Cryptology ePrint Archive: Report 2019/566

Deep Learning based Model Building Attacks on Arbiter PUF Compositions

Pranesh Santikellur and Aritra Bhattacharyay and Rajat Subhra Chakraborty

Abstract: Robustness to modeling attacks is an important requirement for PUF circuits. Several reported Arbiter PUF com- positions have resisted modeling attacks. and often require huge computational resources for successful modeling. In this paper we present deep feedforward neural network based modeling attack on 64-bit and 128-bit Arbiter PUF (APUF), and several other PUFs composed of Arbiter PUFs, namely, XOR APUF, Lightweight Secure PUF (LSPUF), Multiplexer PUF (MPUF) and its variants (cMPUF and rMPUF), and the recently proposed Interpose PUF (IPUF, up to the (4,4)-IPUF configuration). The technique requires no auxiliary information (e.g. side-channel information or reliability information), while employing deep neural networks of relatively low structural complexity to achieve very high modeling accuracy at low computational overhead (compared to previously proposed approaches), and is reasonably robust to error-inflicted training dataset.

Category / Keywords: applications / physically unclonable function, machine learning, deep learning, arbiter puf, PUF

Date: received 26 May 2019

Contact author: pranesh sklr at iitkgp ac in, rschakraborty@cse iitkgp ac in

Available format(s): PDF | BibTeX Citation

Version: 20190527:092510 (All versions of this report)

Short URL: ia.cr/2019/566


[ Cryptology ePrint archive ]