A 4T/Cell Amplifier-Chain-Based XOR PUF With Strong Machine Learning Attack Resilience

This paper presents an amplifier-chain-based XOR physical unclonable function (AC-XOR PUF), with the process- and/or bias-dependent voltage and amplification information of two identical amplifier chains serving as the entropy sources. The current-biased PUF cell using only 4 NMOS transistors achieves a small area with reduced temperature and supply sensitivity. Optimization on both the stage gain and stage number can reduce the input-referred noise (IRN) and improve the PUF reliability. We further employ an XOR gate to process the amplifier-chain outputs for the final response to improve the energy efficiency and uniqueness. The process- and bias-dependent stage amplification and the nonlinear amplifier-chain multiplication, which can significantly increase the number of modeling parameters and introduce a complex decision boundary respectively, can effectively resist machine learning (ML) modeling attacks. Fabricated in standard 65nm CMOS, the proposed AC-XOR PUF occupies an active area of 6845\mu \text{m}^{2} . Without discarding any challenge-response pairs (CRPs), this work features a measured worst case bit error rate (BER) of 5.70% across 1.06\sim 1.55V and - 30\sim 125^{\circ }\text{C} , while demonstrating a reliability (intra-die HD) and uniqueness (inter-die HD) of 0.58% and 49.92%, respectively. It also achieves a ML prediction accuracy of 50.72% using 80\times 80\times 80 artificial neural network (ANN) with 1M CPRs as training set.