Security Evaluation and Enhancement of Bistable Ring PUFs

Xiaolin Xu, Ulrich Rührmair, Daniel E. Holcomb, Wayne P. Burleson

TL;DR

This paper investigates the security of Bistable Ring (BR) Physical Unclonable Functions (PUFs) against machine learning (ML) attacks, finding them vulnerable. To address this, the authors propose an XOR-based enhancement, demonstrating that combining multiple BR PUFs can effectively thwart current ML modeling techniques while also improving uniformity.

What Problem Does This Paper Address?

Physical Unclonable Functions (PUFs) are a promising hardware security primitive that leverage unique manufacturing variations in integrated circuits to create unclonable keys. They operate via a challenge-response mechanism, where a specific input (challenge) elicits a unique output (response). While PUFs offer an alternative to traditional cryptographic methods, their security against sophisticated attacks, particularly machine learning (ML) modeling, is a critical concern. The paper focuses on the Bistable Ring (BR) PUF, a relatively new PUF architecture designed for enhanced security. However, prior work has shown that BR PUFs are susceptible to ML attacks. This paper aims to comprehensively evaluate the BR PUF's resilience against advanced ML techniques and propose a practical method to enhance its security against such threats.

Methods

The authors employ Support Vector Machines (SVMs) as their primary ML tool for modeling attacks. SVMs are chosen for their power in binary classification, which is well-suited for predicting the binary responses of PUFs. They develop analytical models for both the standard BR PUF and its twisted variant (TBR PUF) to understand the underlying mechanisms influencing their responses. For enhancing security, they propose an "XOR BR PUF" strategy. This involves applying the same challenge to multiple individual BR PUFs in parallel and then XORing their individual responses to produce a single, final output. This XORing is hypothesized to obfuscate the individual PUF responses, making it harder for ML models to learn the overall behavior.

Datasets

The paper utilizes Challenge-Response Pairs (CRPs) generated from BR PUF and TBR PUF implementations on Xilinx Spartan-VI FPGA boards. They implement PUFs of various bitlengths (32, 64, 128, and 256 bits). For each implementation, they collect a substantial number of CRPs (up to 1,000,000) to train their ML models. To mitigate the impact of measurement noise, they employ majority voting over 11 repeated measurements for each CRP.

Key Insights and Ideas

  • BR PUFs are surprisingly vulnerable to ML: Despite being designed with security in mind, even sophisticated ML models like SVMs can achieve high prediction rates on BR PUFs, indicating a significant security gap.
  • XORing as a robust defense: The XOR combination of multiple BR PUFs proves to be a highly effective strategy for thwarting ML modeling attacks. This approach significantly increases the complexity required for an ML model to achieve a high prediction rate.
  • Uniformity improvement: Beyond security, the XOR function also demonstrably improves the uniformity of BR PUFs, addressing a known bias issue in the original BR PUF design. This is crucial for PUF reliability and entropy.
  • Scalability of attacks: The number of CRPs required to successfully model a PUF increases with its bitlength, but this scaling is not enough to make single BR PUFs secure against powerful ML techniques.

Key Results

  • SVM attacks on single BR PUFs (and TBR PUFs) achieved prediction rates exceeding 95% with less than 10,000 CRPs for all tested bitlengths (32, 64, 128, 256). This confirms their vulnerability.
  • The proposed XOR BR PUF strategy showed significant resilience. With 4 or more XORed BR PUFs, current SVM modeling attacks were unable to break the complexity, with prediction rates dropping to around 50% (implying random guessing).
  • Reliability of XOR BR PUFs degrades as more PUFs are XORed, a trade-off that needs consideration.
  • Uniformity of BR PUFs is initially biased, but XORing improves it, bringing it closer to the ideal value of 0.5.

Limitations and Future Work

The authors acknowledge that the reliability of XOR BR PUFs decreases as the number of XORed PUFs increases. This is because any unstable sub-response from a single PUF can affect the final XORed output. They suggest that using CRPs that settle quickly can mitigate this issue. Future work proposed includes exploring other ML modeling attacks, such as Evolutionary Strategy and Logistic Regression, to further assess the security of XOR BR PUFs and to investigate other PUF metrics.

Summary updated

BibTeX
@article{xu2015security,
  author = {Xiaolin Xu and Ulrich Rührmair and Daniel E. Holcomb and Wayne P. Burleson},
  title = {{Security Evaluation and Enhancement of Bistable Ring PUFs}},
  journal = {IACR Cryptol. ePrint Arch.},
  year = {2015},
  doi = {10.1007/978-3-319-24837-0\_1},
}
Figure