Data enhancement method and system for side channel analysis

Abstract

The invention provides a data enhancement method and system for side channel analysis, and the method comprises the steps: 1, collecting a power consumption curve of target encryption equipment, and carrying out the preprocessing of the power consumption curve, and obtaining N preprocessed original power consumption curves containing leakage points; 2, randomly sampling in N original power consumption curves containing leakage points to form sample pairs; 3, injecting noise into the original power consumption curve in the sample pair to obtain the sample pair with the injected noise; 4, synthesizing the sample pairs to obtain M synthesized power consumption curves; and 5, carrying out distribution optimization on the synthesized power consumption curve, and combining the synthesized power consumption curve after distribution optimization with the original power consumption curve to form an enhanced power consumption curve sample set. According to the invention, the problem that the obtained information sample leaked by the hardware encryption equipment is small in data volume in the field of side channel analysis is solved, and the sample data volume and the sample diversity can be enriched; and the robustness and generalization ability of the deep learning model are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of side channel analysis, and particularly relates to a data enhancement method and system for side channel analysis. BACKGROUND

[0002] Side channel analysis is a means of using the information leaked by a hardware encryption device when it is working to reveal the internal secret information of the hardware encryption device. In side channel analysis, researchers can use advanced computer deep learning models to analyze the information (such as power consumption, electromagnetic radiation information, etc.) leaked by the hardware device, and thus provide a basis for implementing modeling attacks. However, when building a computer deep learning model, a large amount of sample data is needed, but the information leaked by the actual hardware encryption device is limited, resulting in a small amount of sample data.

[0003] Data enhancement technology can artificially process a small amount of existing samples to generate new samples, increasing the amount of sample data and sample diversity. Common data enhancement techniques include translation, flipping, adding noise, etc., but when such data enhancement techniques are used to enhance the information leaked by the hardware encryption device in side channel analysis, it can result in reduced attack efficiency when implementing modeling attacks and the enhanced sample data may not truly reflect the information leaked by the hardware encryption device.

[0004] Sample Pairing technology is a data enhancement technology applied to image classification, which randomly extracts two pictures from a data set to synthesize a new picture, and takes the label of one of the pictures as the label of the synthesized picture. When Sample Pairing technology is used to enhance sample data in side channel analysis, two power consumption curves can be combined into a new power consumption curve to enrich the amount of sample data, but if the label of one of the two power consumption curves is taken as the label of the synthesized power consumption curve as it is applied to image data enhancement, the problem of incorrect label matching can occur.

[0005] Because the power consumption characteristics generated by the hardware encryption device at different time points and in different encryption operation stages differ significantly, the number of samples of each category in the initial collected power consumption curve sample data set can be severely unbalanced, causing a skewed distribution of sample data categories, which in turn can cause the machine learning model to be trained to be more inclined to high-frequency categories and weaken the learning ability for low-frequency category samples. When the conventional Sample Pairing technology is used to enhance sample data in side channel analysis, it cannot improve the problem of unbalanced sample data. SUMMARY

[0006] To solve at least one of the above technical problems, the present application provides a data enhancement method and system for side channel analysis.

[0007] The first aspect of the present application provides a data enhancement method for side channel analysis, comprising:

[0008] Step 1: Collecting power consumption curves of target encryption devices and preprocessing them to obtain N preprocessed original power consumption curves containing leakage points

[0009] Step 2: Randomly sampling to form sample pairs in the N preprocessed original power consumption curves containing leakage points Where j = 1, 2, …, M; t ∈ [0, T]; Indicates the first original power consumption curve in the jth sample pair, Indicates the second original power consumption curve in the jth sample pair, and M represents the number of sample pairs.

[0010] Step 3: Injecting noise into the original power consumption curves in the sample pairs to obtain noise-injected sample pairs;

[0011] Step 4: Synthesizing the sample pairs to obtain M synthesized power consumption curves

[0012] Step 5: Distributively optimizing the synthesized power consumption curves, and merging the distributively optimized synthesized power consumption curves P new (t) with the original power consumption curves P raw (t) to form an enhanced power consumption curve sample set.

[0013] Preferably, in step 1, the preprocessing of the collected power consumption curves includes normalization, standardization and clipping.

[0014] Preferably, in any of the above solutions, in step 2, the two original power consumption curves in each sample pair belong to the same category and correspond to the same secret key operation phase.

[0015] Preferably, in any of the above solutions, in step 2, the sample pairs are formed by random sampling based on category weights, and the sampling probability is adjusted according to the category distribution imbalance degree γ = max c∈C (N c ) / min c∈C (N c ), where C represents a set of categories, c represents a specific category in C, N c represents the number of samples belonging to the category c, max c∈C (N c ) represents the number of samples in the category with the most samples, and min c∈C (N crepresents the number of samples of the class with the least number of samples in all classes.

[0016] Preferably, in step 3, the original power consumption curves in each sample pair are injected with noise respectively to obtain the injected-noise power consumption curves in the sample pair. wherein represents the first injected-noise power consumption curve in the jth sample pair after injection of noise. represents the second injected-noise power consumption curve in the jth sample pair after injection of noise. is the contribution of the sample pair to the injected noise, is the contribution of the sample pair to the injected noise, and are subject to Gaussian distribution with mean value of 0 and variance of σ. 2

[0017] Preferably, in step 3, the noise and satisfy the condition:

[0018]

[0019] wherein σ min is the minimum value of σ, and SNR target is the target signal-to-noise ratio.

[0020] Preferably, in step 4, the two injected-noise power consumption curves in each sample pair are synthesized to form a synthesized power consumption curve wherein represents the power consumption curve formed after synthesis of the two injected-noise power consumption curves in the jth sample pair, and a represents the synthesis weight, which ranges from 0 to 1.

[0021] Preferably, in step 4, the label of the synthesized power consumption curve is wherein is the label of the first power consumption curve in the sample pair of the synthesized power consumption curve , λ is the label of the second power consumption curve in the sample pair of the synthesized power consumption curve , and the weight coefficient λ (j) is a dynamic mixing factor calculated according to the leakage information.

[0022] Preferably, in step 4, under relevant power consumption analysis, the leakage model of the key byte is used to estimate the contribution of the sample, and the weight coefficient λ is calculated.​(j) .

[0023] Any of the above solutions is preferably that in step 4, the formula is:

[0024]

[0025] The weight coefficient λ is calculated (j) wherein and respectively represent the intermediate values v1 and v2 of the original power consumption curve and corresponding to the theoretical power consumption value under the leakage model.

[0026] Any of the above solutions is preferably that in step 5, the distribution optimization target function of the synthesized power consumption curve P new (t) is:

[0027]

[0028] wherein C represents a set of categories, c represents a specific category in C, p c and q c respectively represent the conditional distribution of the synthesized power consumption curve and the original power consumption curve in the category c, D KL (p c ||q c ) represents the KL divergence of p c and q c , ω is a balance hyperparameter, represents an expected operation, τ c is the class center vector of all samples of the original power consumption curve in which the category is c.

[0029] Any of the above solutions is preferably that in step 5, the loss function of the distribution optimization of the synthesized power consumption curve P new (t) is:

[0030]

[0031] wherein θ is the parameter space of the synthesis process, M c is the number of the synthesized power consumption curve in which the category is c.

[0032] Any of the above solutions is preferably that in step 5, the optimization trajectory of the distribution optimization of the synthesized power consumption curve P new (t) is:

[0033]

[0034] wherein ρ is the intensity coefficient of the equilibrium constraint, and λ is the weight coefficient of the label.

[0035] The second aspect of the application provides a data enhancement system for side channel analysis, comprising a processor, wherein a program is stored on the processor, and the processor is configured to run the program to execute the data enhancement method for side channel analysis.

[0036] Preferably, the data enhancement system for side channel analysis comprises:

[0037] a data acquisition and preprocessing module configured to acquire and preprocess power consumption curves of a target encryption device to obtain N original power consumption curves containing leakage points after preprocessing

[0038] a sample pair selection module configured to randomly sample to form sample pairs from the N original power consumption curves containing leakage points after preprocessing wherein j = 1, 2, …, M; t [0, T]; represents a first original power consumption curve in the jth sample pair, represents a second original power consumption curve in the jth sample pair, and M represents the number of sample pairs.

[0039] a noise injection module configured to inject noise into the original power consumption curves in the sample pairs to obtain noise-injected sample pairs;

[0040] a sample pair synthesis module configured to synthesize the sample pairs to obtain M synthesized power consumption curves

[0041] a distribution optimization and merging module configured to perform distribution optimization on the synthesized power consumption curves P new (t) and merge the distribution-optimized synthesized power consumption curves P new (t) with the original power consumption curves P raw (t) to form an enhanced power consumption curve sample set.

[0042] The data enhancement method and system for side channel analysis have the following beneficial effects:

[0043] 1. The data enhancement method and system solve the problems of limited information leakage of hardware encryption devices and small amount of sample data available for computer deep learning in the field of side channel analysis, enrich the amount of sample data and sample diversity through data enhancement, improve the attack efficiency when implementing modeling attacks, and the enhanced sample data can well reflect the information leakage of hardware encryption devices.

[0044] 2. The enhanced sample data can effectively simulate different noise interference in real environments and enhance the robustness of computer deep learning models.

[0045] 3. The label weight is dynamically adjusted according to the leakage energy distribution of the sample pair, which solves the problem that the label cannot be correctly matched when the sample pair is synthesized into a new sample; meanwhile, through the synthesis sample distribution optimization, the problem of sample class imbalance is solved, and the generalization ability of the computer deep learning model is improved. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 Flowchart of a preferred embodiment of the data enhancement method for side channel analysis according to the present application.

[0047] Figure 2 Principle diagram of balancing sample classes for the embodiment of the data enhancement method for side channel analysis according to the present application as shown in Figure 1

[0048] Figure 3 Result diagram of balancing sample classes for the embodiment of the data enhancement method for side channel analysis according to the present application as shown in Figure 1 DETAILED DESCRIPTION

[0049] In order to better understand the present application, the present application will be described in detail below in conjunction with specific embodiments.

[0050] Embodiment 1

[0051] As shown in Figure 1 A data enhancement method for side channel analysis, comprising:

[0052] Step 1: Collecting the power consumption curve of the target encryption device and pre-processing it to obtain N pre-processed original power consumption curves containing leakage points

[0053] Step 2: Randomly sampling from the N pre-processed original power consumption curves containing leakage points to form sample pairs Where j = 1, 2, …, M; t ∈ [0, T]; represents the first original power consumption curve in the jth sample pair, represents the second original power consumption curve in the jth sample pair, and M represents the number of sample pairs.

[0054] Step 3: Injecting noise into the original power consumption curves in the sample pairs to obtain sample pairs with injected noise;

[0055] Step 4: Synthesizing the sample pairs to obtain M synthesized power consumption curves

[0056] ​​Step 5: distribution optimization is performed on the synthesized power consumption curve, and the distribution-optimized synthesized power consumption curve P new (t) is merged to form an enhanced power consumption curve sample set. raw (t) is merged to form an enhanced power consumption curve sample set.

[0057] In step 1, the collected power consumption curves are preprocessed, including normalization, standardization and clipping. Normalization and standardization processing can effectively eliminate noise and system bias caused by external interference; since the length of the power consumption curve collected by the target encryption device is long, and only the part containing the leakage point is retained during side channel analysis, the irrelevant part is discarded, so that the computer deep learning can focus on the information leakage in the encryption process. It should be noted that in order to facilitate subsequent operations, the length of the power consumption curve after clipping is the same.

[0058] In step 2, the two original power consumption curves in each sample pair belong to the same category and correspond to the same key operation stage, which can ensure that the features of each sample pair remain consistent and the synthesized data can cover similar signal features. In step 2, random sampling is performed based on the category weight to form a sample pair, and the sampling probability is adjusted according to the category distribution imbalance degree γ = max c∈C (N c ) / min c∈C (N c ), where C represents a set of categories, c represents a specific category in C, N c represents the number of samples belonging to category c, max c∈C (N c ) represents the number of samples in the category with the most samples, and min c∈C (N c ) represents the number of samples in the category with the least samples.

[0059] After random sampling in step 2, M sample pairs are formed, and M power consumption curves can be synthesized. In theory, the value of M can be as high as N 2 . If the value of N is 500, the value of M can theoretically reach 250000; considering that the power consumption curves in each sample pair need to belong to the same category and correspond to the same key operation stage, and the power consumption curves in the original power consumption curve sample set belong to different categories and correspond to different key operation stages, the value of M is difficult to reach the highest value N 2 , but this will not affect the geometric progression expansion of the power consumption curve sample set data amount, only the expansion range is smaller, but it is still better than the linear growth of the sample set data amount.

[0060] It should be noted that each of the original power consumption curves P raw In the present embodiment, it is preferable that the label take a value in the range {0, 1, 2, …, 255}, and in other embodiments, the value of the label can be adjusted as needed. The energy leaked by each original power consumption curve is calculated, and it is classified into different categories c according to its energy leaked value, and in the present embodiment, it is preferable that the value of the category c take a value in the range {1, 2, …, 9}, and in other embodiments, the value of the category c can be adjusted as needed. It should be understood that in each category c, there are cases where the labels of the power consumption curves are different.

[0061] In step 3, noise is injected into the original power consumption curves in the sample pair respectively, to obtain the sample pair after injecting noise wherein represents the first injected-noise power consumption curve in the jth sample pair after injecting noise; represents the second injected-noise power consumption curve in the jth sample pair after injecting noise; The noise injected into the first injected-noise power consumption curve, the noise injected into the second injected-noise power consumption curve, and are subject to a Gaussian distribution with a mean of 0 and a variance of σ 2 The noise and satisfy the condition:

[0062]

[0063] wherein σ min is the minimum value of σ, and SNR target is the target signal-to-noise ratio. It should be noted that the value of σ min may be the theoretical minimum value, or on the basis of the theoretical minimum value, the influence of the noise injected into the sample on the model performance and data usability when training the model under different target signal-to-noise ratios can be evaluated to determine a value that is representative and has generalization ability.

[0064] It should be noted that the lengths of the noise and should be equal to and The lengths of the two original power consumption curves are the same; for a certain sample pair, the noise injected into the two original power consumption curves in the sample pair can be the same or different; for different sample pairs, the injected noise can be the same or different. It should be further explained that in order to better simulate the noise interference in the real environment, the injected noise is preferably different while meeting the above conditions, so as to better provide training data in different noise environments for the computer deep learning model and enhance the noise robustness of the model. The injected noise meeting the above conditions can ensure that the injected noise energy does not exceed the specified proportion of the signal energy of the original power consumption curve, thereby ensuring that the generated sample data well retains the original leakage information characteristics on the basis of effectively simulating the noise characteristics in different environments.

[0065] In step 4, the two power consumption curves after injecting noise in each sample pair are synthesized to form a synthesized power consumption curve in, represents the power consumption curve formed by synthesizing the two power consumption curves after injecting noise in the j-th sample pair. α represents the synthesis weight, and its value range is (0, 1). It should be noted that the specific value of α is adjusted according to needs. When its value is 0.5, it represents average synthesis. If you want the synthesized power consumption curve to be more recognizable or retain more features of dominant samples, you can set the value of α to favor more recognizable samples or dominant samples.

[0066] Synthesized power consumption curve The label is in, The synthetic power consumption curve The label of the first power consumption curve in the sample pair of Synthetic power consumption curve The label of the second power consumption curve in the sample pair, the weight coefficient λ (j) It is a dynamic mixing factor calculated based on the leakage information. Under the relevant power consumption analysis, the leakage model of the key bytes is used to estimate the contribution of the sample and calculate the weight coefficient λ (j) Specifically, according to the formula:

[0067]

[0068] Calculate the weight coefficient λ (j) ,in and Represent the original power consumption curves respectively and corresponding to the intermediate values v1 and v2 under the leakage model. It should be noted that the leakage model can adopt an existing model, such as a Hamming weight model (HW), a Hamming distance model (HD), etc., and the intermediate value refers to an intermediate sensitive variable in a key operation process, such as a result after SBox operation in AES encryption.

[0069] In step 5, the synthesized power consumption curve P new The objective function for distribution optimization of (t) is:

[0070]

[0071] where C represents a set of categories, c represents a specific category in C, p c and q c represent the conditional distribution of the synthesized power consumption curve and the original power consumption curve on the category c, respectively, D KL (p c || q c ) represents the KL divergence of p c and q c , ω is a balance hyperparameter, represents an expected operation, τ c is a class center vector of all samples of the category c in the original power consumption curve, i.e., a vector composed of the mean values of the power consumption values of the samples of the category c at each time point, used to represent the leakage pattern of the category c. It should be noted that the KL divergence represents the degree of distribution deviation, and ω is used to control the trade-off between the KL divergence term and the class center vector term, which is set to a constant according to needs, represents the expectation of the synthesized power consumption curve under the condition that the category is c, and is preferably the intra-class average.

[0072] The loss function for distribution optimization of the synthesized power consumption curve P new (t) is:

[0073]

[0074] where θ is a parameter space of the synthesis process, which is composed of the value range of the to-be-optimized parameters (such as synthesis ratio, noise disturbance factor, sampling mechanism, etc.) that control the generation process of the synthesized sample, M c is the number of categories c in the synthesized power consumption curve.

[0075] The optimization trajectory for distribution optimization of the synthesized power consumption curve P new (t) is:

[0076]

[0077] where ρ is a strength coefficient of the balance constraint, and λ is a weight coefficient of the label.

[0078] Because the power consumption characteristics of the hardware encryption device are significantly different at different time points and in different encryption operation stages, the number of samples of each category in the data set of the original power consumption curve can be seriously unbalanced. Figure 2 The principle diagram of balancing the sample categories of the present application is shown. As shown in Figure 2 From the perspective of statistical distribution, the data set of the original power consumption curve presents an approximate binomial distribution, that is, the number of original power consumption curves of some categories (such as S-box output, Hamming weight of key bytes, etc.) is relatively large, while the number of original power consumption curves of other categories is relatively small, which causes the skewed distribution of the data set of the original power consumption curve in categories, which will cause the computer deep learning model to be more inclined to high-frequency categories during training, and weaken the learning ability of low-frequency categories.

[0079] At the same time, because the two power consumption curves in the sample pair are nonlinearly combined in the numerical space to form a synthesized power consumption curve, the label of the synthesized power consumption curve becomes ambiguous, that is, the synthesized power consumption curve can not be accurately attributed to a single category in the original data set. This label inconsistency can interfere with the learning process of the model, affect the convergence of the model, and may cause the category imbalance problem, so that the prediction deviation of some categories is large, thereby reducing the attack success rate.

[0080] Through step 2, in the process of forming the sample pair, the sampling probability is determined according to the imbalance degree of the category distribution, which can make the power consumption curves of the categories with less number of samples preferentially form sample pairs, and form a larger number of sample pairs in order to synthesize a larger number of synthesized power consumption curves, as shown in Figure 2 Further, the power consumption curves synthesized by the sample pairs make up for the difference in category distribution of the original power consumption curves, and further make the number of power consumption curves in the merged sample set approximately the same in each category.

[0081] Through step 4, the weight coefficient λ (j) is dynamically adjusted according to the energy leakage distribution of the two power consumption curves in the sample pair, which can ensure that the label of the synthesized sample curve can still correctly reflect the key features in the key operation process, maintain the stability of the category information, and avoid the adverse effects of label conflicts on model training. At the same time, dynamically adjusting the weight coefficient λ (j) can effectively improve the diversity of power consumption curve samples of the same category by using the natural differences in timing characteristics, amplitude response, etc. between the two original power consumption curves in the sample pair, and enhance the generalization ability of the computer deep learning model to boundary samples.

[0082] Through the distribution optimization in step 5, the synthesized power consumption curve can be dynamically guided to further tilt toward the low-frequency category according to the category distribution deviation, and the quantitative difference of the power consumption curves of each category can be continuously reduced. In this way, the distribution-optimized, synthesized power consumption curve P new (t) and the original power consumption curve P raw (t) Merging to form an enhanced power consumption curve sample set can achieve approximately balanced distribution across categories, improving the generalization ability of computer deep learning models.

[0083] Figure 3 The diagram shows the result of balancing sample categories. Figure 3 As shown, the gray histogram represents the distribution of the original power consumption curve across categories, while the blue histogram represents the distribution of the power consumption curve synthesized from the sample pairs across categories. This distribution differs significantly from the distribution of the original power consumption curve. After merging the original and synthesized power consumption curve datasets, the enhanced power consumption curve sample set is distributed across categories, as shown in the orange histogram. It can be seen that class balance has been significantly improved.

[0084] Example 2

[0085] A data enhancement system for side channel analysis includes a processor having a program stored thereon, and the processor is configured to run the program to perform the data enhancement method for side channel analysis.

[0086] Preferably, the data enhancement system for side channel analysis comprises:

[0087] The data collection and preprocessing module is configured to collect the power consumption curve of the target encryption device and preprocess it to obtain N preprocessed original power consumption curves containing leakage points.

[0088] The sample pair selection module is configured to randomly sample N pre-processed original power consumption curves containing leakage points to form sample pairs. where j = 1, 2, …, M; t∈[0, T]; represents the first original power consumption curve in the j-th sample pair, represents the second original power consumption curve in the jth sample pair, and M represents the number of sample pairs;

[0089] a noise injection module configured to inject noise into an original power consumption curve in a sample pair to obtain a noise-injected sample pair;

[0090] A sample pair synthesis module is configured to synthesize the sample pairs to obtain M synthesized power consumption curves

[0091] a distribution optimization and merging module configured to perform distribution optimization on the synthesized power consumption curve P new (t) and merge the distribution-optimized synthesized power consumption curve P new (t) with the original power consumption curve P raw (t) to form an enhanced power consumption curve sample set.

[0092] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit the present application; although the foregoing embodiments have been described in detail, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some or all of the technical features can be replaced by equivalents, and these replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the present application.

Claims

1. A data enhancement method for side channel analysis, characterized by: include: Step 1: Collect the power consumption curve of the target encryption device and preprocess it to obtain N preprocessed original power consumption curves containing leakage points. Step 2: Randomly sample N pre-processed original power consumption curves containing leakage points to form sample pairs. where j = 1, 2, …, M; t∈[0, T]; represents the first original power consumption curve in the j-th sample pair, represents the second original power consumption curve in the jth sample pair, and M represents the number of sample pairs; Step 3: Inject noise into the original power consumption curve in the sample pair to obtain a noise-injected sample pair; Step 4: Synthesize the sample pairs to obtain M synthetic power consumption curves Step 5: Optimize the distribution of the synthesized power consumption curve and convert the optimized distribution power consumption curve P new (t) and the original power consumption curve P raw (t) Merge to form an enhanced power consumption curve sample set.

2. The data enhancement method for side channel analysis according to claim 1, wherein: In step 2, the two original power consumption curves in each sample pair belong to the same category and correspond to the same key operation phase; Random sampling is performed based on the category weights to form sample pairs, and the sampling probability is determined by the category distribution imbalance γ = max c∈C (N c ) / min c∈C (N c ) is adjusted, where C represents a set of categories, c represents a specific category in C, and N c Indicates the number of samples belonging to category c, max c∈C (N c ) represents the number of samples of the category with the largest number of samples among all categories, min c∈C (N c ) represents the number of samples of the category with the least number of samples among all categories.

3. The data enhancement method for side channel analysis according to claim 2, wherein: In step 3, noise is injected into the original power consumption curves of the sample pairs to obtain the sample pairs after noise injection. in represents the power consumption curve of the first noise injection in the jth sample pair after noise injection; represents the power consumption curve of the second noise injection in the jth sample pair after noise injection; For The injected noise, For The injected noise, and All have mean 0 and variance σ 2 Gaussian distribution; the noise and Requirements: Among them, σ min is the minimum value of σ, SNR target is the target signal-to-noise ratio.

4. The data enhancement method for side channel analysis according to claim 3, wherein: In step 4, the two power consumption curves after injecting noise in each sample pair are synthesized to form a synthesized power consumption curve in, It represents the power consumption curve formed by synthesizing the two power consumption curves after injecting noise in the j-th sample pair. α represents the synthesis weight, and its value range is (0,1).

5. The data enhancement method for side channel analysis according to claim 4, wherein: In step 4, the synthesized power consumption curve The label is in, The synthetic power consumption curve The label of the first power consumption curve in the sample pair of Synthetic power consumption curve The label of the second power consumption curve in the sample pair, the weight coefficient λ (j) It is a dynamic mixing factor calculated based on the leaked information.

6. The data enhancement method for side channel analysis according to claim 5, wherein: In step 4, under the relevant power consumption analysis, the leakage model of the key bytes is used to estimate the contribution of the sample and calculate the weight coefficient λ (j) ;λ (j) The calculation formula is: in and Represent the original power consumption curves respectively and The intermediate values ​​v1 and v2 correspond to the theoretical power consumption values ​​under the leakage model.

7. The data enhancement method for side channel analysis according to claim 6, wherein: In step 5, the synthesized power consumption curve P new (t) The objective function for distribution optimization is: Among them, C represents a set of categories, c represents a specific category in C, and p c and q c Denote the conditional distribution of the synthesized power consumption curve and the original power consumption curve on category c, respectively. KL (p c ||q c ) indicates p c and q c KL divergence of , ω is the balance hyperparameter, represents the expected operation, τ c is the class center vector of all samples of class c in the original power consumption curve.

8. The data enhancement method for side channel analysis according to claim 6, wherein: In step 5, the synthesized power consumption curve P new (t) The loss function for distribution optimization is: Among them, θ is the parameter space of the synthesis process, M c is the number of categories c in the synthesized power consumption curve.

9. The data enhancement method for side channel analysis according to claim 6, wherein: In step 5, the synthesized power consumption curve P new (t) The optimization trajectory for distribution optimization is: Among them, ρ is the strength coefficient of the equilibrium constraint, and λ is the weight coefficient of the label.

10. A data enhancement system for side channel analysis, comprising a processor having a program stored thereon, characterized in that: The processor is configured to run the program to perform the data enhancement method for side channel analysis according to any one of claims 1 to 9.

Images