Machine Learning Assisted Counterfeit IC Detection through Non-destructive Infrared (IR) Spectroscopy Material Characterization
Machine Learning Assisted Counterfeit IC Detection through Non-destructive Infrared (IR) Spectroscopy Material Characterization
复制标题
DOI：
10.1109/ectc51906.2022.00355
复制
发表时间：
2022-05
期刊：
影响因子：
--
通讯作者：
Chengjie Xi;Nathan Jessurun;John True;Aslam A. Khan;M. Tehranipoor;Navid Asadizanjani
中科院分区：
文献类型：
--
作者：
Chengjie Xi;Nathan Jessurun;John True;Aslam A. Khan;M. Tehranipoor;Navid Asadizanjani
如今，由于供应链全球化的不断发展，假冒集成电路（IC）越来越普遍。这种供应链漏洞导致不可靠且不安全的假冒 IC 集成到许多应用（包括消费者、企业和军事领域）的最终用户设备中。人们已经开发出各种方法，例如老化检测传感器、物理不可克隆功能 (PUF) 和硬件计量，以便在此类假冒产品集成到关键系统之前对其进行检测。然而，检测和预防的几个复杂方面限制了它们作为解决假冒问题的权宜之计。因此，迫切需要新颖的检查和保证技术，这些技术需要对设备电路或材料进行最少或不需要额外的改变/修改，同时保持每个样品的低成本。本文通过对假冒和正品 IC 进行初步材料调查，证明了使用 IC 封装材料表征进行假冒检测的可能性。本研究使用漫反射红外傅里叶变换光谱（DRIFT）作为材料表征，材料光谱将用于训练机器学习分类模型。将测试多种机器学习分类方法，例如线性判别分析（LDA）、支持向量机（SVM）、随机森林（RF）和多层感知器（MLP）。借助标准正态变量 (SNV) 数据预处理和 MPL 模型，伪造品与正品样本辨别准确率已达到 92% 以上。这证明了假冒和正品IC样品之间存在封装材料差异。
Nowadays, counterfeit integrated circuits (IC) are increasingly common due to the continuous growth of supply chain globalization. This supply chain vulnerability results in unreliable and insecure counterfeit ICs integrated into the end-user devices in many applications, including consumer, corporate, and military domains. Various methods such as aging detection sensors, Physical Unclonable Functions (PUFs), and hardware metering have been developed to detect such counterfeits before they become integrated into critical systems. However, several complicated aspects of detection and prevention limit their use as a stopgap to the counterfeit problem. Hence, there is a critical need for novel inspection and assurance techniques that require minimal or no additional changes/modifications to the device circuit or material while remaining low-cost per sample. In this paper, the possibility of using IC packaging material characterization for counterfeit detection is proved by a preliminary material survey between the counterfeit and authentic ICs. Diffuse Reflectance Infrared Fourier Transform Spectroscopy (DRIFT) is used as the material characterization in this research, and the material spectrums will be utilized for training machine learning classification models. Several machine learning classification methods will be tested, such as Linear discriminant analysis (LDA), Support Vector Machine (SVM), random forest(RF), and multi-layer perceptron (MLP). With the help of the Standard Normal Variate (SNV) data preprocessing and MPL model, over 92 percent accuracy of counterfeit versus genuine sample discrimination has been achieved. This proves the existence of packaging material differences between counterfeit and authentic IC samples.