Optimization of Combined Power and Modeling Attacks on VR PUFs With Lagrange Multipliers

In this brief, optimized combined power and modeling attacks (CPMAs) are proposed to crack lightweight strong physical unclonable functions (PUFs): voltage regulator (VR) PUFs. The relationship between input challenge and output response of a VR PUF is approximately modeled with polynomials at first. Then an objective function based on the polynomials is established for maximizing the accuracy of the model. Furthermore, the transient power consumption associated with the VR PUF is extracted as the critical constraint. After utilizing Lagrange multipliers to solve the optimization problem related to the VR PUF, the optimum polynomial that is able to assist to maximize the efficacy of CPMAs on the VR PUF is obtained. Ultimately, multi-layer perception (MLP) neural networks are built based on the optimum polynomial to successfully break the VR PUF. As shown in the result, the training accuracy of the proposed CPMAs is boosted over 96.98% while the training accuracy of regular CPMAs is only about 83.69%.