From XOR PUF to CDC XOR PUF: Cost-Effectiveness, Statistical Characteristics, and Security Assessment

Abstract

With the high demand for using the Internet of Things (IoT), their security becomes a critical issue. Traditional cryptographic methods cover a part of IoT security, but not all the cases. Traditional cryptographic methods rely on storing secret keys in non-volatile memory. Hence, those secret keys are vulnerable to many attacks. Also, not all IoT devices can apply a traditional cryptographic method due to the implementation cost or chips' capacity.

Physical Unclonable Functions (PUF) are offered as lightweight security functions that produce secret keys extracted from the disorders that appear in the semiconductors (manufacturing variations). The produced keys by PUFs have the ability to identify a chip when comparing to others. Also, those keys are unclonable due to the hardness of cloning the manufacturing variations in a device.

One well-studied category of the delay-based PUFs is the XOR Arbiter PUFs, which were reported as a robust entity against machine learning attacks. Several studies have tried to show the weaknesses of the XOR Arbiter PUFs in terms of security. 9-XOR Arbiter PUF with a 64-bit challenge can be cracked when training a machine learning model with 350M challenge-response pairs (CRP). Similarly, the best attack on the 8-XOR Arbiter PUF with a 64-bit challenge used 30M CRPs. Considering the results of those studies, XOR Arbiter PUFs resembles robust.

Another XOR Arbiter PUF variant, the Component-Differentially-Challenged XOR PUF (CDC XPUF), was proposed to increase the range of keys PUF can produce based on the number of components. The CDC 5-XPUF with a 64-bit challenge was shown resistant against modeling attacks.

In this work, we investigate the security vulnerability of the XOR Arbiter PUFs and CDC XPUFs using a modified modeling attack that is revealed from the previous studies. Our proposed attack is able to crack the 9-XOR Arbiter PUF 64-bit with less than 5M CRPs in 9 minutes. Also, we modified the transformation function of the proposed attack to be applied to the CDC XPUFs.

Our results show that the XOR Arbiter PUFs are easy to model when assuming that their CRPs are freely accessible. Hence, we expand the study to investigate the n-XOR Arbiter PUF with n ranging from 2 to 10 with different assumptions such as access-restricted PUFs. For the CDC XPUF, our model is able to crack the CDC 4-XPUF 64-bit, but no larger number of components. Since the CDC XPUFs are robust against modeling attacks, we expand the research to include small CDC XPUFs with a small number of components and challenge length to be used in lightweight applications.

Finally, after showing the security assessments of the XOR Arbiter PUFs and CDC XPUFs, their physical properties must be studied carefully before confiding them in real security applications. Those properties include PUFs' fingerprint, reliability, and uniformity. The overall statistical characteristics study shows that the CDC XPUFs are better than XOR Arbiter PUF, especially for the uniqueness and randomness. Note, the experiments of this study are based on synthetic and silicon CRPs.