical unclonable functions resilient to machine learning attacks

In this paper we reconsider the Physical Unclonable Functions based on the traditional approach of optical scattering on random optical media. These devices have the major advantage of utilization of simple and very low-cost technology and therefore the potential to be installed all over the network providing critical cybersecurity operations such authentication, real time cryptographic key generation and generation of trues random sequences. To comply with the requirements of the aforementioned operations, critical issues must be resolved. We propose and implement algorithms for the generation of an almost unlimited number of uncorrelated optical challenges. We show experimentally that the uncorrelated challenges result in optical speckle which after the proper numerical processing generates true random sequences, we determine the optimal illumination conditions in order to achieve the best possible performance in terms of robustness and unpredictability. Moreover, we study the resilience of the PUF against machine learning (ML) attacks. We conclude experimentally that under certain illuminating conditions and using the uncorrelated challenges, the network cannot predict the responses being trained with a very large number of challenge responses (24000 pairs).