Deep learning side channel analysis method for efficiently recovering secret key

Abstract

The invention relates to the technical field of information security, in particular to a deep learning side channel analysis method for efficiently recovering a key, which comprises the following steps: S1, acquiring plaintext, key and encrypted energy leakage information; s2, byte labels set by bytes corresponding to the secret key and energy traces in the interest interval correspondingly form a data set, and the data set is divided into a training set, a verification set and a test set; s3, constructing a single convolutional neural network model according to the relationship between the energy leakage generated in the interest interval and the key, and establishing a side channel analysis deep learning model; s4, inputting the training set into a side channel analysis deep learning model for training to obtain a final model, and collecting training indexes; and S5, evaluating the final model, recovering the byte key by using the final model, and evaluating the attack effect. According to the method, the key of n bytes can be recovered only by training one model, so that the number of models is reduced, the training time is saved, and the efficiency of deep learning side channel analysis is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of information security technologies, and in particular to a deep learning side channel analysis method for efficiently recovering keys. Background Art

[0002] The explosive growth of the Internet of Things (IoT) has not only brought progress to information and communication technology, but also brought security challenges. Since IoT devices tend to process personal data, potential attacks must be considered when protecting these devices. Side channel analysis (SCA) is the most representative of potential attacks. SCA is a technology based on analyzing data obtained from the execution of an algorithm, rather than the algorithm itself. We will call this data leakage, and the types of leakage include execution time, power consumption, and electromagnetic radiation, etc., using the above information to attack and spy on the cryptographic system. Side channel analysis has become a huge threat to information security cryptographic products, and its harm is far greater than traditional mathematical analysis methods. The ultimate goal of SCA is to reveal the secret key of a device. This analysis is performed by exploiting the relationship between the leakage obtained during the execution of the algorithm and the key. The attacker collects the leakage and analyzes them using specific models and appropriate indicators to determine the key. SCA can be transformed into a classification problem. Since classifying given data is an important application of deep learning, deep learning models can automatically understand the patterns, features, and logical relationships that may exist in the extracted data. So deep learning technology can be applied to SCA to improve the efficiency and accuracy of analysis. Power consumption attack is a type of side channel attack. The general core idea is that for a cryptographic encryption device, the amount of energy consumed during its operation is directly related to the intermediate value calculated during the encryption process, and the intermediate value information is directly related to the key and plaintext information input and output by the encryption device during operation. There are many works that combine deep learning methods with power consumption side channel analysis to successfully recover the key.

[0003] In the current side channel attacks based on deep learning, it is believed that since the key of AES-128 encryption method has 16 bytes, the attacker needs to train 16 neural network models to recover the entire key. To perform deep learning side channel analysis on each byte, data collection and data preprocessing are required, and the data position corresponding to each sub-byte is identified for slicing, and each model is established and trained separately, etc., in order to finally recover the key of all bytes, which is time-consuming and requires a lot of operations. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and provide a deep learning side channel analysis method for efficiently recovering keys, comprising the following steps: S1: Run an encryption device to obtain a known plaintext and a known key, and collect energy leakage information obtained by performing a first round of encryption on the known plaintext and the known key, wherein the energy leakage information is raw data formed by power consumption or energy traces, wherein the energy traces encrypted using random plaintext and random keys are data used for modeling, and the energy traces encrypted using random plaintext and fixed keys are data used for testing; S2: The byte labels of the bytes corresponding to the known key correspond to the energy traces in the interest interval one by one to form a data set, and the data set is divided into a training set, a verification set and a test set. The interest interval is to use the correlation power consumption analysis CPA to analyze the energy leakage information generated during AES encryption, and find the position on the energy trace corresponding to each byte of the known key; S3: constructing a single convolutional neural network model according to the relationship between the energy leakage generated in the interest interval and the known key, and establishing a side channel analysis deep learning model according to the single convolutional neural network model; S4: inputting the training set into the side channel analysis deep learning model for training to obtain a final model, and collecting training indicators during the training process; S5: Evaluate the final model, use the final model to recover the key of the byte, and evaluate the attack effect.

[0005] Preferably, the step S1 further comprises: Randomly generate the known plaintext , running the encryption device, the known key Through the first round of encryption operation, the known plaintext Encryption, the intermediate value corresponding to the energy leakage generated during the encryption process is the output of the first round of Sbox, the first round of encryption operation is the byte replacement operation of the AES algorithm, the energy leakage is the energy trace collected during the encryption process, and the calculation formula of the first round of encryption operation is: in, For XOR operation, Corresponding S-box lookup table operation.

[0006] Preferably, in step S2, the use of correlation power consumption analysis CPA to analyze the energy leakage information generated during AES encryption to find the position on the energy trace corresponding to each byte of the known key, that is, the interval of interest, further includes: S211: Aligning the energy leakage information; S212: Calculate the intermediate value according to the byte replacement operation matrix; S213: Calculate the intermediate value The assumed power consumption data corresponding to the matrix is ​​used to calculate the Pearson correlation coefficient between the assumed power consumption data and the original data, and the interval with the maximum correlation coefficient, namely, the interval of interest, is selected, and the original data is subjected to dimensionality reduction processing through CPA to retain the power consumption data information within the interval of interest, and the interval of interest on the original data corresponding to each byte of the known key is selected.

[0007] Preferably, in step S2, forming a data set further comprises: S221: The byte label is outputted by setting the byte of the known key as the attack target. The calculation formula of the byte label is: in, The first byte, The first byte; S222, selecting energy traces in the interval of interest from the original data as the data set through CPA; S223, selecting 70% of the data set as the training set, 10% as the validation set, and 20% as the test set.

[0008] Preferably, in step S3, constructing a single convolutional neural network model, and establishing a side channel analysis deep learning model according to the single convolutional neural network model further includes: S31: Establishing an input layer InputLayer for receiving the power consumption data information; S32: Establishing a network layer for processing the power consumption data information; S33: establishing an output layer Dense with 256 units, wherein the output layer Dense uses a softmax activation function network to output probability distributions of 256 categories, corresponding to 256 sub-keys; S34: Configure the single convolutional neural network model to obtain the side channel analysis deep learning model.

[0009] Preferably, in step S32, the network layer further comprises: The network layer includes a first one-dimensional convolution layer Conv1D, a first one-dimensional average pooling layer AveragePooling1D, a second one-dimensional convolution layer Conv1D, a second one-dimensional average pooling layer AveragePooling1D, data Flatten for converting multi-dimensional feature data into one-dimensional data, and a fully connected layer Dense with 256 units, and the activation function of the fully connected layer Dense is softmax.

[0010] Preferably, the step S4 further comprises: S41: Using the model accuracy acc as the training indicator; S42: Input the training set into the side channel analysis deep learning model for training. After each training, use the current model to test the verification set, and record the model accuracy acc to view the quality of the current model. After reaching the preset number of steps, save the final model.

[0011] Preferably, in step S42, the calculation formula of the model accuracy acc is as follows: , in, represents the validation set, Indicates the first Power consumption curve, Indicates the correct key, represents the set of the power consumption curves when the guessed key is equal to the correct key, and the model accuracy acc is the ratio of the number of power consumption curves when the guessed key is equal to the correct key to the number of power consumption curves in the verification set.

[0012] Preferably, the step S5 further comprises: S51: inputting the test set into the final model to attack and output the model; S52, saving the model output , performing a reverse operation on the model output to obtain the guessed key; S53, use The index evaluates the attack effect of the final model. The fewer attack curves used, the better the model.

[0013] Preferably, the Indicators further include: Said The indicator is the ranking of the correct key among the guessed keys. The calculation formula of the ranking is as follows: , in, is the index value of the guessed key, The output of the model is 16 tensors with 256 elements, corresponding to the probabilities of 256 results of guessing the key of 16 bytes. A tensor sorted in descending order.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention trains a model to recover the entire key without selecting an interest interval for each byte, and at the same time reduces the attack time by 15 / 16 for deep learning side channel analysis, reduces the number of model training, increases the efficiency of side channel attacks, and reduces the cost of side channel attacks; From the data level, the present invention enhances the data. The model can obtain better training results by training all bytes than single byte. Due to the AES encryption algorithm running on the microcontroller, in the same round of training, there is a strong correlation between each byte. This method of processing data also performs better for data sets with strong noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.

[0016] Figure 1 A flowchart of a deep learning side channel analysis method for efficiently recovering keys according to the present invention; Figure 2 A schematic diagram of the interest interval of 16 bytes of AES-128 on the power consumption trace of a deep learning side channel analysis method for efficiently recovering keys according to the present invention; Figure 3 This is a schematic diagram of the rank value results of a CNN model of a deep learning side channel analysis method for efficiently recovering keys in the present invention for a 16-byte attack. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0018] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a", "an", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0019] See also Figure 1 As shown, this embodiment provides a deep learning side channel analysis method for efficiently recovering keys, comprising the following steps: S1: Run an encryption device to obtain a known plaintext and a known key, and collect energy leakage information obtained by performing a first round of encryption on the known plaintext and the known key, wherein the energy leakage information is raw data formed by power consumption or energy traces, wherein the energy traces encrypted using random plaintext and random keys are data used for modeling, and the energy traces encrypted using random plaintext and fixed keys are data used for testing; S2: The byte labels of the bytes corresponding to the known key correspond to the energy traces in the interest interval one by one to form a data set, and the data set is divided into a training set, a verification set and a test set. The interest interval is to use the correlation power consumption analysis CPA to analyze the energy leakage information generated during AES encryption, and find the position on the energy trace corresponding to each byte of the known key; S3: constructing a single convolutional neural network model according to the relationship between the energy leakage generated in the interest interval and the known key, and establishing a side channel analysis deep learning model according to the single convolutional neural network model. Specifically, in this embodiment, different layers of the neural network, loss functions, training steps and other parameters are set; S4: inputting the training set into the side channel analysis deep learning model for training to obtain a final model, and collecting training indicators during the training process; S5: Evaluate the final model, use the final model to recover the key of the byte, and evaluate the attack effect.

[0020] Preferably, the step S1 further comprises: Randomly generate the known plaintext , running the encryption device, the known key Through the first round of encryption operation, the known plaintext Encryption, the intermediate value corresponding to the energy leakage generated during the encryption process is the output of the first round of Sbox, the first round of encryption operation is the byte replacement operation of the AES algorithm, the energy leakage is the energy trace collected during the encryption process, and the calculation formula of the first round of encryption operation is: in, For XOR operation, Corresponding S-box lookup table operation.

[0021] Preferably, this embodiment takes the output of the Sbox of the first round of AES-128 encryption as the attack target, including: randomly generating 80k of 16-byte plaintext, running the encryption device, encrypting with a random key, collecting power consumption data as modeling data, randomly generating 20k of 16-byte plaintext, running the encryption device, encrypting with a fixed key, collecting power consumption data as test data, the intermediate value corresponding to the energy leakage generated during the first round of encryption is the output of the Sbox of the first round, the encryption operation performed is the byte replacement operation of the AES algorithm, and the power consumption trajectory, plaintext, ciphertext and key are added to the data set.

[0022] Preferably, in step S2, the use of correlation power consumption analysis CPA to analyze the energy leakage information generated during AES encryption to find the position on the energy trace corresponding to each byte of the known key, that is, the interval of interest, further includes: S211: Aligning the energy leakage information; S212: Calculate the intermediate value according to the byte replacement operation matrix; S213: Calculate the intermediate value The assumed power consumption data corresponding to the matrix is ​​used to calculate the Pearson correlation coefficient between the assumed power consumption data and the original data, and the interval with the maximum correlation coefficient, namely, the interval of interest, is selected, and the original data is subjected to dimensionality reduction processing through CPA to retain the power consumption data information within the interval of interest, and the interval of interest on the original data corresponding to each byte of the known key is selected.

[0023] Preferably, step S211 further includes: After aligning all data, we get the energy consumption matrix ,The encryption device generates n energy curves during the encryption process, each curve has p energy consumption points, and the original data is written in the form of a matrix. Each row is a power consumption curve, and each column is the value of p power consumption sampling points corresponding to time: .

[0024] Preferably, step S212 further includes: Possible keys , the plain text input is , calculate the pseudo intermediate value matrix , size is ,Calculation formula: Preferably, step S213 further includes: Using the Hamming weight model as the energy mapping model, the assumed intermediate value matrix is ​​mapped to the assumed power consumption data matrix, and the assumed intermediate value is calculated, where the assumed power consumption matrix is ​​calculated. The calculation formula is: ; Calculate the assumed power consumption moment After that, calculate the assumed power consumption matrix The actual power consumption matrix The correlation coefficient of and They are column vectors and The arithmetic mean of is the number of possible keys, is the number of sampling points for each power, and the Pearson correlation coefficient matrix is ​​obtained , the size is , the calculation formula is as follows: ; Please refer to Figure 2 As shown, the point with the largest correlation coefficient is selected as the interest point. The positions of the 16-byte interest points are [96, 192, 288, 384, 480, 576, 672, 768, 864, 960, 1056, 1152, 1248, 1344, 1440,1536], an interval consisting of 95 sampling points including the interest point is selected to form a 16-byte interest interval.

[0025] Preferably, in step S2, forming a data set further comprises: S221: The byte label is outputted by setting the byte of the known key as the attack target. The calculation formula of the byte label is: in, The first byte, The first Bytes. Specifically, in this embodiment, if the first byte of the key is taken as the attack target, the label is the intermediate value of the first byte after the byte replacement calculation. The calculation formula of the label is: , where Indicates the first byte of the key, It represents the first byte of the plaintext. When the key of other bytes is used as the attack target, it is only necessary to replace the key and plaintext with the corresponding bytes, so the label is set to the middle value corresponding to 16 bytes; S222, select the energy trace in the interest interval from the original data through CPA as the data set. Specifically, in this embodiment, select the 16-byte power consumption data interest interval from the original data through CPA as the data set, and divide an original trace into 16 parts using CPA, and each energy trace corresponds to the leakage of the byte key of the cryptographic device during the encryption process. 16 energy traces correspond to 16 different labels, which must be one-to-one corresponding to ensure that the label and the energy trace have sufficient correlation; S223, select 70% of the data set as the training set, 10% as the validation set, and 20% as the test set. Specifically, in this embodiment, there are 100k energy trajectories collected, and we select 70k as the training set, 10k as the validation set, and 20k as the test set.

[0026] Preferably, in step S3, constructing a single convolutional neural network model, and establishing a side channel analysis deep learning model according to the single convolutional neural network model further includes: S31: establishing an input layer InputLayer for receiving the power consumption data information. Specifically, in this embodiment, the input is an energy trajectory in an interest interval corresponding to 16 bytes, each of which has 95 sampling points; S32: Establishing a network layer for processing the power consumption data information; S33: Establish an output layer Dense with 256 units. The output layer Dense uses a softmax activation function network to output probability distributions of 256 categories, corresponding to 256 sub-keys. Specifically, in this embodiment, the output layer Dense has a total of 256 units and outputs a vector with 256 elements. The last layer uses a softmax activation function; S34: Configure the single convolutional neural network model to obtain the side channel analysis deep learning model. Specifically, in this embodiment, RMSporp is set as the optimizer, the learning rate is 0.0005, the batch_size is 200, the number of epochs is 200, the loss function is categorical_crossentropy, and the input of the model is the power consumption data within the 16-byte interest interval of the key. The output of the model is a tensor with 256 elements, which correspond to the probabilities of guessing the key {0, 1, ···, 255}.

[0027] Preferably, in step S32, the network layer further comprises: The network layer includes a first one-dimensional convolution layer Conv1D, a first one-dimensional average pooling layer AveragePooling1D, a second one-dimensional convolution layer Conv1D, a second one-dimensional average pooling layer AveragePooling1D, a data Flatten for converting multi-dimensional feature data into one-dimensional data, and a fully connected layer Dense with 256 units. The activation function of the fully connected layer Dense is softmax. Specifically, in this embodiment, the first one-dimensional convolution layer Conv1D includes 32 filters, the convolution kernel size is 3, the activation function is ReLU, the first one-dimensional average pooling layer AveragePooling1D is a one-dimensional average pooling layer with a pooling window size of 2 and a step size of 1, the second one-dimensional convolution layer Conv1D includes 64 filters, the convolution kernel size is 5, the activation function is ReLU, the second one-dimensional average pooling layer AveragePooling1D is a one-dimensional average pooling layer with a pooling window size of 2 and a step size of 1, and the fully connected layer Dense includes a first fully connected layer Dense and a second fully connected layer Dense.

[0028] Preferably, the step S4 further comprises: S41: Using the model accuracy acc as the training indicator; S42: Input the training set into the side channel analysis deep learning model for training. After each training, use the current model to test the verification set, and record the model accuracy acc to view the quality of the current model. After reaching the preset number of steps, save the final model.

[0029] Preferably, in step S42, the calculation formula of the model accuracy acc is as follows: , in, represents the validation set, Indicates the first Power consumption curve, Indicates the correct key, represents the set of the power consumption curves when the guessed key is equal to the correct key, and the model accuracy acc is the ratio of the number of power consumption curves when the guessed key is equal to the correct key to the number of power consumption curves in the verification set.

[0030] Preferably, the step S5 further comprises: S51: inputting the test set into the final model to attack and output the model. Specifically, in this embodiment, the power consumption data of the test set is input into the final model; S52, saving the model output , perform a reverse operation on the model output to obtain the guessed key. Specifically, in this embodiment, the model output is a one-dimensional floating point array of 256 elements, whose index represents the candidate subkey, and the value at the index represents the probability that the candidate subkey is the correct subkey. The index value is reversed Sboxed and XORed with the plaintext to convert the model output into the probability of guessing the key. Use the argmax function of numpy to sort the probability array of guessed keys in descending order, and the guessed key with the largest probability is the attack result; S53, use The index evaluates the attack effect of the final model. The fewer attack curves used, the better the model.

[0031] Preferably, the Indicators further include: Said The indicator is the ranking of the correct key among the guessed keys. The calculation formula of the ranking is as follows: , in, is the index value of the guessed key, The output of the model is 16 tensors with 256 elements, corresponding to the probabilities of 256 results of guessing the key of 16 bytes. A tensor sorted in descending order.

[0032] Please refer to Figure 3 As shown in the figure, the power consumption data of different bytes of the test set are input into the model to attack different bytes respectively. The horizontal axis is the number of power consumption data used for attack, and the vertical axis is the rank value. The probability of the attack results of several curves is accumulated. For 16 bytes, the rank can finally stabilize to 0 within ten curves, indicating that the model of the present invention can recover the key of all bytes. The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A deep learning side channel analysis method for efficiently recovering keys, characterized in that: The following steps are involved: S1: Acquire a known plaintext and a known key, and collect energy leakage information of a first round of encryption of the known plaintext and the known key, wherein the energy leakage information is raw data formed by power consumption or energy trace; S2: The byte labels of the bytes corresponding to the known key correspond to the energy traces in the interest interval one by one to form a data set, and the data set is divided into a training set, a verification set and a test set. The interest interval is to use the correlation power consumption analysis CPA to analyze the energy leakage information generated during AES encryption, and find the position on the energy trace corresponding to each byte of the known key; S3: constructing a single convolutional neural network model according to the relationship between the energy leakage generated in the interest interval and the known key, and establishing a side channel analysis deep learning model according to the single convolutional neural network model; S4: inputting the training set into the side channel analysis deep learning model for training to obtain a final model, and collecting training indicators during the training process; S5: Evaluate the final model, use the final model to recover the key of the byte, and evaluate the attack effect.

2. A deep learning side channel analysis method for efficiently recovering keys according to claim 1, characterized in that: The step S1 further comprises: Randomly generate the known plaintext , running the encryption device, the known key Through the first round of encryption operation, the known plaintext Encryption, the intermediate value corresponding to the energy leakage generated during the encryption process is the output of the first round of Sbox, the first round of encryption operation is the byte replacement operation of the AES algorithm, the energy leakage is the energy trace collected during the encryption process, and the calculation formula of the first round of encryption operation is: in, For XOR operation, Corresponding S-box lookup table operation.

3. A deep learning side channel analysis method for efficiently recovering keys according to claim 2, characterized in that: In step S2, the energy leakage information generated during AES encryption is analyzed using correlation power consumption analysis CPA to find the position on the energy trace corresponding to each byte of the known key, that is, the interval of interest, further comprising: S211: Aligning the energy leakage information; S212: Calculate the intermediate value according to the byte replacement operation matrix; S213: Calculate the intermediate value The assumed power consumption data corresponding to the matrix is ​​used to calculate the Pearson correlation coefficient between the assumed power consumption data and the original data, and the interval with the maximum correlation coefficient, i.e., the interest interval, is selected. The original data is subjected to dimensionality reduction processing through CPA, the power consumption data information within the interest interval is retained, and the interest interval on the original data corresponding to each byte of the known key is selected.

4. A deep learning side channel analysis method for efficiently recovering keys according to claim 3, characterized in that: In step S2, forming a data set further comprises: S221: The byte label is outputted by setting the byte of the known key as the attack target. The calculation formula of the byte label is: in, The first byte, The first byte; S222, selecting energy traces in the interval of interest from the original data as the data set through CPA; S223, selecting 70% of the data set as the training set, 10% as the validation set, and 20% as the test set.

5. A deep learning side channel analysis method for efficiently recovering keys according to claim 4, characterized in that: In step S3, constructing a single convolutional neural network model, and establishing a side channel analysis deep learning model according to the single convolutional neural network model further includes: S31: Establishing an input layer InputLayer for receiving the power consumption data information; S32: Establishing a network layer for processing the power consumption data information; S33: Establish an output layer Dense with 256 units, and the output layer Dense uses a softmax activation function network to output probability distributions of 256 categories, corresponding to 256 seed keys; S34: Configure the single convolutional neural network model to obtain the side channel analysis deep learning model.

6. A deep learning side channel analysis method for efficiently recovering keys according to claim 5, characterized in that: In step S32, the network layer further includes: The network layer includes a first one-dimensional convolution layer Conv1D, a first one-dimensional average pooling layer AveragePooling1D, a second one-dimensional convolution layer Conv1D, a second one-dimensional average pooling layer AveragePooling1D, data Flatten for converting multi-dimensional feature data into one-dimensional data, and a fully connected layer Dense with 256 units, and the activation function of the fully connected layer Dense is softmax.

7. A deep learning side channel analysis method for efficiently recovering keys according to claim 6, characterized in that: The step S4 further comprises: S41: Using the model accuracy acc as the training indicator; S42: Input the training set into the side channel analysis deep learning model for training. After each training, use the current model to test the verification set, and record the model accuracy acc to view the quality of the current model. After reaching the preset number of steps, save the final model.

8. A deep learning side channel analysis method for efficiently recovering keys according to claim 7, characterized in that: In step S42, the calculation formula of the model accuracy acc is as follows: ， in, represents the validation set, Indicates the first Power consumption curve, Indicates the correct key, represents the set of the power consumption curves when the guessed key is equal to the correct key, and the model accuracy acc is the ratio of the number of power consumption curves when the guessed key is equal to the correct key to the number of power consumption curves in the verification set.

9. A deep learning side channel analysis method for efficiently recovering keys according to claim 8, characterized in that: The step S5 further comprises: S51: inputting the test set into the final model to attack and output the model; S52, saving the model output , performing a reverse operation on the model output to obtain the guessed key; S53, use The index evaluates the attack effect of the final model. The fewer attack curves used, the better the model.

10. A deep learning side channel analysis method for efficiently recovering keys according to claim 9, wherein Indicators further include: Said The indicator is the ranking of the correct key among the guessed keys. The calculation formula of the ranking is as follows: ， in, is the index value of the guessed key, The output of the model is 16 tensors with 256 elements, corresponding to the probabilities of 256 results of guessing the key of 16 bytes. A tensor sorted in descending order.

Images