李刚2,,
1.温州理工学院数据科学与人工智能学院 温州 325035
2.温州大学电气与电子工程学院 温州 325035
基金项目:国家重点研发计划(2018YFB2202100)，国家自然科学基金(61874078, 61904125)，温州市基础性科研项目(G20190006, G20190003)
详细信息
作者简介:马雪娇：女，1991年生，助教，研究方向为物理不可克隆函数攻击与防御
李刚：男，1988年生，讲师，研究方向为密码芯片攻击防御理论及VLSI实现
通讯作者:李刚 ligang@wzu.edu.cn
中图分类号:TN918; TP331计量
文章访问数:244
HTML全文浏览量:96
PDF下载量:34
被引次数:0
出版历程
收稿日期:2021-06-22
修回日期:2021-08-12
网络出版日期:2021-08-23
刊出日期:2021-09-16
ANN Feature Vector Extraction Based Attack Method for Flip-Flop Based Arbiter Physical Unclonable Function
Xuejiao MA1,Gang LI2,,
1. School of Data Science and Artificial Intelligence, Wenzhou University of Technology, Wenzhou 325035, China
2. College of Electrical and Electronic Engineering, Wenzhou University, Wenzhou 325035, China
Funds:The National Key Research and Development Program of China (2018YFB2202100), The National Natural Science Foundation of China (61874078, 61904125), The Wenzhou Basic Scientific Research Projects (G20190006, G20190003)
摘要
摘要:为评估物理不可克隆函数(PUF)的安全性，需针对不同的PUF结构设计相应的攻击方法。该文通过对强PUF电路结构和工作机理的研究，利用人工神经网络(ANN)提出一种针对触发器-仲裁器物理不可克隆函数(FF-APUF)的有效攻击方法。首先，根据FF-APUF电路结构，利用多维数组构建电路延时模型；然后，对FF-APUF的二进制激励进行邻位划分，将划分后的激励转换为十进制并表示为行向量，实现特征向量提取；最后，基于提取的特征向量利用ANN构建攻击模型并通过后向传播算法获得最优参数。实验结果表明，相同条件下攻击预测率均高于其他3种常用的机器学习方法，尤其当激励响应对(CRP)数量较少、激励位数较多时，优势更加明显。当激励位数为128、CRP个数为100和500时，平均攻击预测率分别提高36.0%和16.1%。此外，该方法具有良好的鲁棒性和可扩展性，不同噪声系数下攻击预测率与可靠性相差最大仅0.32%。
关键词:物理不可克隆函数/
触发器-仲裁器物理不可克隆函数/
人工神经网络/
特征向量提取
Abstract:In order to evaluate the security of Physical Unclonable Function (PUF), it is necessary to put forward corresponding attack methods for different PUF structures. By studying the structure and working mechanism of Flip-Flop based Arbiter Physical Unclonable Function (FF-APUF), an effective attack method against FF-APUF is proposed based on Artificial Neural Network (ANN) in this paper. Firstly, according to the circuit structure, the delay model of FF-APUF is established by using multidimensional array. Secondly, all binary challenge bits are divided by two adjacent bits which are converted to a decimal, and then the challenges are expressed as a row vector to extract the feature vector. Finally, based on the extracted feature vectors, the attack model is constructed by ANN, and the optimal parameters are obtained by back propagation algorithm. The experimental results show that the prediction accuracy of the proposed method is higher than other three common machine learning methods under the same conditions. The attack advantage is more obvious, especially when the number of Challenge Response Pairs (CRP) is less and the bit number of challenges is large. For example, when the number of challenge bit is 128, and the number of CRPs is 100 and 500, the average attack prediction accuracy increased by 36.0% and 16.1% respectively. In addition, the proposed method has good robustness and scalability, and the maximum difference of attack prediction rate and reliability is only 0.32% under different noise.
Key words:Physical Unclonable Function (PUF)/
Flip-Flop based Arbiter Physical Unclonable Function (FF-APUF)/
Artificial Neural Network (ANN)/
Feature vector extraction
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=85c1ff8c-6a36-475e-90d2-4ee6bd9648c5
