Gate-level Netlist Hardware Trojan Detection Based On Machine Learning

With modernization and automation,the role of IC in the control and operation of systems has become a critical component.In recent years,the semiconductor industry has enthusiastically adopted the fabless model of chip fabrication i.e.fabricating designs through third party vendors,that is,attackers will use this vulnerability to implant hardware Trojan to destroy hardware security,and this method is concealed and intrusive.Therefore,it is difficult to guarantee that the manufactured products are 100% safe.In this context,how to ensure that the security of the underlying hardware is not compromised has become a hot topic.More and more researchers have begun to invest in the research of hardware Trojan and have proposed many detection methods,mainly including destructive engineering and side channels and logic activation.However,these detection methods have disadvantages,such as unsatisfactory results for large-scale circuit detection,extremely high requirements for detection equipment,easy to be affected by some process noise,and requiring original circuits for comparison.Aiming at the problems in the conventional hardware Trojan detection methods,this paper combines the net features of the gate-level netlist in the hardware circuit with different machine learning classification methods,and proposes a gate-level netlist hardware Trojan detection method based on machine learning.This method classifies the normal net and Trojan net in the circuit according to the features of the net,and transforms the problem of detecting hardware Trojan in the circuit to be tested into a two-class problem of the net in the circuit to be tested.Firstly,focus on the detection of the hardware Trojan at the gate-level netlist stage,select different benchmark test circuits,and extract the feature data set by analyzing the different features between the Trojan net and the normal net in the circuit.Then,according to the size of the data set,different machine learning methods are used.Finally,the classification model is trained based on the feature data set,and the net that needs to be detected is classified to achieve the purpose of detecting hardware Trojan.The experimental results show that the hardware Trojan can be better detected by combining the features of the hardware Trojan net with the machine learning method,and the recognition accuracy of the net in the circuit can be improved.The research content of the thesis includes the following three aspects:(1)Explain the related concepts,basic structure and main classifications of hardware Trojan.Study conventional hardware Trojan detection methods,and summarize the advantages and disadvantages of each detection method.Analyze the different net features of the hardware Trojan circuit and the normal circuit in the gate-level netlist stage of the IC to determine the research direction.(2)Aiming at the small sample circuit data,a detection method based on PSO-SVM algorithm is proposed.Firstly,according to the selected reference circuit,analyze the net features of the Trojan circuit and the normal circuit,and extract the 8-dimensional feature vector.Secondly,use the SMOTE oversampling algorithm to preprocess the extracted feature data set to improve the imbalance problem of the data set category.Next,use the PSO optimization algorithm to find the optimal parameters of the SVM to obtain a higher recognition accuracy.Finally,use the classifier to detect the circuit to be tested and identify the hardware Trojan.The final experimental results show that the use of PSO-SVM can better solve the problem of identifying hardware Trojan in small data samples and improve the accuracy of the overall circuit net recognition.(3)Aiming at the large-scale circuit data,a detection method based on the deep forest algorithm is proposed.Firstly,according to the gate-level netlist features of the reference circuit,increase the net feature to extract the 9-dimensional feature vector.Secondly,increase the multi-granularity scanning process,use different sliding windows to extract the input net feature to generate feature examples of different dimensions to enhance the correlation between the features of the net,and use the feature vector output by the final sliding window as the input vector of the cascade forest.In the cascade forest,the feature information of the input feature vector is sufficient through multiple different types of forest classifiers learn to improve the overall performance of the model.The final experimental results show that the use of the deep forest algorithm can better solve the detection problem of hardware Trojan in big data samples and improve the detection accuracy of the overall circuit net.