Advancing PUF Security Machine Learning Assisted Modeling Attacks
Niraj Prasad Bhatta, Fathi Amsaad · 2024
Physically Unclonable Functions (PUFs) works as an essential component for hardware security, using their inherent unpredictability to prevent unauthorized access. However, arbiter PUFs, which are widely used, are sensitive to modeling attacks, especially when applying machine learning methods. This study shows the usefulness of various machine learning algorithms in compromising the security of arbiter-PUFs. For this study, we review how effectively logistic regression, K-nearest neighbor, nave bayes, gradient boosting, and random forest classifiers can predict PUF reactions. We make use of a dataset of 12,000 challenge-response pairs (CRPs) for a 64-stage arbitrator-PUF. The Random Forest Classifier emerged as the most proficient, producing the best precision to duplicate the behavior of the arbitrator-PUF with 89% in the attack part. To address this problem, a defense mechanism that integrates random fluctuations into challenge bits was developed with the purpose of undermining the prediction powers of the machine learning models. Our findings reveal a significant reduction in attack accuracy after the adoption of protection measures, highlighting the potential of noise injection as a technique to increase PUF security with the accuracy of 81%. The conclusions of this research not only underline the vital need for effective defense strategies against machine learning attacks on PUFs, but also offer a feasible strategy to defend these crucial security components.