AU2021104460A4

AU2021104460A4 - an electromagnetic attack method of AES cryptographic chip based on neural network

Info

Publication number: AU2021104460A4
Application number: AU2021104460A
Authority: AU
Inventors: Tao Dong; Zhuoxian ZHANG
Original assignee: Southwest University
Current assignee: Southwest University
Priority date: 2021-07-23
Filing date: 2021-07-23
Publication date: 2021-09-16
Anticipated expiration: 2029-07-23

Abstract

The invention provides an electromagnetic attack method and system of AES cryptographic chip based on neural network. The method includes: constructing training data set; the training data set includes multiple types of data; construct the neural network model; preprocessing the training data set; the neural network model is trained through the training set after preprocessing; preprocessing the attack data; 32 groups of trained neural network models were used to classify and predict the preprocessed data to be attacked, and 32 groups of keys were obtained. Each key group includes 4 bits of key. The 32 groups of keys are combined to obtain a 128-bit key of the AES cryptographic chip. The invention can attack the key directly. There is no need to determine where any intermediate step occurs, and no need for plaintext or ciphertext. It can directly analyze the value of the key.

Description

1. Technical Field

The invention relates to the field of electromagnetic attack, in particular to an electromagnetic

attack method of AES cryptographic chip based on neural network.

2. Background

Electromagnetic analysis attack is one of the side channel attacks. It obtains the key information

of the cryptographic chip by collecting and analyzing the electromagnetic leakage. Common

electromagnetic analysis attack methods include simple electromagnetic analysis attack, correlation

electromagnetic analysis attack, differential electromagnetic analysis attack, template attack and so on.

These attack methods usually use Hamming weight or Hamming distance model to attack the

intermediate value of a certain step in the encryption process. So these have the following two

disadvantages. The first is to determine the position of one of the step encryption wheels. The second is

that the plaintext or ciphertext used in the encryption must be known in order to deduce the key.

3. The Invention Content

The purpose of the present invention is to provide an electromagnetic attack method and system of

AES cryptographic chip based on neural network. Direct attack against the key, it does not need to

determine the location of the middle step of the encryption wheel, and does not need plaintext or

ciphertext. It can directly analyze the value of the key.

To realize the above purpose, the invention provides the following scheme:

An electromagnetic attack method of AES cryptographic chip based on neural network, including:

Construct training data set; the training data set includes multiple types of data;

Construct the neural network model;

Preprocessing the training data set;

The neural network model is trained through the training set after preprocessing;

Preprocessing the attack data;

32 groups of trained neural network models are used to classify and predict the preprocessed data

to be attacked, and 32 groups of keys are obtained; each group of keys includes 4-bit keys;

The 32 groups of keys are combined to obtain a 128-bit key of the AES cryptographic chip.

Optionally, the construction of the training data set, specifically including:

The key randomly generated by non-attack field is used to encrypt the electromagnetic leakage

data;

The plaintext generated by the random algorithm is used to encrypt the electromagnetic leakage

data;

For each type corresponding to the attack field, the same amount of electromagnetic leak data is

collected.

Optionally, the said training data set is preprocessed, specifically including:

Data enhancement and normalization processing are carried out on the training data set.

Optionally, data enhancement processing is performed on the training data set, specifically

including:

The position of each data in each kind of data D in the training data set is randomly scrambled to

generate a data set.

Select the first L data from dataset D' and superposition it to generate an enhanced data B'.

Delete the selected L data and cycle until the number of remaining data in dataset D' is less than L,

then put the remaining data into D' . Once again, randomly scramble the selected stack operation, and

cycle until the generated enhanced data reaches k pieces.

Optionally, the neural network model includes an input layer, a first residual layer, a second

residual layer, a maximum pooling layer, a one-dimensional transition layer, a full connection layer and

an output layer.

Optionally, this may also include:

The trained neural network model is tested through a test data set.

The invention also provides an AES cryptographic chip electromagnetic attack system based on

neural network, including:

The training set construction module is used to construct the training data set; the training data set

includes multiple types of data;

The model building module is used to build the neural network model;

The first preprocessing module is used for preprocessing the training data set;

The training module is used to train the neural network model through the pre-processed training

set;

The second preprocessing module is used to preprocess the attack data;

In the classification and prediction module, 32 groups of trained neural network models are used to classify and predict the preprocessed data to be attacked, and 32 groups of keys are obtained. Each key group includes 4 bits of key;

A combination module is used to combine the 32 groups of keys to obtain a 128-bit key of the

AES cryptographic chip.

Optionally, the training set building module specifically includes:

The first encryption unit encrypts the electromagnetic leakage data by using the keys randomly

generated by non-attack fields;

The second encryption unit encrypts the electromagnetic leakage data through the plaintext

generated by the random algorithm;

The acquisition unit is used to collect the same amount of electromagnetic leak data for each type

of attack field.

Optionally, the first preprocessing module, specifically including:

A data enhancement unit for data enhancement processing of the training data set;

The normalization processing unit is used for the normalization processing of the enhanced

training data set.

Optionally, the neural network model comprises an input layer, a first residual layer, a second

an output layer.

According to the specific embodiment provided by the invention, the invention discloses the

following technical effects:

The invention provides an electromagnetic attack method and system of AES cryptographic chip

based on neural network, which is directly constructed against the key, rather than the intermediate

value, and the result is the key. Therefore, the invention can directly model and analyze the key without

the need for plaintext or ciphertext to derive the key.

4. Instruction With Pictures

In order to more clearly illustrate the examples of the invention or the technical solutions in the

prior art, the following will briefly introduce the drawings needed in the embodiments. It is obvious

that the drawings in the following description are only some examples of the invention. For those

skilled in the art, other drawings can be obtained from these drawings without paying creative labor.

Fig. 1 is the flowchart of the electromagnetic attack method of the AES cryptographic chip based on the neural network in an embodiment of the invention;

Fig. 2 is the schematic diagram of the model construction of the neural network of the

embodiment of the invention;

Fig. 3 is a structural schematic diagram of the neural network model of an embodiment of the

invention;

Fig. 4 is the structural schematic diagram of ClResBlock-1, an embodiment of the invention;

Fig. 5 is a structural schematic diagram of ClResBlock-2, an embodiment of the invention;

Fig. 6 is the structural schematic diagram oflResBlock, an embodiment of the invention;

Fig. 7 is a schematic diagram of the process of data preprocessing in embodiments of the

invention;

Fig. 8 is the flow diagram of the data enhancement method of the embodiment of the invention;

Fig. 9 is a schematic diagram of the data enhancement process for each type of data in

embodiment of the invention;

Fig. 10 is a schematic diagram of the neural network model training of an embodiment of the

invention;

Fig. 11 is a schematic diagram of the embodiment of the invention using the trained neural

network model to classify and predict the preprocessed data to be attacked;

Fig. 12 is a schematic diagram of the test accuracy of the neural network model of the

embodiment of the invention;

Fig. 13 is a block diagram of the electromagnetic attack system of AES cryptographic chip based

on neural network as an embodiment of the invention.

5. Specific Implementation Method

The technical scheme in the embodiment of the invention will be described clearly and completely

in combination with the drawings in the embodiment of the invention. Obviously, the described

embodiment are only part of the embodiment of the present invention, not all of them. Based on the

embodiment in the invention, all other embodiment obtained by ordinary technicians in the art without

making creative work belong to the protection scope of the invention.

AES cryptographic chip based on neural network. The direct attack against the key does not need to

determine the location of a certain step in the middle, nor does it need plaintext or ciphertext, so it can directly analyze the value of the key.

In order to make the above purposes, characteristics and advantages of the invention more obvious

and understandable, the invention is further explained in detail in combination with the attached

drawings and specific methods of implementation.

As shown in Figure 1, an electromagnetic attack method of AES cryptographic chip based on

neural network, including:

Step 101: Build the training data set; the training data set includes multiple types of data.

Prepare the appropriate training data set. In order to reduce the influence of the non-attacking

target field and plaintext in the key on the classification effect of the neural network classification

model constructed, a network model specially facing the target field is trained, and the electromagnetic

leakage data used for training the model meets the following conditions:

The non-attack field of the key used for encryption is randomly generated.

The plaintext used for encryption is generated by a random algorithm, that is, the plaintext used

for each encryption is different.

Each type corresponding to the attack field, from 0000 to 1111, collects the same amount of

electromagnetic leakage data.

Step 102: Build a neural network model.

The 128-bit keys of the Advanced Encryption Standard (AES) are divided into 32 groups of four

bits. Each group contains 16 types of data. Sixteen deep neural network classification models were

built for each group. The specific network model construction method is shown in Figure 2.

During constructing a neural network model, the invention designs a neural network model. Its

structural diagram is shown in Fig. 3. The neural network model includes the input layer, the first

residual layer (the segment containing two residual blocks), the second residual layer (the segment

containing m residual blocks), the maximum pooling layer, the one-dimensional layer, the full

connection layer and the output layer. Choosing different m and n values, we can build different

depth network structure.

The schematic diagram of CIRresBlock-1 (Convolved Direct Connection Residual Block) is

shown in Figure 4. In the figure, x input represents the network input, conv represents the convolution

operation, BN represents the batch normalization operation, ReLU represents the ReLU activation

function, Max-pool represents the maximum Pool operation, and Stack represents the data stacking operation.

The structural schematic diagram of CIRresBlcck-2 is shown in Figure 5. The meaning of each

module in the figure is the same as that in the structural schematic diagram of ClresBlock-1.

The structural schematic diagram of IResBock (Identity Res Block) is shown in Figure 6. The

meaning of each module in the figure is consistent with that in the structural schematic diagram of

ClResBlock-1.

Step 103: Preprocess the training data set.

The data preprocessing process in the invention mainly includes the data enhancement process and

the normalization process. The data enhancement scheme effectively solves the problems of limited

data set and low signal-to-noise ratio. Normalization can effectively solve the problem of small value

of original data. The schematic diagram of the preprocessing process is shown in Fig. 7. x; represents

the original data of article i , a total of n pieces of original data, ai represents the data of article i

obtained through the data enhancement scheme, and a total of m pieces of enhanced data, and pi

represents the data after normalization of each piece of enhanced data.

The flow diagram of the data enhancement method is shown in Figure 8. The original total data

set D, is divided into training data set D and test data set b without cross, and then the data is

enhanced according to the data category. The training data set is used to train the network model, and

the test data set is used to test the accuracy of the network model, which is specially used to test

whether the training model is available. The data in test and training data sets are all known to be

classified. The data to be attacked mentioned above refers to the data of unknown classification. The

invention uses the network model trained by the training data set to predict its classification and obtain

its key.

The following example D is used to illustrate the data enhancement operation (the processing of

the test data set and the attack data is the same as the training data set).

2 The training data set is D= IDI,D , D 3,..., D"], where D' represents the data of class i , contains

n types of data. Each type of data D' contains m pieces of data, expressed as

D=[D', , ,..., D . L represents the number of data strips needed for each stacking. The stacking

data set B = [BB 2, B 3,..., B" is generated through the flow chart shown in Figure 8. B' represents the superposition data of class i ,which contains a total of n types of data. Each type of data B' contains k pieces of data (k satisfies the restriction condition k < C, ), expressed as

B =( B', B ,,B',..., B,'

. As shown in Figure 9, the data enhancement process for each type of data is as follows. Firstly,

the position of each data in D' is randomly scrambled to generate data set D' . Then select the first L

data from D' and stack it to generate an enhanced data B . Delete the selected L pieces of data,

cycle until the number of remaining data in D' is less than L, and put the remaining data in D' . Then

randomly scramble them again to perform the selective overlay operation, so as to loop until the

number of generated enhanced data reaches k.

The normalization process adopts the method of mean standard deviation normalization, and the

calculation formula is as follows:

a-" P= a-p

where a is the data to be normalized, p is the normalized data, p is the mean value of x

and a is the standard deviation of x. In the present invention, each piece of data in the enhanced

data set needs to be normalized separately. a={a,a 2 ,...,a,} corresponds to a specific

one-dimensional data vector. First, the mean value y and standard deviation a are calculated. Then subtract the mean value y from each value a, in a, and divide it by the standard deviation

a to get the corresponding data p after a normalization.

Step 104: The neural network model is trained through the preprocessed training set. As shown in

Figure 10, the trained classification model ModelX can be obtained finally, where X represents the

position of the corresponding 4-bit key of the model in the 128-bit key of AES. For example, Modell

means that the classification result of the model is the 4-bit key of group 1, and Mode32 means that

the classification result of the model is the 4-bit key of group 32.

Step 105: Preprocess the attack data. The preprocessing process of the attack data is the same as

that of the training data set.

Step 106: 32 groups of trained neural network models were used to classify and predict the

preprocessed data to be attacked, and 32 groups of keys were obtained. Each key group consists of four

bits.

Step 107: Combine the 32 key groups to obtain the 128-bit key of the AES cipher chip. As shown

in Figure 11.

The test accuracy of 32 groups of 16 types of neural network models constructed using this

method is shown in Fig. 12. It can be seen from the figure that the accuracy of each model is greater

than 80%, indicating that the neural network model constructed by the invention is very effective for

the classification of electromagnetic leakage signals. Therefore, the 128-bit key of AES can be

effectively analyzed when the 32 network models trained by the invention are used to carry out actual

attacks.

following technical effects:

1)The object of attack is the whole encryption process, and there is no difficulty in positioning. In

traditional electromagnetic analysis attacks, the target is usually a certain step (or a certain round) in

the encryption process. In practice, it is difficult to accurately locate a step (or a round) in the

encryption process. The electromagnetic attack method used in this method directly analyzes the whole

encryption process and does not have the problem of difficult localization.

2)Modeling and analyzing the key directly, without knowing the plaintext or ciphertext in advance

for key derivation. This is because the electromagnetic attack model of this method is built directly

against the entire execution of the key, rather than the median value of a particular round. Therefore,

when the model is matched, the result is the key.

3)The amount of raw data required to build the model is small. Because of the data enhancement

scheme designed in this method, a large amount of enhanced data with high SNR can be generated by

using less original data with low SNR.

4)The accuracy of model construction is high. Based on residual neural network, CIResBlock-1

and CIResBlock-2 convolutional direct connected residual blocks are designed. In the residual blocks

with unequal input and output sizes, there is an approximate direct connection between the input and

output, so that the whole network error can be more effectively propagated in the whole network. Thus,

the accuracy of neural network classification model is effectively improved. In addition, the high

accuracy is also related to the data enhancement operation improving the signal-to-noise ratio and

increasing the number of training sets.

As shown in Fig. 13, the present invention also provides an electromagnetic attack system of AES cryptographic chip based on neural network, including:

The training set construction module 1301 is used to construct the training data set; the training

data set includes multiple types of data.

The training set construction module 1301 specifically includes:

generated by non-attack fields.

generated by the random algorithm.

The acquisition unit is used to collect the same amount of electromagnetic X data for each type

corresponding to the attack field.

The model building module 1302 is used to build the neural network model. The neural network

model includes an input layer, a first residual layer, a second residual layer, a maximum pooling layer,

a one-dimensional layer, a full connection layer and an output layer.

The first preprocessing module 1303 for preprocessing the training data set.

The first preprocessing module 1303 specifically includes:

A data enhancement unit for data enhancement processing of the training data set:

The normalized processing unit is used to normalize the enhanced training data set.

The training module 1304 is used for training the neural network model through the preprocessed

training set.

The second preprocessing module 1305 is used for preprocessing the attack data.

The classification and prediction module 1306 uses 32 groups of trained neural network models to

classify and predict the preprocessed data to be attacked, and obtains 32 groups of keys; Each set of

keys includes a 4-bit key.

The combination module 1307 is used to combine the 32 groups of keys to obtain a 128-bit key of

the AES cryptographic chip.

[0106]The embodiment in this specification are described in a progressive manner. Each

embodiment focuses on the differences from other embodiment. The same or similar parts of the

various embodiment can be referred to each other. For the systems disclosed by the embodiment, since

they correspond to the methods disclosed by the embodiment. Therefore, the description is relatively

simple. Please refer to the description of the method section for relevant points.

In this paper, specific examples are applied to illustrate the principle and implementation mode of

the invention. The description of the above embodiment is only for helping to understand the method

and the core idea of the present invention. At the same time, for the general technical personnel in the

field, according to the idea of the invention, there will be changes in the specific implementation mode

and application scope. In summary, the content of this specification should not be construed as a

limitation of the invention.

1. An electromagnetic attack method of AES cryptographic chip based on neural network is

characterized by:

Construct training data set; the training data set includes a plurality of types of data;

The neural network model is constructed;

Preprocessing the training data set;

The neural network model is trained through the preprocessed training set;

The attack data is preprocessed;

Combine the 32 sets of keys to obtain a 128-bit key of the AES cryptographic chip.

2. According to the electromagnetic attack method of AES cryptographic chip based on neural

network mentioned in Claim 1, its characteristics lie in that the training data set is constructed,

specifically including:

Use a key randomly generated from the non-attack field to encrypt the electromagnetic leakage

data;

The plaintext generated by a random algorithm is used to encrypt the electromagnetic leakage

data;

For each type corresponding to the attack field, the same amount of electromagnetic leakage data

is collected.

3. According to the electromagnetic attack method of AES cryptographic chip based on neural

network mentioned in Claim 1, its characteristic lies in the preprocessing of the training data set, which

specifically includes:

Perform data enhancement and normalization processing on the training data set.

4. According to the electromagnetic attack method of AES cryptographic chip based on neural

network mentioned in Claim 3, its characteristic lies in the data enhancement processing of the training

data set, which specifically includes:

Randomly scramble the position of each piece of data in each type of data D in the training data

set to generate a data set D'

Select the first L data from the data set D' , superimpose to generate an enhanced data B , and delete the selected L data, and loop until the data set D' ; when the number of remaining data in the data is less than L , put the remaining data into D' ; randomly scramble again to carry out selective overlay operation, and cycle until the generated enhanced data reaches k pieces.

5. According to the electromagnetic attack method of AES cryptographic chip based on the neural

network mentioned in Claim 1, its characteristics lie in that the neural network model includes the input

layer, the first residual layer, the second residual layer, the maximum pool layer, the one-dimensional

layer, the full connection layer and the output layer.

6. According to the electromagnetic attack method of AES cryptographic chip based on neural

network mentioned in Claim 1, its characteristics are as follows:

Test the trained neural network model through the test data set.

7. A neural network based AES cryptographic chip electromagnetic attack system, which is

characterized by:

The training set building module is used to build a training data set; the training data set includes

multiple types of data;

Model building module for building neural network models:

The first preprocessing module is used to preprocess the training data set;

The training module is used to train the neural network model through the preprocessed training

set;

The second preprocessing module is used to preprocess the data to be attacked;

In the classification and prediction module, 32 groups of trained neural network models are used

to classify and predict the preprocessed data to be attacked, and 32 groups of keys are obtained; each

key group includes 4 bits of key;

The combination module is used to combine the 32 sets of keys to obtain the 128-bit key of the

AES cryptographic chip.

8. According to the electromagnetic attack system of AES cryptographic chip based on neural

network mentioned in Claim 7, its characteristics lie in that the training set building module

specifically includes:

generated by non-attack fields.

The second encryption unit encrypts the electromagnetic leakage data through the plaintext generated by the random algorithm.

The collection unit is used to collect the same amount of electromagnetic leakage data for each

type corresponding to the attack field.

9. According to the electromagnetic attack system of AES cryptographic chip based on neural

network mentioned in Claim 7, its features lie in the first preprocessing module mentioned above,

which specifically includes:

The normalization processing unit is used to normalize the enhanced training data set.

10. According to the electromagnetic attack system of AES cryptographic chip based on the neural

network mentioned in Claim 7, its characteristics lie in that the neural network model includes the input

layer, the first residual layer, the second residual layer, the maximum pooling layer, the

one-dimensional layer, the full connection layer and the output layer.