BayesianPUFNet: Training Sample Efficient Modeling Attack for Physically Unclonable Functions

摘要

This paper proposes a deep neural network namedBayesianPUFNet that can achieve high prediction accuracy even with fewchallenge-response pairs (CRPs) available for training. Generally, modelingattacks are a vulnerability that could compromise the authenticity of physicallyunclonable functions (PUFs); thus, various machine learning methodsincluding deep neural networks have been proposed to assess the vulnerabilityof PUFs. However, conventional modeling attacks have not consideredthe cost of CRP collection and analyzed attacks based on the assumption thatsufficient CRPs were available for training; therefore, previous studies mayhave underestimated the vulnerability of PUFs. Herein, we showthat the applicationof Bayesian deep neural networks that incorporate Bayesian statisticscan provide accurate response prediction even in situations where sufficientCRPs are not available for learning. Numerical experiments show thatthe proposed model uses only half the CRP to achieve the same response predictionas that of the conventional methods.