Using Statistical Models to Improve the Reliability of Delay-Based PUFs

Abstract

Physical Unclonable Functions (PUFs) use random physical variations to map input challenges to output responses in a way that is unique to each chip. PUFs are promising low cost security primitives but unreliability of outputs limits the practical applications of PUFs. This work addresses two causes of unreliability: environmental noise and device aging. To improve reliability, we constructively apply Machine Learning modeling, and use the models to predict and then discard challenge-response pairs (CRPs) that will be unreliable with respect to noise and aging on a given PUF instance. The proposed method provides flexibility to control error rate by deciding what percentage of challenges to discard. Our experiments find that a PUF with nominal reliability of 91% can be made fully reliable by discarding only 20% of challenges.