A Novel Configurable RO-Obfuscated PUF Design with Machine Learning Immunity

Strong physical unclonable function (PUF) is a low-cost hardware security primitive to protect Internet-of-Things (IoT) devices. However, it may be attacked by machine learning (ML). Various PUF models designed in complex structures, such as nonlinearity or challenge-response pair (CRP) obfuscation, have been presented to combat these risks. However, these methods mainly increase area overhead, and the prediction rate is still very high. Therefore, this paper proposes a structure obfuscated PUF named SO-PUF. Our proposal derives from the configurable ring oscillator (CRO) PUF. The CRO can be transformed into two different structures depending on the challenges by randomly deleting one of the two NOT gates in each stage. Thus, half of the CRPs generated by SO-PUF are invalid, which will confuse the attackers. We have implemented and verified the performance of SO-PUF on the Xilinx-6XC6SLX25 microboard. Experimental results show that compared with the dual-mode PUF, the SO-PUF improves uniqueness by 0.71 %, temperature reliability by 38.4 %, and voltage reliability by 1.92 %. In addition, the SO-PUF also reduces the prediction rates of LR, SVM, and ANN modeling attacks by 0.26 %, 0.38 %, and 5.62 %, respectively. The results prove that SO-PUF has better resistance to ML attacks than dual-mode PUF.