Research On Hardware Trojan Detection Method Based On Machine Learning

With the rapid development of information society and the accelerated application of artificial intelligence technology,people's demand for IC chips is increasing dramatically.However,due to the complexity of chip design and manufacturing,chip manufacturers can not fully control each link,which provides the possibility for some attackers to maliciously modify and destroy the integrated circuit.This kind of defective module,which is deliberately manufactured and inserted into the chip by the attacker,changes the function or performance of the chip and triggers under some special conditions,is called Hardware Trojan.Due to the potential threat of hardware Trojans to chip security,scholars at home and abroad have proposed a variety of hardware Trojan detection methods.The detection method based on failure analysis has a high detection accuracy.In this paper,the method based on machine learning is proposed to be used.However,due to the limited conditions,it is impossible to obtain a large number of practical circuit images of the chip for reverse engineering processing required by machine learning training in a short time.In view of this limitation,this paper replaces reverse engineering image with transistor level layout image and gate level network table image,and realizes the detection of transistor level and gate level Hardware Trojan by machine learning image recognition method.Experiments show that the machine learning image detection method has better detection performance for these two types of hardware Trojans,and the feasibility of the failure analysis detection method based on machine learning is verified.The main contents of this paper are as follows:(1)In order to build a transistor level Hardware Trojan model,we studied the Hardware Trojan with modified polarity and extracted the image set of machine learning training verification from it.Firstly,the device modeling and Simulation of this type of hardware Trojan are carried out to verify its correctness;secondly,the Trojan unit library is designed according to the corresponding layout design rules based on smic130 standard unit library;finally,the Trojan model is set up in AES(Advanced Encryption In standard)carrier circuit,156 logical unit images in standard AES and Trojan AES circuit layout are extracted by automatic positioning method to form a verification image set,and 600 logical unit images in standard unit library and Trojan unit library are extracted and processed by similarity to form a training image set.(2)The image recognition methods based on SVM(Support Vector Machine)and convolution neural network are studied.For support vector machine method,LBP feature(Local Binary Pattern),SIFT feature(Scale Invariant Feature Transform),HOG feature(Histogram of Oriented Gradient)of image are extracted respectively,and then the extracted image feature is sent to support vector machine for classification;for convolution neural network method,the network model is constructed and the key parameters are determined through experimental comparison.Finally,the detection efficiency of the two methods on transistor level layout image set is compared,among which convolution neural network is more efficient,the detection time of single layout image is 146 ms,and the detection accuracy is 97.4%.(3)The target detection algorithm of Faster R-CNN(Region based Revolution Neural Networks)is applied to the location detection of gate level netlist image.Firstly,three design strategies are proposed to reduce the trigger probability,power consumption and logic resources of the Trojan horse and control the Trojan horse.After that,the designed Trojan horse is inserted into the iscas89 carrier circuit,and 800 gate level netlist images are extracted to form the training verification image set.Finally,through the fast r-cnn algorithm to the gate level netlist image set training verification,obtains 98.0% Trojan horse detection accuracy and 96.7% Trojan horse classification accuracy.