Resilience evaluation of memristor based PUF against machine learning attacks

Area of Science:

• Hardware Security
• Nanotechnology
• Machine Learning

Background:

• Physical unclonable functions (PUFs) are crucial for hardware security, generating unique keys from process variations.
• CMOS-based PUFs are common but vulnerable to machine learning (ML) modeling attacks.
• Memristor-based PUFs (MR-PUFs) offer an alternative using nanotechnology, providing inherent randomness and nonlinearity.

Purpose of the Study:

• To develop and apply an ML analysis and attack framework to evaluate the security of memristor-based PUFs (MR-PUFs).
• To assess the randomness prediction resiliency of MR-PUFs against various ML algorithms.
• To validate the cryptographic randomness of MR-PUFs through NIST testing and ML-based attack simulations.

Main Methods:

• Constructed an ML attack framework utilizing Logistic Regression, SVM, GMM, K-means, Random Forest, XGBoost, and LSTM.
• Evaluated MR-PUF randomness prediction using these ML models with efficient time and data complexity.
• Performed holistic NIST cryptographic randomness testing on the MR-PUF outputs.

Main Results:

• ML models demonstrated low accuracy (within 50%) and ROC (within 0.5) in predicting MR-PUF randomness.
• The study confirmed the failure of these ML models to predict the PUF's random data.
• Attack execution times were efficient, with linear and quadratic complexities observed.

Conclusions:

• Memristor-based PUFs exhibit significant resilience against sophisticated machine learning modeling attacks.
• The inherent randomness and nonlinear characteristics of memristors contribute to robust hardware security.
• MR-PUFs present a promising and secure alternative to traditional CMOS-based PUFs for cryptographic applications.