Research On Hardware Trojan Detection Technology Based On Convolutional Neural Network

In recent years,the application of integrated circuits has become more widespread and has penetrated all aspects of people's lives.In the R&D process of integrated circuits,it is inevitable to use designs and third-party services in different countries or regions.Attackers use vulnerabilities in the industry chain as the entry point for attacks and use Hardware Trojan to achieve the purpose.Hardware Trojan not only bring hidden dangers to the civilian field but also threaten national defense and national security.Therefore,it is of practical significance to carry out research on hardware Trojan detection technology and accurately locate the type of hardware Trojan.At present,there is no unified standard for hardware Trojan detection.Different types of hardware Trojan horses use different detection technologies.When detecting multiple hardware Trojan,the detection time is long and consumes manpower,material,and financial resources.In response to the above problems,this paper designs a hardware Trojan detection model based on convolutional neural networks and long short-term memory networks.First,use the NLTK word segmentation tool to preprocess the hardware Trojan samples,and use the generative adversarial network to generate normal samples to form a data set.Then design a vector representation method based on character embedding and word embedding,reduce the dimensionality,convert each sample document into a two-dimensional feature matrix.Finally,combining the advantages of convolutional neural networks and long short-term memory networks,a detection model combining the two networks is designed.Through training and learning the extracted feature files,the detection of hardware Trojan is realized Experiments show that the designed hardware Trojan detection model can achieve a detection accuracy of 82.108%.At the same time,the parameters of the detection model are dynamically adjusted to realize multiple classifications of hardware Trojans.The main contributions of this article are as follows:(1)Sample preprocessing: This paper writes a python script file to process the collected original hardware Trojan samples,and delete the redundant information in the samples,and then use the NLTK word segmentation tool to segment the data set,write regular expressions to retain the special symbols in the document,and form a preprocessed document.(2)Vector conversion representation: The hardware Trojan horse samples can only be input into the neural network for training and learning when they are converted into a twodimensional matrix.This paper uses a vector conversion method based on the combination of character embedding and word embedding to convert the pre-processed hardware Trojan samples into vector representations.Compared with the ordinary character embedding method,it can extract the key information between phrases;compared with the ordinary word embedding method,the special symbols in the sample can also be retained.Use the character-level information in the sample to construct the word embedding vector,which can effectively represent the rare phrases in the sample and the symbols and words that did not appear during training.Each hardware Trojan sample can be expressed in a suitable vector form,which can reflect the characteristic information contained in the hardware Trojan sample to a greater extent.(3)Hardware Trojan detection: Because of the different detection methods of different hardware Trojan and the problem of high resource consumption,this paper designs a hardware Trojan detection model based on convolutional neural networks and long shortterm memory networks by analyzing the advantages of convolutional neural networks and long and short-term memory networks(HT-CNN+LSTM),in which the convolutional neural network can better extract the local features of the hardware Trojan,the long short-term neural network connects the context,and combines the semantics to extract the characteristics of the hardware Trojan sample to realize the detection of the hardware Trojan sample.(4)Multi-classification of hardware Trojans: Based on the detection model of hardware Trojans,this paper dynamically adjusts the HT-CNN+LSTM model parameters,designs the multi-classification model of hardware Trojans,and realizes the classification of seven kinds of hardware Trojans,which can achieve 91.584% accuracy.