Hardware Trojan Detection Based On Supervised Learning

In last several years,cloud computing,artificial intelligence and Internet of things are changing with each passing day.Chips have been widely used in computing,automotive,defense,aerospace,medical,telecommunications,networking,household appliances,portable devices,and wireless applications.Chips have become a country’s basic strategic resource.Due to the increasing complexity of hardware design,in order to reduce the burden on hardware designers,third-party hardware intellectual property(IP)modules are usually introduced.However,the widespread use of third-party IP has caused security problems,such as hardware Trojan(HT)inserted by attackers.At present,there is no single detection method that can completely solve this problem.This paper aims at using the Ring Oscillation Network(RON)to detect the shortcomings of the HT method with lower accuracy and higher FPR,and combined with supervised learning and integrated learning to improve the hardware Trojan horse detection method based on RON.Firstly,a detection method of ring oscillator network based on NAND structure is proposed.The ring oscillator based on the NOT structure is composed of a NAND gate and an even number of inverters.If all inverters are replaced with NAND,a NAND structure ring oscillator is formed.Each stage of the NAND structure ring oscillator is connected to the power supply voltage,so it is more sensitive to voltage drops than the NOT structure ring oscillator,so it is easier to detect hardware Trojan.The experiment verifies the effectiveness of the method by comparing the detection methods of NOT structure ring oscillator and NAND structure ring oscillator.Then,a detection method based on supervised learning and NAND ring oscillation network is proposed.Select 20 hardware Trojans from Trust-Hub’s benchmark Trojan library,build the required RON structure,and collect data in the experimental environment where these Trojans and ISCAS’89 benchmark circuit are implanted in FPGA.Then use K nearest neighbors(KNN),support vector machines(SVM),Naive Bayesian(NB)and the C4.5classification model based on discretization to classify them.In order to further increase the accuracy and reduce FPR,SMOTE oversampling is used on the data set to optimize the classifier.Ultimately,a detection method based on ensemble learning and NAND ring oscillation network is proposed.The experiment compares different ensemble methods: heterogeneous ensemble,bagging,and mixed use of the two methods,and verifies that ensemble learning has higher accuracy and lower FPR in this classification problem.Experimental results show that the SVM classifier with the highest accuracy rate reaches 99.3%.NB classifier with the lowest FPR value is only 1.3%.KSNC classifier with better overall performance was also obtained,with an accuracy rate of 99.1% and an FPR of 2.1%.