8. Conclusion and Future Work

Nowadays, PUFs are among the main building blocks of the security measures, which are considered necessary to not only mitigate attacks against ICs but also remedy the shortcomings of the corresponding traditional countermeasures. However, after more than a decade of the invention of PUFs, the design of a PUF fulfilling the necessary, and natural requirements is still a challenging task. Being unpredictable and unclonable are involved in these requirements, albeit being unsatisfied as demonstrated by the results of various attacks. ML attacks provide evidence supporting that an adversary can launch a cost-efficient attack to predict the response of the PUF to an unseen, arbitrarily chosen challenge. Although several studies focus on the security assessment of PUFs by mounting ML attacks, to the best of our knowledge, our study is the first work that addresses the issue with the final model delivery after the learning phase. In this regard, this thesis proposes a generic PAC learning framework, which has been successfully applied to known, and widely accepted families of PUFs, namely, Arbiter, XOR Arbiter, RO-, and BR-PUFs. More specifically, our framework clearly states how these PUF families can be learned, for given levels of accuracy and confidence. Interestingly enough, the maximum number of CRPs required to mount our attack can be calculated beforehand.