Research On Hardware Trojan Detection Based On RTL Feature Extraction And Machine Learning

At present,the global IC(Integrated Circuit)industry is entering a period of disruptive technological innovation,and China's IC industry also has ushered in significant development opportunities.However,with the increasing complexity of IC design,more and more design manufacturers have started to adopt third-party IP(Intellectual Property)cores to shorten the development cycle and reduce the cost of chip design.But at the same time,it has also brought security risks.If the third-party IP core used during the chip design is embedded with an HT with malicious functions,it will have a serious impact on many pivotal areas including finance,national defense and so on.Therefore,it is quite important to carry out research on HT detection for third-party IP cores.This paper proposes an HT detection method based on RTL(Register Transfer Level)feature extraction and machine learning for third-party IP cores by analyzing the characteristics of Verilog codes and combining machine learning algorithms.The steps are as follows:First of all,in order to analyze the Verilog codes more conveniently and accurately,this paper proposes a set of node division criteria,which can be used to refine the Verilog codes into nodes that are statement blocks with no other branch structure inside for subsequent analysis;Secondly,according to the synthesizable Verilog syntax,this paper defines the RTL features of the node from the perspective of the signals and operations performed by the statements of the node.This paper also proposes the calculation method of each feature,so as to quantify the nodes into a numerical model for mathematical analysis.Then,node division and feature extraction are performed on several circuits about one type of IP core with HTs,and obtain RTL feature data of all nodes divided.Then the nodes' RTL feature data are combined with SVM or Random Forest algorithms to train the HT classification of this type of IP core.Finally,the best classifier through machine learning training can be implemented to detect RTL HTs for corresponding types of third-party IP cores.The HT detection method proposed in this paper is implemented based on Python and it is performed on AES-128 and RS232 circuits with HTs of Trust-Hub to conduct test verification.Two machine learning algorithms are applied respectively which are SVM and Random Forest.The results show that the HT detection method proposed in this paper has achieved the accuracy of99% and 92.4% for the AES-128 circuit and RS232 circuit by using SVM,and it has also achieved the accuracy of 100% and 91.8% for the AES-128 circuit and RS232 circuit by using Random Forest.So it can be proved that the HT detection method based on RTL feature extraction and machine learning proposed in this paper is an effective method,which is suitable of HT detection for third-party IP cores.