Research On Machine Learning Based Hardware Trojan Detection Method

In recent years,hardware Trojan has become an emerging security issue in the field of information security.This thesis first reviews the background of the emergence of hardware Trojans and existing hardware Trojan detection works.Next,three innovative research works are presented.The major works and contributions are as follows:1)Golden chips-free hardware Trojan detection method based on co-training techniqueThis thesis proposes a golden chips-free hardware Trojan detection method based on co-training technique.The proposed method can provide reliable hardware Trojan detection capability for manufactured ICs exploiting inaccurate simulation models and unlabeled manufactured ICs.The proposed method first trains two classification algorithms using simulated ICs in the IC design flow.In the testing phase,the trained two classifiers detect the manufactured unlabeled ICs and then add labels to these ICs.Next,these two classifiers select ICs which pass statistical examination.Then,the selected ICs are added to the training set of the other classifier for further training.After several iterations,the proposed method can improve the detection accuracy steadily.In order to address the issues of process variations and noises,the partial least squares method is used to preprocess the raw data of manufactured ICs.The experimental results show that the proposed method can detect unknown hardware Trojans with high detection accuracy and high hardware Trojan detection rate.2)Golden models-free hardware Trojan detection method based on clustering analysisThis thesis proposes a golden models-free hardware Trojan detection technique based on clustering analysis.The golden models-free hardware Trojan detection problem is formulated into two detection models: partition detection model and anomaly detection model.The mutual information based cluster labeling index is proposed to determine the label of each cluster.Experimental results show that the proposed method can achieve good detection performance in the detection scenario where there is no golden model for reference,which is a difficult task for existing hardware Trojan detection methods.3)A novel clustering ensemble based hardware Trojan detection method against untrustworthy testing partiesThis thesis is the first to consider that untrustworthy testing parties may be untrustworthy.We propose two attack models and an adversarial test data modification method for untrustworthy testing parties.Then,a novel clustering ensemble based hardware Trojan detection method is proposed to defeat untrustworthy testing parties.In the testing phase,each testing party obtains its basic detection results of unlabeled ICs.Then,we exploit the proposed hybrid cluster ensemble method to integrate the basic detection results of each testing party and then obtain the final ensemble detection results.In order to address the issues of process variations and noises,the unsupervised correlation-based feature selection method is exploited to preprocess the raw data of manufactured ICs.In addition,the following key issues are analyzed: the number of necessary testing parties,the time overhead and computational overhead of the proposed method,how to use the proposed diversity analysis method to select the basic clustering algorithms,and the reason why the proposed clustering ensemble method is superior to the majority voting method.The experimental results show that the proposed method can resist the malicious modification of untrustworthy testing parties robustly,and can detect hardware Trojans with high detection accuracy.