Deep learning side channel attack method, system and device based on Boosting integration method and medium

Abstract

The invention discloses a deep learning side channel attack method, system and device based on a Boosting integration method and a medium, and belongs to the technical field of cryptography security assessment, and the method comprises the steps: obtaining a side channel signal sequence and a real key tag; performing wavelet denoising on the side channel signal sequence to obtain a denoised signal sequence; extracting feature vectors according to the de-noised signal sequence to construct a training sample set; constructing a plurality of weak learners adopting a multi-layer sensor structure; giving initial sample weights to samples in the training sample set, and training a plurality of weak learners in sequence in multiple rounds of iteration; carrying out weighted fusion on the output of the plurality of trained weak learners to obtain integrated prediction probability distribution; and calculating accumulated scores of the candidate keys according to the integrated prediction probability distribution, sorting the accumulated scores, and obtaining a position index of a correct key as a guess entropy. According to the method, key leakage features are reserved through wavelet denoising, difficult case learning is reinforced through dynamic sample weighting, and robustness and generalization ability are improved through multi-model fusion.

Description

Technical Field

[0001] This invention relates to the field of cryptographic security assessment technology, specifically to a deep learning side-channel attack method, system, device, and medium based on the Boosting ensemble method. Background Technology

[0002] Side-channel attacks are security threats that target the physical implementation of cryptographic devices. They infer key information by analyzing physical information (such as power consumption, electromagnetic radiation, and time) leaked during the operation of the encryption device. With the widespread use of the Internet of Things and embedded devices, side-channel attacks have become an important tool for security assessment of cryptographic chips.

[0003] Traditional side-channel attack methods mainly include differential power analysis and correlation power analysis. These methods rely on prior knowledge of the physical implementation of cryptographic algorithms and manually designed attack models, and their effectiveness decreases significantly when facing protected devices or high-noise environments. In recent years, machine learning methods have been introduced into the field of side-channel attacks, improving attack efficiency by automatically learning the mapping relationship between side-channel signals and keys.

[0004] Machine learning-based side-channel attack methods have evolved from shallow models to deep learning. Early methods used shallow models such as support vector machines and random forests, requiring manual feature engineering and exhibiting limited generalization ability. Deep learning methods achieve end-to-end learning through convolutional neural networks or multilayer perceptrons, avoiding complex feature design. However, single deep learning models suffer from the following problems: First, there is insufficient learning of difficult-to-classify samples. In side-channel attacks, the key leakage information in some trajectories is weak, and a single model cannot fully learn the information contained in these difficult samples, resulting in reduced attack efficiency.

[0005] Second, the model lacks robustness. A single model is prone to overfitting specific noise patterns in the training data. When faced with different devices, different measurement environments, or adversarial interference, the model's performance drops significantly, and its generalization ability is limited.

[0006] Third, noise interference in the side-channel signal reduces the feature extraction quality of the model. During the acquisition process, power consumption signals are mixed with measurement noise, electromagnetic interference, and other irrelevant information. Traditional methods directly input the original signal or the signal after simple filtering into the model, which cannot effectively separate key-related leakage from background noise, resulting in low quality of feature representations learned by the model.

[0007] Ensemble learning methods improve overall performance by combining the predictions of multiple weak learners, achieving significant results in traditional machine learning. However, existing ensemble learning side-channel attack methods primarily apply ensemble strategies to shallow models, failing to fully leverage the powerful feature learning capabilities of deep learning. Combining Boosting strategies with deep learning models, dynamically adjusting sample weights to focus on difficult-to-classify samples, and simultaneously utilizing the weighted fusion of multiple deep learning weak learners to enhance model robustness, is an effective way to improve the efficiency of side-channel attacks. Summary of the Invention

[0008] In view of the above-mentioned problems, the present invention provides a deep learning side-channel attack method, system, device and medium based on the Boosting ensemble method.

[0009] Therefore, the technical problems solved by this invention are: how to dynamically adjust sample weights through a Boosting ensemble strategy so that the deep learning model focuses on difficult-to-classify samples with weak key leakage, thereby improving the learning effect on difficult examples; how to retain key-related leakage features in the side-channel signal while suppressing noise through wavelet denoising preprocessing, thereby improving the quality of feature extraction; and how to improve the robustness and generalization ability of the model under different devices and noise environments through weighted fusion of multiple deep learning weak learners.

[0010] To address the aforementioned technical problems, this invention provides the following technical solution: a deep learning side-channel attack method based on the Boosting ensemble method, comprising, The side channel signal sequence and the corresponding real key tag are obtained. The side channel signal sequence is obtained by collecting the power consumption signal during the operation of the encryption device. Wavelet denoising is performed on the side channel signal sequence to obtain a denoised signal sequence. Feature vectors are extracted from the denoised signal sequence, and a training sample set is constructed based on the feature vectors and the real key labels. Construct multiple weak learners, which adopt a multilayer perceptron structure; The samples in the training sample set are assigned initial sample weights. Multiple weak learners are trained sequentially in multiple iterations. In each iteration, a weak learner is trained using the current sample weights. The sample weights are then updated based on the classification error of the trained weak learners, so that misclassified samples will receive higher weights in subsequent iterations. The outputs of multiple trained weak learners are weighted and fused to obtain the ensemble prediction probability distribution. The cumulative score of the candidate keys is calculated based on the integrated prediction probability distribution. The candidate keys are then sorted according to the cumulative score to obtain a key sorting list. The position index of the correct key in the key sorting list is obtained as the guessing entropy.

[0011] As a preferred embodiment of the deep learning side-channel attack method based on the Boosting ensemble method described in this invention, wherein: the step of performing wavelet denoising on the side-channel signal sequence to obtain a denoised signal sequence includes: determining the wavelet basis function and the number of wavelet decomposition layers using a wavelet decomposition parameter determination method; The side channel signal sequence is subjected to multi-level discrete wavelet decomposition. In each decomposition, the output of the previous layer is convolved and downsampled by low-pass and high-pass filters respectively to obtain the approximation coefficients and detail coefficients of the current layer. A thresholding method is used to suppress noise in the detail coefficients of each layer. The thresholding method adjusts the detail coefficients according to the relationship between the detail coefficients and the preset threshold. Wavelet reconstruction is performed based on the approximation coefficients and detail coefficients after thresholding, and the denoised signal sequence is obtained through upsampling and filter convolution operations.

[0012] The beneficial effects of this preferred technical solution are as follows: by selecting appropriate wavelet basis functions and decomposition levels through wavelet decomposition parameter determination, the side channel signal is decomposed into different frequency components using the multi-scale characteristics of wavelet transform, and the noise components in the high-frequency detail coefficients are suppressed by using threshold processing. While removing measurement noise and electromagnetic interference, the key-related leakage information contained in the low-frequency approximation coefficients is preserved, avoiding the loss of useful signal components that may be lost by traditional filtering methods, and improving the feature extraction quality and key recovery efficiency of subsequent deep learning models.

[0013] As a preferred embodiment of the deep learning side-channel attack method based on the Boosting ensemble method described in this invention, the multilayer perceptron structure includes: an input layer receiving feature vectors; The first hidden layer performs a linear transformation on the output of the input layer using a weight matrix and a bias term. The result of the linear transformation is then subjected to a non-linear transformation using an activation function to obtain the output of the first hidden layer. The second hidden layer performs a linear transformation on the output of the first hidden layer using a weight matrix and a bias term. The result of the linear transformation is then subjected to a non-linear transformation using an activation function to obtain the output of the second hidden layer. The output layer performs a linear transformation on the output of the second hidden layer using a weight matrix and a bias term to obtain the logit value corresponding to each category.

[0014] As a preferred embodiment of the deep learning side-channel attack method based on the Boosting ensemble method described in this invention, the initialization method of each layer weight matrix is ​​as follows: the initialization interval of the first hidden layer weight matrix is ​​determined, and the upper and lower bounds of the initialization interval are calculated based on the number of neurons in the input layer. Determine the initialization interval of the weight matrix of the second hidden layer. The upper and lower bounds of the initialization interval are calculated based on the number of neurons in the first hidden layer. Determine the initialization interval of the output layer weight matrix. The upper and lower bounds of the initialization interval are calculated based on the number of neurons in the second hidden layer. Randomly sample values ​​from the initialization intervals corresponding to each layer and assign them to the corresponding weight matrix.

[0015] As a preferred embodiment of the deep learning side-channel attack method based on the Boosting ensemble method described in this invention, the step of training a weak learner using the current sample weights in each iteration and then updating the sample weights according to the classification error of the trained weak learner includes: constructing a weighted loss function, wherein the loss contribution of each training sample in the weighted loss function is weighted by the corresponding sample weights. The current weak learner is trained by minimizing the weighted loss function, and the prediction results of the trained weak learner for each sample are obtained. The classification accuracy of each sample is determined based on the difference between the predicted results and the true labels; The weights of each sample are updated based on the classification accuracy; the weights of misclassified samples are increased, while the weights of correctly classified samples are decreased.

[0016] The beneficial effects of this preferred technical solution are as follows: by constructing a weighted loss function, the model pays more attention to samples with higher weights during training. The sample weights are dynamically updated based on the classification results of the current weak learner, and the weights of samples with weak key leakage and difficult to classify correctly are increased. This allows the subsequent weak learners to focus on learning these difficult samples. After rounds of iteration, multiple weak learners form complementary prediction capabilities, avoiding the attack efficiency loss caused by a single model ignoring difficult-to-classify samples, and improving the overall learning effect of the model on side-channel signals under different signal-to-noise ratio conditions.

[0017] As a preferred embodiment of the deep learning side-channel attack method based on the Boosting ensemble method described in this invention, the step of weightedly fusing the outputs of multiple trained weak learners to obtain the ensemble prediction probability distribution includes: obtaining the logit vector of each weak learner's output to the input sample; The fusion logit vector is obtained by weighted summing of the logit vectors of each weak learner according to the fusion weight coefficients corresponding to each weak learner. The fusion weight coefficients are determined by minimizing the prediction loss on the validation set. The ensemble prediction probability distribution is obtained by performing softmax normalization on the fused logit vector.

[0018] As a preferred embodiment of the deep learning side-channel attack method based on the Boosting ensemble method described in this invention, the step of calculating the cumulative score of the candidate keys according to the ensemble prediction probability distribution, sorting the candidate keys according to the cumulative score to obtain a key sorting list, and obtaining the position index of the correct key in the key sorting list as the guessing entropy includes: for each side-channel signal sequence, obtaining the prediction probability value of each category from the ensemble prediction probability distribution; For each candidate key, based on the median label corresponding to the candidate key in the current side channel signal sequence, extract the probability value corresponding to the median label from the predicted probability value, and calculate the logarithm of the probability value as the score of the candidate key in the current side channel signal sequence. The scores of each candidate key on all side channel signal sequences are summed to obtain the cumulative score of the candidate key; Sort all candidate keys in descending order of cumulative score to form a key sorting list; The position index of the correct key in the key sorting list is used as the guess entropy.

[0019] This invention provides a deep learning side-channel attack system based on the Boosting ensemble method.

[0020] To address the aforementioned technical problems, this invention provides the following technical solution: a deep learning side-channel attack system based on the Boosting ensemble method, comprising: a signal acquisition module for acquiring side-channel signal sequences and corresponding real key tags; The signal preprocessing module is used to perform wavelet denoising on the side channel signal sequence to obtain a denoised signal sequence, and to extract feature vectors from the denoised signal sequence to construct a training sample set. The weak learner building module is used to build multiple weak learners with a multilayer perceptron structure. The Boosting training module is used to assign sample weights to samples in the training sample set, train multiple weak learners in multiple iterations and update sample weights according to classification error. The integration and fusion module is used to weight and fuse the outputs of multiple trained weak learners to obtain an integrated prediction probability distribution. The key evaluation module is used to calculate the cumulative score of candidate keys based on the integrated prediction probability distribution and sort them to obtain a key sorting list. The position index of the correct key in the key sorting list is obtained as the guessing entropy.

[0021] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the deep learning side-channel attack method based on the Boosting ensemble method.

[0022] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the deep learning side-channel attack method based on the Boosting ensemble method.

[0023] The beneficial effects of this invention are as follows: By performing wavelet denoising on the side-channel signal sequence and utilizing the multi-scale decomposition characteristics of wavelet transform to separate the signal into different frequency components, high-frequency noise components are suppressed while low-frequency key leakage information is preserved. Compared to directly using the original signal or simple filtering methods, this method can remove measurement noise and electromagnetic interference while avoiding the loss of useful signals, improving the feature extraction quality of the deep learning model and providing clearer input features for subsequent key recovery. By dynamically adjusting the sample weights based on the classification error of the weak learners in multiple iterations, misclassified samples receive higher weights in subsequent iterations, achieving focused learning of difficult-to-classify samples with weak key leakage. Compared to the traditional method of training all samples with equal weights, this avoids the problem of insufficient learning of difficult examples caused by the model being dominated by easily classified samples. By training multiple weak learners under different sample weight distributions, complementary predictive capabilities are formed, improving the overall attack efficiency of the model on side-channel signals under different signal-to-noise ratio conditions. By weighted fusion of the outputs of multiple weak learners after training and optimizing the fusion weight coefficients using the validation set, the feature representations and prediction capabilities learned by each of the multiple weak learners are integrated. Compared with a single deep learning model, this reduces the risk of overfitting the model to specific noise patterns in the training data and improves the robustness and generalization ability of the model under different measurement devices, different environmental noise, and adversarial interference conditions, enabling the attack method to adapt to a wider range of practical application scenarios. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating the overall process of a deep learning side-channel attack method based on the Boosting ensemble approach, as provided in one embodiment of the present invention.

[0026] Figure 2 The figure shows simulation experimental data of a deep learning side-channel attack method based on the Boosting ensemble method provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0028] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a deep learning side-channel attack method based on the Boosting ensemble method, including: Step 1: Obtain the side channel signal sequence and the corresponding real key tag. The side channel signal sequence is obtained by collecting the power consumption signal during the operation of the encryption device. Step 2: Perform wavelet denoising on the side channel signal sequence to obtain a denoised signal sequence; Step 3: Extract feature vectors from the denoised signal sequence, and construct a training sample set based on the feature vectors and the real key labels; Step 4: Construct multiple weak learners, which adopt a multilayer perceptron structure; Step 5: Assign initial sample weights to the samples in the training sample set, and train multiple weak learners in multiple iterations. In each iteration, train a weak learner using the current sample weights and update the sample weights according to the classification error of the trained weak learners, so that misclassified samples will get higher weights in subsequent iterations. Step 6: Weight and fuse the outputs of the trained weak learners to obtain the ensemble prediction probability distribution; Step 7: Calculate the cumulative score of the candidate keys based on the ensemble prediction probability distribution, sort the candidate keys according to the cumulative score to obtain the key sorting list, and obtain the position index of the correct key in the key sorting list as the guessing entropy.

[0029] Side-channel attacks infer keys by analyzing physical information leaked during encryption operations in cryptographic devices. Power consumption side-channel signals reflect the differences in energy consumption of the device's internal circuitry when processing different data; these differences are related to the intermediate value processed, which is determined by both the plaintext and the key. Traditional side-channel attack methods rely on manually designed statistical models and prior knowledge, and their effectiveness is limited against protected devices or in high-noise environments. Deep learning methods automatically extract key-related features from side-channel signals through end-to-end learning; however, single deep learning models suffer from insufficient learning ability for difficult-to-classify samples and lack of robustness.

[0030] This invention combines a deep learning model with a Boosting ensemble method, utilizing wavelet denoising preprocessing to improve signal quality, a dynamic sample weighting mechanism to enhance hard example learning, and weighted fusion of multiple weak learners to improve model robustness. In step 2, wavelet denoising leverages the multi-scale decomposition properties of wavelet transform to decompose the side-channel signal into different frequency components, suppressing high-frequency noise while preserving low-frequency key leakage information, providing high-quality input for subsequent feature extraction. In step 5, Boosting iterative training dynamically adjusts sample weights, ensuring each newly trained weak learner pays more attention to samples misclassified by previous weak learners. These samples often correspond to subtle key leakage trajectories. Through multiple iterations, multiple weak learners form complementary predictive capabilities, covering attack scenarios under different signal-to-noise ratios. In step 6, weighted fusion optimizes the fusion weight coefficients on the validation set, integrating the feature representations learned by multiple weak learners, reducing the risk of overfitting in a single model, and improving the model's generalization ability under different measurement devices and noise environments.

[0031] Example 2, an embodiment of the present invention, provides a deep learning side-channel attack method based on the Boosting ensemble method, based on the previous embodiment, including: Step 2: Wavelet denoising is performed on the side channel signal sequence to obtain the denoised signal sequence, including the following steps A1-A4: A1: The wavelet basis function and the number of wavelet decomposition levels are determined using the wavelet decomposition parameter determination method; A2: Perform multi-level discrete wavelet decomposition on the side channel signal sequence. In each decomposition, the output of the previous layer is convolved and downsampled by low-pass and high-pass filters respectively to obtain the approximation coefficients and detail coefficients of the current layer. A3: A thresholding method is used to suppress noise in the detail coefficients of each layer. The thresholding method adjusts the detail coefficients according to the relationship between the detail coefficients and the preset threshold. A4: Wavelet reconstruction is performed based on the approximation coefficients and detail coefficients after thresholding, and the denoised signal sequence is obtained through upsampling and filter convolution operations.

[0032] In this embodiment, in step A1, the wavelet decomposition parameters are determined by selecting the db4 wavelet basis function from the Daubechies wavelet family. The db4 wavelet basis has a fourth-order vanishing moment characteristic, which can effectively separate smooth and abrupt components in the side-channel signal. Its compact support characteristic ensures the computational efficiency of decomposition and reconstruction. The number of wavelet decomposition layers is determined according to the number of sampling points and frequency characteristics of the side-channel signal. When the signal length is long, a larger number of decomposition layers is selected to obtain more detailed frequency division. For typical side-channel signals, the number of decomposition layers is set to 3 to 5. Three-layer decomposition is suitable for scenarios with short signal length or low noise level, while five-layer decomposition is suitable for scenarios with long signal length or requiring more detailed frequency analysis. Too few decomposition layers will result in insufficient frequency resolution, making it difficult to effectively separate noise and signal; too many decomposition layers will increase computational complexity and may over-decompose useful signal components.

[0033] In an optional implementation, in step A1, the wavelet decomposition parameters can be determined by selecting the sym8 wavelet basis function from the Symlet wavelet family. The Symlet wavelet is an approximately symmetrical version of the Daubechies wavelet, possessing better symmetry and reducing phase distortion introduced by wavelet transform, making it suitable for side-channel attack scenarios where the signal's time-domain characteristics need to be preserved. The number of decomposition levels is adaptively determined based on the signal-to-noise ratio characteristics of the signal. By calculating the energy distribution of the detail coefficients after each decomposition level, further decomposition stops when the energy proportion of a certain level's detail coefficients falls below a preset threshold.

[0034] In another optional implementation, in step A1, the wavelet decomposition parameters can be determined by selecting the coif3 wavelet basis function from the Coiflet wavelet family as the wavelet basis function. The Coiflet wavelet possesses the vanishing moment characteristics of both wavelet and scaling functions, enabling good frequency localization in both the wavelet and scaling domains, making it suitable for complex scenarios where key leakage information is distributed across multiple frequency bands. The number of decomposition layers is determined based on the total length of the side-channel signal. The base-2 logarithm of the total signal length is subtracted from the fixed offset to obtain the number of decomposition layers, ensuring that the coarsest-scale approximation coefficients still contain enough sampling points to retain low-frequency information.

[0035] In this embodiment, step A3 involves adjusting the detail coefficients of each layer using a soft thresholding method. For each detail coefficient, the difference between its absolute value and a preset threshold is calculated. When the absolute value of a detail coefficient is less than or equal to the preset threshold, the detail coefficient is set to zero, achieving complete suppression of small-amplitude noise components. When the absolute value of a detail coefficient is greater than the preset threshold, the sign of the detail coefficient is retained, and the result of subtracting the threshold from its absolute value is used as the processed detail coefficient, achieving preservation of large-amplitude signal components while reducing noise impact. Soft thresholding avoids the discontinuities that may be introduced by hard thresholding through continuous shrinking operations, making the reconstructed signal smoother. The preset threshold is calculated based on the standard deviation of the detail coefficients and the signal length, ensuring that the threshold can adapt to side-channel signals with different noise levels.

[0036] In an optional implementation, in step A3, the thresholding method can be achieved by adjusting the detail coefficients of each layer using a hard thresholding approach. For each detail coefficient, its absolute value is compared with a preset threshold. When the absolute value of a detail coefficient is less than the preset threshold, the coefficient is considered to be mainly contributed by noise and is directly set to zero. When the absolute value of a detail coefficient is greater than or equal to the preset threshold, the coefficient is considered to contain useful signal components and its original value remains unchanged. Hard thresholding achieves simple and fast noise suppression through binarization, and is suitable for scenarios where noise and signal are clearly distinguishable in amplitude. The preset threshold is determined by estimating the noise variance and combining it with the probability distribution of the detail coefficients, so that most noise coefficients are set to zero while signal coefficients are preserved.

[0037] In another optional implementation, in step A3, the threshold processing can also be achieved by adjusting the detail coefficients of each layer using a compromise thresholding method. This compromise thresholding method combines the continuity of a soft threshold with the preservation of a hard threshold. When the absolute value of a detail coefficient is less than a preset threshold, it is set to zero; when the absolute value is greater than the preset threshold, a shrinkage function between the soft and hard thresholds is used for adjustment. Specifically, an adjustment parameter is defined to control the degree of shrinkage. When the adjustment parameter is zero, it is equivalent to a hard threshold; when the adjustment parameter is one, it is equivalent to a soft threshold; and intermediate values ​​achieve different degrees of shrinkage.

[0038] Step 4: Construct multiple weak learners. The weak learners adopt a multilayer perceptron structure, and the weak learners include the following structures B1-B4: B1: The input layer receives feature vectors; B2: The first hidden layer performs a linear transformation on the output of the input layer through a weight matrix and a bias term, and then performs a non-linear transformation on the result of the linear transformation through an activation function to obtain the output of the first hidden layer. B3: The second hidden layer performs a linear transformation on the output of the first hidden layer using a weight matrix and a bias term. The result of the linear transformation is then subjected to a non-linear transformation using an activation function to obtain the output of the second hidden layer. B4: The output layer performs a linear transformation on the output of the second hidden layer using the weight matrix and bias terms to obtain the logit value corresponding to each category.

[0039] Specifically, the initialization method for the weight matrices of each layer is as follows: determine the initialization interval of the weight matrix of the first hidden layer, and calculate the upper and lower bounds of the initialization interval based on the number of neurons in the input layer; Determine the initialization interval of the weight matrix of the second hidden layer. The upper and lower bounds of the initialization interval are calculated based on the number of neurons in the first hidden layer. Determine the initialization interval of the output layer weight matrix. The upper and lower bounds of the initialization interval are calculated based on the number of neurons in the second hidden layer. Randomly sample values ​​from the initialization intervals corresponding to each layer and assign them to the corresponding weight matrix.

[0040] In step 4, the feature vector received by the input layer comes from the wavelet-denoised signal sequence, and the dimension of the feature vector is equal to the length of the selected signal segment. The first and second hidden layers undergo linear transformations using weight matrices and bias terms, followed by nonlinear transformations using the ReLU activation function. The output layer does not use an activation function and directly outputs the logit value, the dimension of which is equal to the number of categories in the classification task. For side-channel attacks on the AES encryption algorithm, the number of categories is 256, corresponding to the 256 possible values ​​for one byte. The structural parameters of multiple weak learners can be the same or different. Through different weight initializations and sample weight distributions, each weak learner learns different feature representations and prediction patterns.

[0041] The initialization interval of the weight matrix is ​​calculated based on the number of input neurons in that layer. For the first and second hidden layers, the weight matrix uses the Kaiming initialization method, with an initialization interval of [interval to be filled in]. Where fan_in is the number of input neurons in this layer. For the output layer, the weight matrix uses the Xavier initialization method, with an initialization interval of... ,in and These represent the number of neurons in the first and second hidden layers, respectively. Random samples are taken from the initialization intervals corresponding to each layer and assigned to the corresponding weight matrix.

[0042] Step 5: Assign initial sample weights to the samples in the training sample set, and train multiple weak learners sequentially in multiple iterations. In each iteration, a weak learner is trained using the current sample weights, and the sample weights are updated based on the classification error of the trained weak learners, so that misclassified samples will receive higher weights in subsequent iterations. This includes the following steps C1-C4: C1: Construct a weighted loss function, in which the loss contribution of each training sample in the weighted loss function is weighted by the corresponding sample weights; C2: Train the current weak learner by minimizing the weighted loss function to obtain the prediction results of the trained weak learner for each sample; C3: Determine the classification accuracy of each sample based on the difference between the predicted results and the true labels; C4: Update the weights of each sample based on the classification correctness. Increase the weight of misclassified samples and decrease the weight of correctly classified samples.

[0043] In this embodiment, step 5 updates the sample weights by employing a weighted cross-entropy loss function to influence the training process. For each sample in the training sample set, the cross-entropy loss of that sample is calculated. The cross-entropy loss measures the difference between the predicted probability distribution and the true label. The cross-entropy loss of each sample is multiplied by the corresponding sample weight to obtain the weighted loss contribution of that sample. The weighted loss contributions of all samples are summed, and the L2 regularization term of the weight matrix is ​​added to form the complete weighted loss function. The weighted loss function is minimized using the gradient descent algorithm to update the weight parameters of the weak learner. After training, the trained weak learner is used to predict all training samples, and the predicted class is compared with the true label to determine the classification correctness of each sample. The sample weights are updated according to the classification results. For misclassified samples, their weights are multiplied by a growth factor greater than 1; for correctly classified samples, their weights are multiplied by a decay factor less than 1. After the weights are updated, all sample weights are normalized so that the sum of the weights is 1, ensuring that the weight distribution is always an effective probability distribution.

[0044] In an optional implementation, in step 5, the sample weights are updated by using a gradient boosting residual fitting method. In each iteration, the residual between the current ensemble model's predicted output for each sample and the true label is calculated; the residual represents information that the current model has not yet fitted. The residual is used as a new training target, and the weak learner in the current iteration is trained to fit these residuals. Samples with larger residuals indicate that the current model has a larger prediction error for them. By having the new weak learner focus on fitting these samples with large residuals, implicit weighting of hard-to-classify samples is achieved. The absolute or squared value of the residual can be converted into equivalent sample weights, making the gradient boosting method mathematically equivalent to the explicit sample weighting method. After training, the newly trained weak learner is added to the ensemble model with a certain learning rate, updating the ensemble model's predicted output. In each iteration, the weak learner focuses on fitting the residuals of the previous model, ultimately forming a refined and progressively approximating overall prediction.

[0045] In another optional implementation, in step 5, updating the sample weights can also be achieved by adjusting the sample weights using the exponential weight update method of the AdaBoost algorithm. After training the current weak learner, the weighted error rate of the weak learner on the training set is calculated. The weighted error rate is the sum of the weights of misclassified samples divided by the sum of the weights of all samples. The confidence coefficient of the current weak learner is calculated based on the weighted error rate. The confidence coefficient is half of the weighted accuracy divided by the natural logarithm of the weighted error rate. For misclassified samples, their weights are multiplied by an exponential function value, where the exponent of the exponential function is the confidence coefficient, achieving exponential growth of the weights of misclassified samples. For correctly classified samples, their weights are multiplied by an exponential function value, where the exponent of the exponential function is the negative confidence coefficient, achieving exponential decay of the weights of correctly classified samples.

[0046] Step 6: Weighted fusion of the outputs of multiple trained weak learners to obtain the ensemble prediction probability distribution includes the following steps D1-D3: D1: Obtain the logit vector output by each weak learner for the input sample; D2: The fusion logit vector is obtained by weighted summing of the logit vectors of each weak learner according to the fusion weight coefficients corresponding to each weak learner. The fusion weight coefficients are determined by minimizing the prediction loss on the validation set. D3: Softmax normalization is applied to the fused logit vector to obtain the ensemble prediction probability distribution.

[0047] In this embodiment, step 6 involves weighted fusion by integrating the outputs of multiple weak learners using a logit-level linear weighted fusion method. For each input sample, the output layer of each weak learner generates a logit vector, the dimension of which is equal to the number of classes. The logit vector of each weak learner is multiplied by its corresponding fusion weight coefficient, which is a scalar value representing the importance of the weak learner in the ensemble model. All weighted logit vectors are summed dimensionally to obtain the fused logit vector. The fusion weight coefficient is determined through optimization on the validation set. Using the fusion weight coefficient as the optimization variable and the cross-entropy loss on the validation set as the optimization objective, the gradient descent algorithm is used to iteratively update the fusion weight coefficient until the loss converges. To prevent a single weak learner from having an excessively large weight and dominating the fusion result, an L2 regularization term for the fusion weight coefficient is added to the optimization objective. The fused logit vector is converted into a probability distribution using a softmax function. The softmax function takes the exponent of each element of the logit vector and then divides it by the sum of all element exponents to obtain a probability distribution with a sum of 1, which is used as the ensemble prediction probability distribution for subsequent key recovery.

[0048] In an optional implementation, in step 6, weighted fusion can be performed by integrating the outputs of multiple weak learners using a probabilistic weighted average fusion method. For the logit output of each weak learner, softmax normalization is performed independently to obtain the predicted probability distribution of that weak learner. The predicted probability distribution of each weak learner is multiplied by the corresponding fusion weight coefficient, which satisfies the non-negativity constraint and sums to 1, representing the confidence weight of each weak learner. All weighted probability distributions are summed dimensionally to obtain the integrated predicted probability distribution. The fusion weight coefficients are determined through optimization on the validation set. Since the optimization problem of probabilistic weighted fusion is non-convex, constrained optimization algorithms such as projective gradient descent or sequential minimum optimization are used to solve it. Compared with logit-level fusion, probabilistic fusion normalizes the outputs of each weak learner, making the output scale of different weak learners uniform, but may lose the absolute magnitude information of the logit value.

[0049] In another optional implementation, in step 6, weighted fusion can also be performed by integrating the outputs of multiple weak learners using a weighted voting fusion method. For each weak learner, the class with the highest probability from its predicted probability distribution is selected as the voting result for that weak learner. A voting weight is assigned to each weak learner, determined based on its classification accuracy on the validation set; weak learners with higher accuracy have greater voting weights. For each class, the sum of the voting weights of all weak learners voting for that class is calculated to obtain the weighted voting score for that class. The weighted voting scores for each class are normalized so that their sum equals 1, resulting in the integrated predicted probability distribution. The weighted voting method is simple and intuitive, suitable for scenarios where the performance of weak learners varies significantly. By setting the voting weights, the influence of poorly performing weak learners can be effectively suppressed. However, weighted voting only utilizes the highest probability class information of each weak learner and does not fully utilize the complete probability distribution information.

[0050] Step 7: Calculate the cumulative score of the candidate keys based on the ensemble prediction probability distribution, sort the candidate keys according to the cumulative score to obtain the key sorting list, and obtain the position index of the correct key in the key sorting list as the guessing entropy, including the following steps E1-E5: E1: For each side-channel signal sequence, obtain the predicted probability value for each category from the integrated prediction probability distribution; E2: For each candidate key, based on the median value label corresponding to the candidate key in the current side channel signal sequence, extract the probability value corresponding to the median value label from the predicted probability value, and calculate the logarithm of the probability value as the score of the candidate key in the current side channel signal sequence. E3: The scores of each candidate key on all side channel signal sequences are summed to obtain the cumulative score of the candidate key; E4: Sort all candidate keys according to their cumulative scores from highest to lowest to form a key sorting list; E5: Find the position index of the correct key in the key sorting list as the guess entropy.

[0051] In step 7, for each candidate key, the intermediate value of the encryption algorithm is calculated based on the known plaintext and the candidate key. A byte of this intermediate value is used as the intermediate value label for that candidate key in the current side-channel signal sequence. The ensemble prediction probability distribution provides the probability of each category (i.e., each intermediate value). The probability value corresponding to the intermediate value label is extracted from this distribution, and the logarithm of this probability value is calculated as the score of the candidate key in the current side-channel signal sequence. The scores of each candidate key in all side-channel signal sequences are summed to obtain the cumulative score of the candidate key. All candidate keys are sorted from highest to lowest according to their cumulative scores. The index of the correct key in the sorted list is the guess entropy, starting from 0.

[0052] Example 3, an embodiment of the present invention, provides a deep learning side-channel attack method based on the Boosting ensemble method, based on the previous embodiment, including: In step 1, side channel signals (power consumption, electromagnetic fields, etc.) are collected and ASCADf is used as the training set.

[0053] In step 2, the wavelet denoising preprocessing includes: processing the original sampled sequence... conduct Layer discrete wavelet decomposition yields the th Approximation coefficient of the layer With detail coefficient Satisfies the decomposition expression:

[0054] in, To determine the number of decomposition layers, a selection is typically made. ; Motherwavelet bases can include Daubechies (dbN, e.g., db4), Symlet, Coiflet, etc., with db4 / db6 being the most commonly used. Decomposition is achieved through convolutional downsampling, and is represented as:

[0055] in, These are low-pass and high-pass reconstruction / decomposition filters, respectively.

[0056] This implementation limits the weak learner to a three-layer multilayer perceptron (MLP), each of which includes an input layer, two hidden layers, and an output layer. The mathematical form and implementation details are given below in sub-steps.

[0057] Model Structure and Forward Computation: Let the input of a single weak learner be the feature vector. (From the denoised windowed output), the network parameters are the weight matrix and bias terms:

[0058] in The number of units in the first and second hidden layers. For the number of categories, For input dimensions.

[0059] Forward propagation:

[0060]

[0061]

[0062] If Dropout / BatchNorm is used, insert at the appropriate position:

[0063]

[0064] in This is the dropout ratio.

[0065] Parameter initialization and randomization include the following steps: Weight initialization: hidden layer Initialize using Kaiming / He:

[0066] Output layer Initialize using Xavier / Glorot:

[0067] Training loss, weighted sample training and optimization strategies Weighted cross-entropy (for AdaBoost / reweighting scenarios): If the training sample set is ,in Given the sample weights (from the Boosting iteration), the weighted loss of a single weak learner is defined as:

[0068] in , is the L2 regularization coefficient.

[0069] Weighted variant of loss (residual fitting / GradientBoosting): When using the GradientBoosting approach, a weak learner can be trained to fit the negative gradient (residual) of the current model. The residuals are used as the regression target and trained with squared loss or vector regression loss.

[0070] Weighted training implementation: During training, the sample weights are adjusted. It can be used directly for weighted loss or achieve an equivalent effect through repeated sampling.

[0071] Output: Obtain a single weak learner and form a set of weak learners: After completing the above initialization and training, a set of trained three-layer MLP weak learners is obtained. ,in This represents the number of weak learners, and their outputs (logits or calibrated probabilities) are saved for subsequent Boosting training and fusion.

[0072] Preferred weight initialization and regularization strategies for a three-layer MLP include the following options: using Kaiming / He initialization for hidden layer weights, using Xavier initialization for output layer weights, and optional Dropout, BatchNorm, or L2 regularization to prevent overfitting.

[0073] In the Boosting iterative training (dynamically adjusting sample weights to focus on hard-to-classify samples), the input data are shown in Tables 1, 2, and 3: Table 1 Required Input Data for Boosting Iterative Training

[0074] Table 2 Optional Input Data for Boosting Iterative Training

[0075] Table 3: Specific parameters for each module / level

[0076] Weighted fusion is achieved through logit-level linear fusion: Let there be T trained three-layer MLP weak learners. The weak learners work on the input... Output logits vector ( (Number of categories, e.g., 256). Define scalar weights. (Positive or negative, or restricted to non-negative) The logits of each weak learner are linearly weighted, and the fused logits are:

[0077] The final predicted probability is obtained through softmax:

[0078] Weight Through the validation set Learning by minimizing cross-entropy loss:

[0079] Regularization terms can be introduced, such as Controlling the weights prevents a single weak learner from dominating the fusion result. To ensure non-negativity or probabilistic meaning, the weights can be adjusted. Perform softmax parameterization: and directly to the parameters Perform gradient optimization.

[0080] Probability-level weighted average: If the weak learner outputs a calibrated probability vector Then, a weighted average of the probabilities can be directly applied:

[0081] Among them, weight Weight The result is obtained by minimizing the cross-entropy on the validation set:

[0082] For easier optimization, unconstrained parameters can be used. Perform softmax mapping: and to Perform gradient descent.

[0083] Key prediction and performance evaluation include the following steps: Guessing entropy calculation: before After trajectories are calculated, the cumulative score of all keys is determined. Sorted in descending order, correct keys The index in the sorted results (counting from 0) is the [index number]. The formula for the guessed entropy corresponding to each trajectory is:

[0084] in Indicates: set After sorting in descending order, the elements The position index (index 0 indicates the first position, which is the best guess).

[0085] Initialization parameters: Total number of tracks read ; Initialize key score array (Length 256, corresponding to 256 keys), all with an initial value of 0 (the initial cumulative score of all keys is 0); Initialize the guessed entropy sequence (Empty list), used to store the guessed entropy after each additional trajectory.

[0086] Iterate through each trajectory, accumulate the score, and calculate the guess entropy: For the Trajectory ( From 0 to First, extract the predicted probability distribution of the current trajectory. (A probability array of length M), then iterate through all 256 keys. ( (from 0 to 255): Find the current key In the Labels corresponding to each trajectory ; Calculate the current key Logarithmic contribution under this trajectory: Add the contribution to the key. In the total score: ; For the cumulative score array Arrange them in descending order to obtain the sorted key index sequence; Find the correct key in the sorted sequence. The position index (i.e., ranking) is added to the ranking. sequence; Return the guessed entropy sequence: After traversing all trajectories, return... A sequence (of length N), where each element corresponds to "before use". "Ranking of the correct key when there is a trajectory".

[0087] Wavelet denoising is used to preserve key leakage features, and a boosting strategy is used to enhance learning of difficult examples. Weighted fusion of multiple weak learners improves the robustness of the model and adapts it to high-noise and cross-device scenarios.

[0088] Example 4, refer to Figure 2 This invention provides a deep learning side-channel attack method based on the Boosting ensemble method, as one embodiment of the present invention. To verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0089] The experiment uses the ASCADf dataset to compare the key recovery performance of the Boosting ensemble method of this invention with that of the traditional single-model method.

[0090] The evaluation metric is Guessing Entropy (GE). The lower the GE value, the better the key recovery effect. GE=0 means that the target key is completely recovered.

[0091] The data file includes two sets of GE data: `GE_per_trace_boosting.csv` records the GE values ​​of the Boosting integration method of this invention under 1-2000 side channel trajectories; `GE_per_trace_original.csv` records the GE data of the single model method, including GE_original (baseline single model) and GE_50steps / GE_80steps / GE_100steps (single model variants with different iteration steps).

[0092] Visualization of experimental data, such as Figure 2 As shown, the analysis is as follows: Verification of wavelet denoising preprocessing: The method of this invention reduced GE to 4 with only 10 traces, and GE=0 and remained stable with 25 traces; while the single model method still maintained GE above 20 even after 200 traces (GE_original is 24, GE_100steps is 92), and the fluctuation was large.

[0093] Wavelet denoising effectively filters out noise interference, enabling the model to quickly capture key-related features. GE convergence speed far exceeds that of single-model methods.

[0094] Validation of Boosting iterative training: The GE of the method of this invention drops rapidly from the initial 55, GE=1 at 22 traces, and converges completely to 0 after 25 traces; while GE_original, which is the best performing single model method, still has some data with GE greater than 1 at 2000 traces (such as GE=1 at 1969 traces), and the overall convergence speed is extremely slow.

[0095] The Boosting iterative weighting mechanism reduced the number of traces required for key recovery from thousands to 25, validating the effectiveness of hard example learning.

[0096] Verification of weighted fusion: After 25 traces, the GE of the method of this invention remains stable at 0 for a long time, with only brief fluctuations (GE=1) in a very few traces (such as 70 and 77), and the overall stability is extremely strong; while the GE value of the single model method always fluctuates greatly (GE_100steps is 143 when there are 50 traces and is still 92 when there are 200 traces), and there is no trend of stable convergence to 0.

[0097] Multi-model fusion offsets the bias and variance of a single model, keeping the attack effect stable under different trace numbers.

[0098] Example 5 is an embodiment of the present invention. This embodiment provides a deep learning side-channel attack system based on the Boosting ensemble method, including: a signal acquisition module for acquiring side-channel signal sequences and corresponding real key tags; The signal preprocessing module is used to perform wavelet denoising on the side channel signal sequence to obtain a denoised signal sequence, and to extract feature vectors from the denoised signal sequence to construct a training sample set. The weak learner building module is used to build multiple weak learners with a multilayer perceptron structure. The Boosting training module is used to assign sample weights to samples in the training sample set, train multiple weak learners in multiple iterations and update sample weights according to classification error. The integration and fusion module is used to weight and fuse the outputs of multiple trained weak learners to obtain an integrated prediction probability distribution. The key evaluation module is used to calculate the cumulative score of candidate keys based on the integrated prediction probability distribution and sort them to obtain a key sorting list. The position index of the correct key in the key sorting list is obtained as the guessing entropy.

[0099] This embodiment also provides an electronic device applicable to the deep learning side-channel attack method based on the Boosting ensemble method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the deep learning side-channel attack method based on the Boosting ensemble method proposed in the above embodiment.

[0100] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the deep learning side-channel attack method based on the Boosting ensemble method proposed in the above embodiments.

[0101] The storage medium proposed in this embodiment and the deep learning side-channel attack method based on the Boosting integration method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0102] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning side-channel attack method based on the Boosting ensemble approach, characterized by: include, The side channel signal sequence and the corresponding real key tag are obtained. The side channel signal sequence is obtained by collecting the power consumption signal during the operation of the encryption device. Wavelet denoising is performed on the side channel signal sequence to obtain a denoised signal sequence. Feature vectors are extracted from the denoised signal sequence, and a training sample set is constructed based on the feature vectors and the real key labels. Construct multiple weak learners, which adopt a multilayer perceptron structure; The samples in the training sample set are assigned initial sample weights. Multiple weak learners are trained sequentially in multiple iterations. In each iteration, a weak learner is trained using the current sample weights. The sample weights are then updated based on the classification error of the trained weak learners, so that misclassified samples will receive higher weights in subsequent iterations. The outputs of multiple trained weak learners are weighted and fused to obtain the ensemble prediction probability distribution. The cumulative score of the candidate keys is calculated based on the integrated prediction probability distribution. The candidate keys are then sorted according to the cumulative score to obtain a key sorting list. The position index of the correct key in the key sorting list is obtained as the guessing entropy.

2. The deep learning side-channel attack method based on the Boosting ensemble method as described in claim 1, characterized in that: The process of wavelet denoising the contra-channel signal sequence to obtain a denoised signal sequence includes: determining the wavelet basis function and the number of wavelet decomposition layers using a wavelet decomposition parameter determination method; The side channel signal sequence is subjected to multi-level discrete wavelet decomposition. In each decomposition, the output of the previous layer is convolved and downsampled by low-pass and high-pass filters respectively to obtain the approximation coefficients and detail coefficients of the current layer. A thresholding method is used to suppress noise in the detail coefficients of each layer. The thresholding method adjusts the detail coefficients according to the relationship between the detail coefficients and the preset threshold. Wavelet reconstruction is performed based on the approximation coefficients and detail coefficients after thresholding, and the denoised signal sequence is obtained through upsampling and filter convolution operations.

3. The deep learning side-channel attack method based on the Boosting ensemble method as described in claim 2, characterized in that: The multilayer perceptron structure includes: an input layer that receives feature vectors; The first hidden layer performs a linear transformation on the output of the input layer using a weight matrix and a bias term. The result of the linear transformation is then subjected to a non-linear transformation using an activation function to obtain the output of the first hidden layer. The second hidden layer performs a linear transformation on the output of the first hidden layer using a weight matrix and a bias term. The result of the linear transformation is then subjected to a non-linear transformation using an activation function to obtain the output of the second hidden layer. The output layer performs a linear transformation on the output of the second hidden layer using a weight matrix and a bias term to obtain the logit value corresponding to each category.

4. The deep learning side-channel attack method based on the Boosting ensemble method as described in claim 3, characterized in that: The initialization method for the weight matrices of each layer is as follows: determine the initialization interval of the weight matrix of the first hidden layer, and calculate the upper and lower bounds of the initialization interval based on the number of neurons in the input layer; Determine the initialization interval of the weight matrix of the second hidden layer. The upper and lower bounds of the initialization interval are calculated based on the number of neurons in the first hidden layer. Determine the initialization interval of the output layer weight matrix. The upper and lower bounds of the initialization interval are calculated based on the number of neurons in the second hidden layer. Randomly sample values ​​from the initialization intervals corresponding to each layer and assign them to the corresponding weight matrix.

5. The deep learning side-channel attack method based on the Boosting ensemble method as described in claim 4, characterized in that: The step of training a weak learner using the current sample weights in each iteration and then updating the sample weights based on the classification error of the trained weak learner includes: constructing a weighted loss function, in which the loss contribution of each training sample in the weighted loss function is weighted by the corresponding sample weights; The current weak learner is trained by minimizing the weighted loss function, and the prediction results of the trained weak learner for each sample are obtained. The classification accuracy of each sample is determined based on the difference between the predicted results and the true labels; The weights of each sample are updated based on the classification accuracy; the weights of misclassified samples are increased, while the weights of correctly classified samples are decreased.

6. The deep learning side-channel attack method based on the Boosting ensemble method as described in claim 5, characterized in that: The step of weightedly fusing the outputs of multiple trained weak learners to obtain the integrated prediction probability distribution includes: obtaining the logit vector of each weak learner's output to the input sample; The fusion logit vector is obtained by weighted summing of the logit vectors of each weak learner according to the fusion weight coefficients corresponding to each weak learner. The fusion weight coefficients are determined by minimizing the prediction loss on the validation set. The ensemble prediction probability distribution is obtained by performing softmax normalization on the fused logit vector.

7. The deep learning side-channel attack method based on the Boosting ensemble method as described in claim 6, characterized in that: The step of calculating the cumulative score of the candidate keys based on the integrated prediction probability distribution, sorting the candidate keys according to the cumulative score to obtain a key sorting list, and obtaining the position index of the correct key in the key sorting list as the guessing entropy includes: for each side channel signal sequence, obtaining the prediction probability value of each category from the integrated prediction probability distribution; For each candidate key, based on the median label corresponding to the candidate key in the current side channel signal sequence, extract the probability value corresponding to the median label from the predicted probability value, and calculate the logarithm of the probability value as the score of the candidate key in the current side channel signal sequence. The scores of each candidate key on all side channel signal sequences are summed to obtain the cumulative score of the candidate key; Sort all candidate keys in descending order of cumulative score to form a key sorting list; The position index of the correct key in the key sorting list is used as the guess entropy.

8. A deep learning side-channel attack system based on the Boosting ensemble method, employing the deep learning side-channel attack method based on the Boosting ensemble method as described in any one of claims 1 to 7, characterized in that, include: The signal acquisition module is used to acquire the side channel signal sequence and the corresponding real key tag; The signal preprocessing module is used to perform wavelet denoising on the side channel signal sequence to obtain a denoised signal sequence, and to extract feature vectors from the denoised signal sequence to construct a training sample set. The weak learner building module is used to build multiple weak learners with a multilayer perceptron structure. The Boosting training module is used to assign sample weights to samples in the training sample set, train multiple weak learners in multiple iterations and update sample weights according to classification error. The integration and fusion module is used to weight and fuse the outputs of multiple trained weak learners to obtain an integrated prediction probability distribution. The key evaluation module is used to calculate the cumulative score of candidate keys based on the integrated prediction probability distribution and sort them to obtain a key sorting list. The position index of the correct key in the key sorting list is obtained as the guessing entropy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning side-channel attack method based on the Boosting ensemble method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning side-channel attack method based on the Boosting ensemble method as described in any one of claims 1 to 7.

Images