Analysis and Machine Learning Vulnerability Assessment of XOR-Inverter based Ring Oscillator PUF Design

Physical Unclonable Functions (PUFs) exploit the manufacturing process variations inherent in silicon-based chips to generate unique secret keys. Although PUFs are supposed to be unclonable or unbreakable, researchers have found that ROPUFs, in general, are vulnerable to machine learning attacks. In this paper, we analyze the vulnerability of an XOR inverter based ROPUF to machine learning attacks. The XOR ROPUF is designed and implemented using an Artix-7 FPGA. The Challenge Response Pairs (CRPs) data obtained from the XOR-Inverter ROPUF is trained using different machine learning models including Logistic Regression, kernel Support Vector Machine, and ANN based modeling. From the study, it is found that the ANN models can be used to train the XOR-inverter based PUF CRPs with a training accuracy close to 99%. In terms of prediction accuracy, the ANN models can predict 69.41% of the CRPs. Further, in order to thwart machine learning attacks, we propose a new modified XOR-Inverter ROPUF design which drastically reduces the prediction accuracy of machine learning modeling attacks.