On the Scaling of Machine Learning Attacks on PUFs with Application to Noise Bifurcation

Abstract. Physical Unclonable Functions (PUFs) are seen as a promis-ing alternative to traditional cryptographic algorithms for secure and lightweight device authentication. However, most strong PUF propos-als can be attacked using machine learning algorithms in which a pre-cise software model of the PUF is determined. One of the most popular strong PUFs is the XOR Arbiter PUF. In this paper, we examine the machine learning resistance of the XOR Arbiter PUF by replicating the attack by Rührmaier et al. from CCS 2010. Using a more efficient imple-mentation we are able to confirm the predicted exponential increase in needed number of responses for increasing XORs. However, our results show that the machine learning performance does not only depend on the PUF design and and the number of used response bits, but also on the specific PUF instance under attack. This is an important observation for machine learning attacks on PUFs in general. This instance-dependent behavior makes it difficult to determine precise lower bounds of the re-quired number of challenge and response pairs (CRPs) and hence such numbers should always be treated with caution. Furthermore, we examine a machine learning countermeasure called noise bifurcation that was recently introduced at HOST 2014. In noise bifurca-tion, the machine learning resistance of XOR Arbiter PUFs is increased at the cost of using more responses during the authentication process. However, we show that noise bifurcation has a much smaller impact on the machine learning resistance than the results from HOST 2014 sug-gest