Research On Hardware Trojan Detection Approach Based On Gate-Level Circuit Features

With the globalisation of the semiconductor industry,hardware Trojan has become a serious threat to the security of integrated circuits.Hardware Trojan detection technology has become a hot research topic nowadays.Among many detection approaches,hardware Trojan detection technique based on feature recognition is widely concerned.However,existing research results still have some problems,such as low number of hardware Trojan features and poor performance of classification models.Based on this,this paper carries out research on hardware Trojan detection methods based on gate-level circuit features and machine learning.By analysing the gate-level features of hardware Trojans,a comprehensive hardware Trojan feature library is established,and the performance of the classification model is enhanced by using machine learning algorithms.The major work and innovations are as follows:1.Based on the gate-level netlists in the Trust-Hub hardware Trojan library,the gate-level features of hardware Trojan are studied in terms of signal concealment,trigger rarity and malicious functionality.6 signal features and 17 structure features are summarised and proposed to establish a comprehensive hardware Trojan feature library and lay the foundation for the research of detection algorithms.2.Aiming at the signal features of hardware Trojan,this paper proposes an unsupervised hardware Trojan detection method based on KM-LOF.By extracting the testability values of the gate-level netlists,a four-dimensional signal feature vector is established,and then the distribution characteristics of hardware Trojans in the feature space are analysed.The k-means clustering algorithm is applied to cluster the overall signal samples to complete the primary screening of Trojan signals.The screening results are optimised based on the LOF algorithm.The hardware Trojan detection process based on signal features and KM-LOF is established.14 SAED benchmark circuits are used to validate the efficacy of the proposed approach.The results show that the method can achieve effective identification of strongly hidden hardware Trojans.Average hardware Trojan detection rate achieves 99.72%.3.Aiming at the problem of inconspicuous signal features of weakly hidden hardware Trojans and the low detection rate of hardware Trojan in existing methods,this paper proposes a hardware Trojan detection method based on hybrid features and PSO-SVM.Combining the signal features and some structure features,a hardware Trojan hybrid feature set is established.The BorderlineSMOTE algorithm is applied to expand the Trojan samples in the feature dataset to form a balanced training dataset.The performance of the hardware Trojan classification model is enhanced by optimising the SVM model parameters based on the PSO algorithm.17 SAED benchmark circuits are used to validate the efficacy of the proposed approach.The results show that the method is compatible with the detection of both strongly and weakly hidden hardware Trojans.The detection rate of hardware Trojans has been further improved.4.Aiming at the problems of redundant features in the hardware Trojan feature library,as well as the low accuracy and weak generalization ability of existing classification models,this paper proposes a hardware Trojan detection method based on feature selection and Ada Boost.Based on the principle of mutual information,the necessity of hardware Trojan feature selection is analysed.A wrapped hardware Trojan feature selection algorithm based on CFS and model evaluation is proposed to obtain the optimal hardware Trojan feature set.The Ada Boost algorithm is applied to build an integrated hardware Trojan classifier.The hardware Trojan detection architecture based on feature selection and Ada Boost is formed.22 SAED benchmark circuits and60 LEDA benchmark circuits are used to validate the efficacy of the proposed approach.The results show that the method is applicable to different types and sizes of hardware Trojans.The comprehensive performance of the classification model is further improved.