摘要：

物理不可克隆函数(PUF)在资源受限的信息安全领域起着至关重要的作用，然而广泛使用的仲裁器PUF(APUF)及其变体因结构简单和防御维度单一，面临机器学习建模攻击的威胁，同时具有高防御能力的PUF设计通常伴随着较高的硬件成本。为应对这些挑战，提出一种新型的多维协同PUF(MS PUF)设计，旨在平衡强大的抗建模攻击能力和低硬件开销。该设计以APUF为基础，融合了弱PUF、线性反馈移位寄存器(LFSR)和多路复用器(MUX)，通过异或操作混淆输入信号并动态控制MUX输出，增强了PUF响应的安全性和不可预测性。在此设计中，MUX的输出有两种选择：一是直接采用弱PUF序列，二是通过分组异或处理并采用由弱PUF初始化的LFSR生成的序列。此外，MS PUF通过引入逐层异或混淆机制，构筑了一个多层次、多维度的协同安全防御策略。实验结果表明，MS PUF在均匀性、唯一性和可靠性等关键性能指标上表现优异，且硬件开销低，在防御逻辑回归(LR)、支持向量机(SVM)、人工神经网络(ANN)、卷积神经网络(CNN)以及全连接长短时记忆(FC-LSTM)网络等多种机器学习建模攻击时，MS PUF的预测准确率均接近50%，展示了出色的防御能力。

关键词: 仲裁器物理不可克隆函数, 机器学习建模攻击, 硬件开销, 多维协同PUF, 逐层异或混淆机制

Abstract:

Physically Unclonable Functions (PUFs) play a crucial role in the domain of information security with limited resources. However, the widely used Arbiter PUF (APUF) and its variants are threatened by machine learning modeling attacks because of their simple structure and singular defense dimensions. Moreover, PUF designs with high defensive capabilities are usually accompanied by high hardware costs. To address these challenges, a novel Multi-dimensional Synergistic PUF (MS PUF) design is proposed in this study, aiming to balance the strong resistance against machine-learning modeling attacks with low hardware overhead. This design is based on the APUF and incorporates weak PUFs, a Linear Feedback Shift Register (LFSR), and a Multiplexer (MUX), enhancing the security and unpredictability of PUF responses through Exclusive OR (XOR) operations to obfuscate input signals and dynamically control the MUX output. In this design, the output of the MUX has two options: one is the direct use of weak PUF sequences, and the other is the use of sequences generated by the LFSR initialized by weak PUFs after grouped XOR processing. Moreover, by introducing a layered XOR confusion mechanism, a multi-level, multi-dimensional, and synergistic security defense strategy has been developed. Experimental results indicate that the MS PUF exhibits outstanding performance in key performance indicators such as uniformity, uniqueness, and reliability, with a low hardware overhead. The prediction accuracy of the MS PUF approaches 50% when defending against a variety of machine-learning modeling attacks, such as Logistic Regression (LR), Support Vector Machine (SVM), Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Fully Connected Long Short-Term Memory (FC-LSTM) network, demonstrating its excellent defense capability.

Key words: Arbiter Physical Unclonable Function(APUF), machine learning modeling attack, hardware overhead, Multi-dimensional Synergistic PUF(MS PUF), layered Exclusive OR(XOR)confusion mechanism