Vulnerability Analysis Employing a Generative Adversarial Network Model for a Ring Oscillator Physical Unclonable Function
Talha Hussain Syed, Akshay Raghavendra Kulkarni, Mohammed Y. Niamat · 2025
Physical Unclonable Functions (PUFs) are critical to authentication and assurance of integrated circuits (ICs), by leveraging inherent manufacturing process variations in the semiconductor chips to generate unique Challenge-Response Pairs (CRPs). This research investigates the susceptibility of FPGA based Ring Oscillator Physical Unclonable Function (ROPUF) by predicting its CRPs using machine learning models and Generative Artificial Intelligence (Gen-AI). An essential component of this research is identifying potential security risks to prevent the incorporation of vulnerable components into secure applications. In this paper, we consider a plausible scenario where an adversary illicitly acquires 5% of the CRP dataset, and a Generative Adversarial Network (GAN) is used to generate synthetic CRPs that are subsequently added to augment the training dataset by $35 \%$. With this augmented data, machine learning models demonstrate a notable improvement in predictive accuracy. Decision Tree (DT) prediction accuracy increases from 60.17% to 67.15%, K-Nearest Neighbors (KNN) from $64.67 \%$ to $72.17 \%$, Random Forest Classifier (RF) from $67.83 \%$ to $75.74 \%$, and Extreme Gradient Boosting (XGB) from $\mathbf{7 3. 8 3 \%}$ to $\mathbf{8 3. 4 1 \%}$. This enhanced predictive capability reveals that machine learning models even with partial datasets augmented using GAN models can predict the original CRPs with increased accuracy.