A Novel Attack on Machine-Learning Resistant Physical Unclonable Functions

Many of today's proposed designs for strong physical unclonable functions (PUFs) are constructed from strong PUF building blocks that are known to be themselves vulnerable to machine-learning (ML) attacks, such as the Arbiter PUF or Ring Oscillator PUF. These designs aim to make the global design ML-resistant by obscuring the true challenge and/or responses corresponding to the vulnerable building blocks, thereby making it more difficult for ML attacks to infer underlying model parameters. In this work, we propose a model-free ML attack, inspired by deep learning-based image imputation algorithms, that successfully breaks these popular countermeasures. Our attack is unique in that it uses not only the observed challenge-response pairs (CRPs) of the device being attacked but also uses CRPs collected from other PUFs of the same type, which results in improved bit-error rates (BERs). We show that our attack breaks several proposed ML-resistant PUF architectures based on obfuscation of vulnerable strong PUF building blocks.