PAC Learning of XOR Arbiter PUFs
The content of this chapter is mainly based on Ganji et al., Trust and Trustworthy Computing, Springer, Berlin, pp. 22–39, 2015, (). In this chapter, besides the development of a PAC learning framework for XOR Arbiter PUFs, a theoretical limit for ML attacks as a function of the number of the chains and the number of Arbiter PUF stages has been established. Furthermore, we show that our approach deals with the noisy responses in an efficient fashion so that in this case, the maximum number of CRPs collected by the attacker is polynomial in the noise rate. Our rigorous mathematical approach matches the results of experiments, which can be found in the literature. Last but not least, on the basis of learning theory concepts, this chapter explicitly states that the current form of XOR Arbiter PUFs may not be considered as an ultimate solution to the problem of insecure Arbiter PUFs.