Efficient hybrid side‐channel/machine learning attack on XOR PUFs

Abstract

A novel hybrid side-channel (SC)/machine learning attack is explored in this Letter to leak the confidential information of non-linear physical unclonable functions (PUFs): XOR arbiter PUFs. In the proposed hybrid attack, SC analyses are utilised to pre-process the input challenge of XOR arbiter PUFs to add high correlation among all the input challenge bits. Subsequently, a convolutional neural network (CNN) attack is performed on the correlated input challenge bits to extract the critical feature among the neighbour data to significantly improve its training/testing accuracy. As shown in the results, after applying the SC analyses to add correlation for the input challenge of an XOR arbiter PUF, the training/testing accuracy of the hybrid attack can be boosted over 0.98. In contrast, the training/testing accuracy of a regular CNN attack on the XOR arbiter PUF is around 0.64 due to the lack of the corresponding correlation.