CN114626106A

CN114626106A - Hardware Trojan horse detection method based on cascade structure characteristics

Info

Publication number: CN114626106A
Application number: CN202210159164.3A
Authority: CN
Inventors: 刘鸿瑾; 陈嘉伟; 张绍林; 施博; 李天文; 李宾; 王巧凤; 王红霞; 白星; 李亚妮; 李康; 史江义
Original assignee: Beijing Sunwise Space Technology Ltd
Current assignee: Beijing Sunwise Space Technology Ltd
Priority date: 2022-02-21
Filing date: 2022-02-21
Publication date: 2022-06-14

Abstract

The invention discloses a hardware Trojan horse detection method based on cascade structure characteristics, relates to the technical field of hardware safety, and solves the problems of high false judgment rate and complex work of the existing hardware Trojan horse detection technology, and comprises the following steps: step S1: extracting a characteristic vector of the cascade structure characteristic of the hardware Trojan horse circuit from a gate-level netlist file based on Verilog; step S2: constructing and training a neural network model; step S3: using the trained neural network model to perform Trojan horse detection; the method has the advantage of high Trojan horse detection accuracy of the hardware circuit.

Description

Hardware Trojan horse detection method based on cascade structure characteristics

Technical Field

The invention relates to the technical field of hardware safety, in particular to the technical field of hardware Trojan horse detection methods based on cascade structure characteristics.

Background

Recently, with the globalization of the integrated circuit industry, hardware security has become an urgent issue. Typically, to reduce development costs and shorten the marking time, developers will use intellectual property cores and EDA tools provided by third parties. However, the services provided by third party providers are not necessarily reliable. A malicious vendor may implant a hardware trojan in their IP. The existence of the hardware trojan can cause the integrated circuit chip to have wrong functions, reduce the performance, leak confidential information and even be damaged. A hardware trojan is typically composed of a trigger and a payload. When the trigger is activated, the payload will perform a malicious function. The hardware trojan is generally designed at a node with low triggering probability in order to avoid detection. This makes it difficult to detect a hardware trojan in the circuit.

In recent years, machine learning is widely applied to the Trojan detection of the gate-level netlist, wherein the more common machine learning models are: SVM, K-means, random forest, etc. Although the accuracy of the Trojan detection of the gate-level netlist is greatly improved by the addition of the machine learning, the missing detection rate and the misjudgment rate of the traditional machine learning model are still high. The problems and defects of the prior art are as follows: in the prior art, the missing detection rate or the false judgment rate of the hardware Trojan horse detection technology is too high, namely when a Trojan horse circuit is completely detected, many normal circuits can be judged by mistake; when all normal circuits are detected as normal, many trojan circuits are missed; the existing Trojan horse detection method based on the neural network is large in required feature quantity, and the feature extraction work is complicated.

The method solves the problems and defects in the prior art, and has certain difficulty because the hardware Trojan horse circuit has extremely small scale and extremely low triggering rate, avoids the traditional detection method by using the node with small triggering probability, and hides the node in a normal circuit by using the advantage of small scale; in addition, the requirement of the machine learning method on the characteristics is high, the same model is trained by using different characteristics, a great effect can be achieved, and the characteristics are selected to accord with the characteristics of both the model and the hardware Trojan horse.

Disclosure of Invention

The invention aims to: the problem of current hardware Trojan horse detection technology misjudgment rate height and work loaded down with trivial details is solved. In order to solve the technical problem, the invention provides a hardware Trojan horse detection method based on cascade structure characteristics.

The invention specifically adopts the following technical scheme for realizing the purpose:

a hardware Trojan horse detection method based on cascade structure characteristics comprises the following steps:

step S1: extracting a characteristic vector of a cascade structure of a hardware Trojan horse circuit from a gate-level netlist file based on Verilog;

step S2: constructing and training a neural network model;

step S3: and (5) using the trained neural network model to perform Trojan detection.

Preferably, in step S1, the extracting the feature vector of the cascade structure of the hardware trojan horse circuit includes:

classifying gate-level modules of the netlist according to functions, and setting the size of the feature vector according to the number of classified types;

and traversing through a depth-first search algorithm to obtain the feature vector of the cascade structure.

Preferably, the classifying the gate-level modules of the netlist according to functions and setting the size of the feature vector according to the number of classified types includes:

classifying the gate-level modules according to the functional information of the gates to obtain the classification types of the gate-level modules;

combining the classifications in pairs to obtain a static structure characteristic type;

and setting the size of the feature vector according to the number of the feature types of the static structure.

Preferably, the method for obtaining the feature vector of the cascade structure through traversal by the depth-first search algorithm includes:

step S401: initializing, selecting a target gate, setting the target gate as an access gate, and setting the depth as an initial value 0;

step S402: moving to an unaccessed adjacent door of the access door, wherein the adjacent door becomes a new access door, adding 1 to the depth, counting the connection type, and adding 1 to the number of the connection type vectors at the corresponding positions;

step S403: judging whether the access door has an adjacent door which is not accessed, if so, returning to the step S402; if not, the maximum depth is judged to be reached, the previous gate is returned and serves as an access gate, the depth is reduced by 1, then whether the access gate has an adjacent gate which is not accessed is judged, if yes, the step S402 is returned, and if not, the process is ended to obtain the feature vector of the target gate.

Preferably, in the step S2, the neural network model includes a fully-connected neural network and a decision tree;

the fully-connected neural network comprises an input layer, a hidden layer and an output layer; the input layer is used for inputting the feature vectors into a neural network; the hidden layer and the output layer are used for acquiring and outputting a calculation result;

and the decision tree is used for judging whether the gate-level netlist contains the Trojan horse or not according to the calculation result.

Preferably, in step S3, the performing Trojan horse detection using the trained neural network model includes:

performing feature extraction on the netlist to be processed to obtain a feature vector of the whole netlist;

putting the feature vectors into the trained neural network model one by one to obtain model output;

and (4) detecting the output result of the model, judging that the Trojan horse is contained if the output result is 1, and judging that the Trojan horse is normal if the output result is 0.

The invention has the following beneficial effects:

the performance of hardware Trojan horse detection is improved, and the safety of integrated circuit chip design is enhanced; a new idea is provided for the application of the neural network in the field of hardware Trojan horse detection; the feature extraction method is very simple and can be realized only by the most basic depth-first search, the feature vector contains the gate-level connection structure features of the circuit near the target gate, the gate-level fan-in and fan-out information is abandoned, but the fan-in and fan-out information is contained in the feature vector in a phase-change manner due to the small search depth; the characteristic method of the invention omits the redundant information of other characteristic methods, so that a better neural network model is obtained under the condition that the input characteristic data is smaller; the feature extraction method can be used for extracting and detecting features of any gate-level module of the gate-level netlist, so that the module position of the Trojan horse circuit in the gate-level netlist can be determined.

Drawings

FIG. 1 is a schematic flow diagram of the present invention;

FIG. 2 is a symbolic illustration of a gate level module;

FIG. 3 is a schematic diagram of a flip-flop circuit of embodiment 1;

fig. 4 is a schematic diagram of the neural network of embodiment 1.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. The components of embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.

Thus, the following detailed description of the embodiments of the present invention, presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Example 1

As shown in fig. 1, the present embodiment provides a hardware Trojan horse detection method based on a cascade structure feature, including the following steps:

in the step, the local cascade structure characteristics of the gate-level netlist can be extracted, and the structure characteristic information is changed into a characteristic vector form, so that the training and learning of the neural network are facilitated.

Step S2: constructing and training a neural network model;

through the steps, a neural network model can be constructed, then the neural network model is trained through the known Trojan horse netlist, and then each weight value in the neural network model is determined, so that the neural network model can distinguish the characteristics in the hardware Trojan horse feature vector, and the trained model can detect the hardware Trojan horse on the unknown netlist.

In this embodiment, preferably, in the step S1, the extracting the feature vector of the cascade structure of the hardware trojan horse circuit includes:

As a preferred scheme of this embodiment, the classifying the gate-level modules of the netlist according to functions and setting the size of the feature vector according to the number of types of classification includes:

In the present embodiment, the gate level module types are classified into 14 types according to the function information of the gate, specifically referring to table 1, and the symbolic expression of the commonly used gate level module refers to fig. 2;

TABLE 1 netlist gate level module functional classification

Then, after the classification, the static structure feature types are expressed, such as NOR-NOR, AND-NOR AND AND-OR, namely the static structure feature types are combined in pairs in the table 1, so that the static structure feature types are 196 types in total;

the size of the feature vector is set to 196, which represents the number of elements in the feature vector.

In addition, the method for obtaining the feature vector of the cascade structure through traversal by the depth-first search algorithm does not include:

Taking the trigger circuit of HT in s35932-T100 as an example, as shown in FIG. 3. Assuming that the netlist gate-level modules are classified into only four types of AND, NOR, INV AND OTHER according to the functional classification, the size of the feature vector is set to be 16. The maximum depth of the depth-first search algorithm is set to 2, a nor gate in the circuit diagram is taken as a target gate, a feature vector of the gate is extracted, and the feature vector is shown in table 2.

Let the feature vector of a single sample be x ═ x₁,x₂,x₃,…,x_n]Wherein x is_iRepresents the number corresponding to a certain feature type of the sample, and n represents the number of the feature types, i.e. the size of the feature vector. The feature vector of table two can be expressed as x ═ 0,8,0,0,2,0,0,0,1,0,0,0,0,0]。

TABLE 2 feature vectors

Type of feature	Feature vector
		AND-AND	0
AND-NOR	8
		AND-INV	0
AND-OTHER	0
		NOR-AND	2
NOR-NOR	0
		NOR-INV	0
NOR-OTHRE	0
		INV-AND	1
INV-NOR	0
		INV-INV	0
INV-OTHER	0
		OTHER-AND	0
OTHER-NOR	0
		OTHER-INV	0
OTHER-OTHER	0

Referring to fig. 4, as a preferred solution, in the step S2, the constructing and training the neural network model includes:

in step S2, the neural network model includes a fully-connected neural network and a decision tree;

the fully-connected neural network comprises an input layer, a hidden layer and an output layer; the fully-connected neural network comprises an input layer, a hidden layer and an output layer; the input layer is used for inputting the feature vectors into a neural network; the hidden layer and the output layer are used for acquiring and outputting a calculation result; and the decision tree is used for judging whether the gate-level netlist contains the Trojan horse or not according to the calculation result. The forward propagation calculation process is as follows:

c＝(y_j·_j ⁽²⁾)+b⁽²⁾,j∈(1,2,3,…,n)

y_o＝Sigmoid(c)

wherein

Is the weight of the hidden layer or layers,

for hiding the bias value of the layer, W_j ⁽²⁾As a weight of the output layer, b⁽²⁾Is the bias value of the output layer. The training of these values is achieved by back propagation.

Since the Sigmoid function is:

it can be seen that since the output of Sigmoid is (0, 1), the calculation result y_oThe value of (A) can be taken as the probability that the input sample is a Trojan circuit, i.e., y_oThe closer the value of (d) is to 1, the greater the probability that the sample is a Trojan horse.

After the probability is obtained, a threshold value needs to be determined to classify the sample. This is done here using a decision tree. Will y_oAnd inputting the input into a decision tree, and classifying the decision tree to obtain output y. y can only be equal to 0 or 1,0 representing the gate as a normal circuit; 1 represents the gate as a trojan circuit.

The model training process of this embodiment may be as follows:

step one, using several known Trojan netlist, namely netlist of known Trojan position, extracting characteristics of the Trojan netlist to obtain samples, wherein Trojan label is 1, normal label is 0, and sample balance processing is carried out on the samples to ensure that the proportion of the Trojan sample and the normal sample is approximate; training by using the extracted samples, so that the neural network can output a probability result close to 1 when encountering the Trojan horse sample; and step three, forming a new sample by using the probability result output by the trained full-connection layer network and the corresponding label, and then training the decision tree by using the new sample, so that a proper threshold value is trained, and the Trojan horse sample and the normal sample are distinguished. The output of the decision tree is 0 or 1,0 is a normal circuit, and 1 is a Trojan circuit.

Finally, in step S3, the performing the Trojan horse detection using the trained neural network model may include:

Taking the reference circuit of Trust-HUB as an example, 15 reference circuits are taken as tests from Trust-HUB, and the reference circuit information is shown in the following table 3:

TABLE 3 Trust-HUB reference Circuit information

Netlist name	Normal circuit	Trojan horse circuit
			RS232-T1000	202	13
RS232-T1100	204	12
			RS232-T1200	202	14
RS232-T1300	204	9
			RS232-T1400	202	13
RS232-T1500	202	14
			RS232-T1600	202	12
s15850-T100	2155	27
			s35932-T100	5426	15
s35932-T200	5422	16
			s35932-T300	5426	36
s38417-T100	5329	12
			s38417-T200	5329	15
s38417-T300	5329	44
			s38584-T100	6473	9

Using 14 netlists as training set and 1 netlist as test set. The true positive rate TPR and the true negative rate TNR were used as evaluation indices. The TPR and TNR are calculated as follows:

the classification results can be classified into true negative TN, false positive FP, false negative FN and true positive TP. TN is the number of normal circuits correctly identified as normal circuits. FP is the number of normal circuits that are misidentified as trojan circuits. FN is the number of trojan circuits that are misidentified as normal circuits. TP is the number of correctly identified Trojan horse circuits. The calculation formulas of the true positive rate TPR and the true negative rate TNR are respectively as follows:

TPR＝TP/(TP+FN)；

TNR＝TN/(TN+FP)。

this example uses the Keras library of python to construct a model. The model is trained using 14 of the netlists as a training set. The trained model is tested with the remaining 1 netlist as a test set. The final experimental results are shown in table 4:

TABLE 4 results of the experiment

Netlist names	TPR	TNR
			RS232-T1000	100	99.5
RS232-T1100	100	100
			RS232-T1200	100	100
RS232-T1300	100	99.5
			RS232-T1400	100	99.5
RS232-T1500	100	100
			RS232-T1600	100	99
s15850-T100	100	91.1
			s35932-T100	93.3	100
s35932-T200	87.5	100
			s35932-T300	94.4	100
s38417-T100	83.3	100
			s38417-T200	73.3	100
s38417-T300	86.4	100
			s38584-T100	66.7	95.3
Mean value of	92.3	98.9

As can be seen from Table 4, the TPR of the hardware Trojan horse detection method of the present embodiment is between 66.7% and 100%, and the TNR is between 91.1% and 100%. The mean TPR was 92.3% and the mean TNR was 98.9%. On the premise of repeated training of the neural network model, the method has good hardware Trojan horse detection capability.

Claims

1. A hardware Trojan horse detection method based on cascade structure characteristics is characterized by comprising the following steps:

step S1: extracting a characteristic vector of the cascade structure characteristic of the hardware Trojan horse circuit from a gate-level netlist file based on Verilog;

step S2: constructing and training a neural network model;

2. The method according to claim 1, wherein in step S1, the extracting the feature vector of the cascaded structure of the hardware Trojan horse circuit comprises:

3. The hardware Trojan horse detection method based on the cascade structure features as claimed in claim 2, wherein the step of classifying the gate-level modules of the netlist according to functions and setting the size of the feature vectors according to the number of classified types comprises:

combining the gate-level modules in a pairwise manner to obtain a static structure feature type;

4. The hardware Trojan horse detection method based on the cascade structure features as claimed in claim 2, wherein the method for obtaining the feature vector of the cascade structure by traversing through a depth-first search algorithm comprises:

5. The hardware Trojan horse detection method based on cascade structure characteristics as claimed in claim 1, wherein in the step S2, the neural network model comprises a fully connected neural network and a decision tree;

6. The hardware Trojan horse detection method based on cascade structure features of claim 1, wherein in the step S3, the Trojan horse detection using the trained neural network model comprises: