Machine Learning Vulnerability Analysis of FPGA-based Ring Oscillator PUFs and Counter Measures

Abstract

Physical Unclonable Functions (PUFs) exploit the manufacturing process variations inherent in silicon-based chips to generate unique secret keys. Although PUFs are supposed to be unclonable or unbreakable, researchers have found that they are vulnerable to machine learning (ML) attacks. In this article, we analyze the vulnerability of different FPGA-based Ring Oscillator PUFs (ROPUFs) to machine learning attacks. The challenge-response pairs (CRPs) data obtained from different ROPUFs is trained using different machine learning algorithms. From the study, it is found that the Artificial Neural Network (ANN) models can be used to train the ROPUFs with a training accuracy of 99.9% and a prediction accuracy of 62% when 5,000 CRPs are used for a challenge-response ROPUF. In this article, we assume a realistic situation where a small set of the CRP dataset (approximately 15% maximum) is unscrupulously obtained by the hacker. A prediction accuracy of 62% makes the PUF vulnerable to machine learning attacks. Therefore, a secondary goal of this article is the design of a ROPUF capable of thwarting machine learning modeling attacks. The modified XOR-inverter ROPUF drastically reduces the prediction accuracy from 62% to 13.1%, thus making it increasingly difficult for hackers to attack the ROPUF.

Security Analysis of Digital-Based Physically Unclonable Functions: Dataset Generation, Machine Learning Modeling, and Correlation Analysis

“…This modeling approach does not require any auxiliary information and solely relies on computations performed on the CRPs themselves [8]. As a consequence, PUFs are susceptible to ML-based attacks [9], especially when CRPs are accessible outside the chip without any protection mechanisms in place. Constructed PUF models can take the form of numerical models derived from collected data or ML models trained on a sufficient number of CRPs [10].…”

Designing a photonic physically unclonable function having resilience to machine learning attacks

“…1,[11][12][13][14][15] Machine learning (ML)-based attacks that predict responses for as-of-yet-unseen challenges have become an increasing threat to PUFs of all variants. 5,[16][17][18][19][20][21][22][23][24][25][26][27][28][29][30][31] These attacks do not directly represent the transformation between challenges and responses, but rather predict the result of that transformation after learning from a set of CRPs collected from a given PUF. Machine-learning-based attacks, including regression mappings, state vector machines, k-nearest neighbors, random forests, genetic algorithms, artificial neural networks, and generative adversarial networks, have proven effective against various conventional PUFs, and additional work has developed theoretical relationships between the number of CRPs used for training and response prediction accuracy.…”