Suh et al. utilized PUF in a RFID system [33] , where each response of the PUF tag is stored in the backend database yet drew during authentication. If the new response generated from the tag is the same as that stored in the database, then one-way authentication of the tag valid to the reader is achieved. Though, the method proposed in [13] is not considered in this research due to the manufacturing cost as well as the vulnerability to Machine Learning (ML) attacks [35] . In recent years, such a kind of attacks has severe impact on PUF, since the delay parameters of PUF mainly determine the challenge- response behavior of the arbiter PUF. Machine Learning (ML) algorithms collect and analyze the challenge-response pairs (CRPs) of the arbiter PUF and randomly generates n PUF models with different delay parameters. Furthermore, these models are trained by collected CRPs, so the PUF model with the challenge response behavior closest to the original arbiter PUF is selected among n PUF models. Additionally, the delay parameters of the selected PUF model are randomly mutated so new PUF models are generated, and continuously repeated as many times as needed until the final trained PUF model produces a similar response as the original arbiter PUF