Machine Learning Assisted Challenge Selection for Modeling Attack Resistance in Strong PUFs

Physical Unclonable Functions (PUF) have been proposed as security primitives in many applications. However, previous studies indicate that delay based strong PUFs are vulnerable under machine learning (ML) based modeling attacks as such algorithms can achieve extremely high prediction accuracy with small number of training data. In this paper, we will first show that these results slightly overestimate the achievable prediction accuracy, as the additive delay assumption may not be true for boundary cases. Next, it is shown that we can use ML algorithms to select such boundary cases to prevent attackers from building an accurate prediction model, which significantly improves the unpredictability of strong PUFs.