Deciphering the Unreplicable: Machine Learning Paradigms for Assessing and Manipulating Physical Unclonable Functions

Physically unclonable functions, such as inscription key generators, take advantage of their inherent physical distinctions. Even with promises of immunity to modelling and challenge- response pair prediction, the introduction of machine learning techniques presents a problem. Machine learning is a great tool for data extraction and system modelling because of its analytical strength and adaptability, especially when it comes to preventing attacks on systems that are physically difficult to recreate. This work extends earlier approaches by first creating a 64-bit binary dataset with Field Programmable Gate Array design tools, then modifying the dataset labels and optimising performance measures. The study employs four machine learning strategies, each with a distinct accuracy rate. K-fold logistic regression attains an accuracy of 50.11 percent, which is marginally more than support vector machines' precision of 49.50 percent. By comparison, a random forest classifier employing K-fold yields an accuracy rate of 49.00 percent; however, deep learning performs exceptionally well, with an astounding precision rate of 82.01 percent. This study attempts to evaluate PUF susceptibility by presenting accuracy, sensitivity, and specificity values.