MA-GRNN: a high-efficient modeling attack approach utilizing generalized regression neural network for XOR arbiter physical unclonable functions

In this paper, we propose a novel modeling attack approach to predict the responses of XOR arbiter physical unclonable functions (XOR APUFs), which improves the prediction accuracy and reduces the computational time. The high-dimensional mathematical model of XOR APUF is established and its weakness is analyzed. Furthermore, a modeling attack approach based on the generalized regression neural network (MA-GRNN) is introduced to approximate the responses of XOR APUFs. As a proof-of-concept, four popular machine learning algorithms are utilized to evaluate the attack efficacy of 3-XOR, 4-XOR, 5-XOR and 6-XOR APUF schemes. Experimental results show that the MA-GRNN achieves a high prediction accuracy compared to the other three modeling attack approaches while requiring less computational time simultaneously.