Matrix Encryption based Anti-Machine Learning Attack Algorithm for Strong PUF
Ziyu Zhou, Gang Li, Pengjun Wang, Ming Ye · 2021 IEEE 14th International Conference on ASIC (ASICON) · 2021
Physically unclonable function (PUF) has broad application prospect in the field of hardware security. However, due to the inevitable correlation between the challenge and response of strong PUF, it is vulnerable to machine learning (ML) modeling attack. In order to improve the anti-ML attack ability of strong PUF, this paper proposes an anti-ML attack algorithm for strong PUF based on matrix encryption (ME). It divides the responses generated by the original challenge into groups, and outputs the encrypted data by matrix multiplication as the final response. Since matrix multiplying encryption cannot inversely infer plaintext through ciphertext and the attacker cannot obtain the actual response generated by PUF through inverse transformation. Therefore, the proposed encryption method can be disclosed to the public. The experimental results show that even if nearly one million CRPs of 64-bit APUF, XOR-APUF and MPUF are collected to predict attacks after ME, the prediction accuracy of several ml attacks is about 50%, which is equivalent to random guessing.