Fully automatic high-quality input differential screening method and system for differential neural network analysis

Abstract

This invention relates to the interdisciplinary field of cryptography and deep learning, and particularly to a fully automated method and system for selecting high-quality input differentials in differential neural network analysis. The method first determines input parameters such as the target object of the block cipher, the number of samples in the dataset, and the ciphertext pair dataset generation function. Then, it iterates through all single-bit input differences, generating corresponding positive sample ciphertext pair datasets, and performs feature extraction and quantization analysis to obtain the quality evaluation value of each single-bit input difference. Next, it selects high-quality single-bit input differences to form a candidate set, and generates multi-bit input differences through XOR combination. Finally, it quantifies and evaluates the multi-bit input differences, integrates and sorts the quality evaluation values ​​of all differences, and selects a preset number of optimal input differences. This invention requires no manual parameter tuning, does not rely on specific cryptographic algorithm structures, has strong adaptability, and can efficiently select high-quality input differences and optimize the training dataset for neural network discriminators.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of cryptography and deep learning, and particularly to a fully automated, high-quality input differential filtering method and system for differential neural network analysis. The core application is the optimization and filtering of input differentials in the training dataset of a neural network discriminator. It can be widely used in scenarios such as IoT devices and embedded systems that deploy lightweight block ciphers (such as SPECK-32 / 64, SIMON-32 / 64, SIMECK32 / 64), providing an efficient tool for security evaluation of cryptographic algorithms. Background Technology

[0002] With the rapid development of deep learning in the field of cryptanalysis, differential neural network analysis (DNN) is a cryptanalysis technique that combines deep learning with traditional differential cryptanalysis. Its core is to train a neural network discriminator to identify the non-random propagation characteristics of the plaintext pair's input differential after encryption, thereby achieving key recovery or algorithm security assessment. DNN has become a highly competitive cryptanalysis technique. Its key advantage lies in the fact that the neural network discriminator can automatically learn the differential propagation characteristics of the encryption algorithm, eliminating the need for cryptanalysts to manually derive complex differential paths, significantly lowering the technical application threshold.

[0003] The performance of a neural network discriminator directly depends on the quality of the training dataset, and the selection of input differences (the XOR difference between plaintext pairs, a core element in constructing the difference dataset, whose quality directly determines the performance of the neural network discriminator) is a crucial step in dataset optimization. High-quality input differences allow the differential features of plaintext pairs to be fully exposed after encryption, helping the model quickly learn effective discrimination rules. Currently, the mainstream input difference selection methods mainly include the following categories:

[0004] 1. Traditional differential cryptanalysis methods: rely on the expertise of cryptanalysts to manually mine high-probability differential features, which has an extremely high knowledge threshold and is inefficient;

[0005] 2. Small-scale training method: Input difference candidates are generated by random bit flipping and rely on small-scale neural network evaluation, which has the problems of insufficient search space traversal and insufficient reliability of results;

[0006] 3. Automated constraint model method: Searching for differential paths by constructing mathematical constraint models requires customizing models for different cryptographic algorithms, which is time-consuming, labor-intensive, and highly dependent on knowledge.

[0007] 4. Data Analysis Method (GIDE): A data analysis-driven method based on PCA and clustering, which lowers the knowledge threshold, but still has significant shortcomings and cannot meet the needs of practical applications.

[0008] Existing input difference filtering methods suffer from the following key drawbacks in practical applications of differential neural network analysis, resulting in their inability to meet requirements in terms of efficiency, automation, and reliability:

[0009] 1. Significant efficiency bottleneck: The core modules of the GIDE algorithm have extremely uneven time consumption ratios, with the clustering and evaluation modules becoming performance bottlenecks; moreover, it only supports differential search for inputs with a single Hamming weight, and the entire process needs to be run repeatedly in scenarios with multiple Hamming weights, resulting in an exponential increase in time costs;

[0010] 2. Insufficient automation: Key parameters of GIDE need to be determined through manual pre-experiment debugging, which relies on the experimental experience of operators and cannot achieve full automation.

[0011] 3. The reliability of the results depends on parameter tuning: GIDE's high reliability is only achieved under specific pre-tuned parameters. Improper parameter settings will lead to deviations in the screening results and reduce the training effect of the neural network discriminator.

[0012] 4. Cumbersome search for multiple Hamming weight differences: The optimal input difference may be distributed in different Hamming weight ranges, but GIDE needs to run the algorithm separately for each Hamming weight, which is complicated and time-consuming. Summary of the Invention

[0013] This invention addresses the problems of low efficiency, insufficient automation, cumbersome operation in multi-Hamm weight scenarios, and parameter-dependent reliability in existing input differential filtering methods for differential neural network analysis. It proposes a fully automated, high-quality input differential filtering method and system for differential neural network analysis, achieving low knowledge dependence, high efficiency, high reliability, and full automation of input differential filtering, thereby optimizing the dataset quality for differential neural network analysis.

[0014] To achieve the above objectives, the technical solution adopted is:

[0015] This invention provides a fully automated high-quality input differential filtering method for differential neural network analysis, comprising the following steps:

[0016] Determine the target object, input parameters, and dataset generation function: The target object is a block cipher; the input parameters include the number of dataset samples n, the number of input differences to be filtered m, and the number of single-bit input difference candidates k used to generate multi-bit input differences; the dataset generation function GenDataset is used to generate the corresponding ciphertext pair dataset based on the given input differences;

[0017] Single-bit difference evaluation: Traverse all single-bit input differences, call the dataset generation function GenDataset to generate the corresponding positive sample ciphertext pair dataset, perform feature extraction and quantization analysis on the positive sample ciphertext pair dataset, and obtain the quality evaluation value of each single-bit input difference;

[0018] Multi-bit differential generation: Based on the quality assessment value of single-bit input differentials, high-quality single-bit input differentials are selected to form a candidate set, and multi-bit input differentials are generated by XOR combination.

[0019] Unified evaluation and optimal difference selection: The generated multi-bit input differences are quantitatively evaluated to obtain their quality evaluation values. The quality evaluation values ​​of all single-bit and multi-bit input differences are integrated and sorted according to the values. Based on the sorting, a preset number of optimal input differences are selected.

[0020] According to the fully automatic high-quality input differential filtering method of differential neural network analysis of the present invention, the block size of the block cipher is bs, the key size is ks; the total number of single-bit input differentials is bs, and in each single-bit input differential there is only 1 bit that is 1, and all other (bs-1) bits are 0.

[0021] According to the fully automatic high-quality input differential filtering method of differential neural network analysis of the present invention, the process of generating the ciphertext pair dataset by the dataset generation function GenDataset is as follows: a large number of plaintext pairs (P1, P2) are generated using the input difference Δ, satisfying P1⊕P1=Δ; these plaintext pairs are encrypted through r rounds to obtain the corresponding ciphertext pairs (C1, C2), and the set of all (C1, C2) constitutes the ciphertext pair dataset.

[0022] According to the fully automated high-quality input differential filtering method of differential neural network analysis of the present invention, the single-bit differential evaluation specifically includes:

[0023] For each single-bit input differential Δ single Call the dataset generation function GenDataset to generate each Δ single The corresponding dataset D contains n positive sample ciphertext pairs. single ;

[0024] Calculate dataset D single The covariance matrix A;

[0025] Perform PCA eigenvalue decomposition on the covariance matrix A to obtain the eigenvalue set. Where q = 2 × bs, q corresponds to the dimension of the ciphertext pair, and bs represents the block size of the block cipher;

[0026] Calculate the variance of the eigenvalue set Λ ;

[0027] Single-bit input differential Δ single Its corresponding eigenvalue variance Associated storage to the evaluation set Δset .

[0028] According to the fully automated high-quality input differential filtering method of differential neural network analysis of the present invention, the variance of the feature set Λ is further... The calculation formula is as follows:

[0029]

[0030] Where, μ Λ The variance is the mean of the feature set Λ. The larger the variance, the stronger the non-randomness of the dataset, and the higher the quality of the corresponding input difference.

[0031] According to the fully automated high-quality input differential filtering method of differential neural network analysis of the present invention, the multi-bit differential generation specifically includes:

[0032] The evaluation set Δ set Single-bit input differential in eigenvalue variance Arrange the data in descending order and select the top k high-quality single-bit input differences to form the candidate set Δ. top-single ;

[0033] For candidate set Δ top-single The elements in the array are XORed together to generate 2-bit and 3-bit differences;

[0034] The generated 2-bit and 3-bit differences are stored in a multi-bit difference set Δ. multi .

[0035] According to the fully automatic high-quality input differential filtering method of differential neural network analysis of the present invention, the calculation formula for generating the 2-bit difference is further as follows: The formula for generating the 3-bit difference is as follows: .

[0036] According to the fully automated high-quality input differential filtering method of differential neural network analysis of the present invention, the unified evaluation and optimal differential selection specifically include:

[0037] Traversing the multi-bit difference set Δ multi Each multi-bit input differential Δ m Repeat the single-bit differential evaluation step to calculate Δ. m eigenvalue variance and input multi-bit differential Δ m Its eigenvalue variance Store to evaluation set Δ set ;

[0038] The evaluation set Δ setAll input differences are sorted in descending order of their eigenvalue variances, and the top m differences are selected as the optimal input difference output.

[0039] Furthermore, the present invention also provides a fully automated high-quality input differential filtering system for differential neural network analysis, comprising:

[0040] The parameter and function definition module is used when the target object is a block cipher; the input parameters include the number of dataset samples n, the number of input differences to be filtered m, and the number of single-bit input difference candidates k used to generate multi-bit input differences; the dataset generation function GenDataset is used to generate the corresponding ciphertext pair dataset based on the given input differences;

[0041] The single-bit difference evaluation module is used to traverse all single-bit input differences, call the dataset generation function GenDataset to generate the corresponding positive sample ciphertext pair dataset, perform feature extraction and quantization analysis on the positive sample ciphertext pair dataset, and obtain the quality evaluation value of each single-bit input difference.

[0042] The multi-bit difference generation module is used to select high-quality single-bit input differences based on the quality evaluation value of single-bit input differences to form a candidate set, and generate multi-bit input differences through XOR combination.

[0043] The optimal difference selection module is used to quantize and evaluate the generated multi-bit input differences to obtain their quality evaluation values, integrate the quality evaluation values ​​of all single-bit and multi-bit input differences, sort them according to the values, and select a preset number of optimal input differences based on the sorting.

[0044] The beneficial effects achieved by adopting the above technical solution are:

[0045] 1. Lightweight quality quantification index: The PCA feature value ratio variance is directly used as the quality quantification index of the difference dataset. The non-randomness of the feature value distribution reflects the quality of the input difference, fundamentally improving the efficiency bottleneck.

[0046] 2. Differential Combination Generation Strategy: It was discovered that high-quality multi-bit input differentials can be generated by XOR combination of high-quality single-bit input differentials, eliminating the need to repeat the entire process for each Hamming weight, thus significantly improving the screening efficiency for multi-Hamming weight scenarios.

[0047] 3. Fully automated process design: No manual debugging of any core parameters is required. Only the basic parameters of the target cryptographic algorithm and the dataset configuration need to be input to automatically complete the entire process of single-bit evaluation, multi-bit generation and unified screening, completely eliminating manual intervention.

[0048] 4. Cross-algorithm adaptability: It does not depend on the specific structure of the target cryptographic algorithm, but is evaluated only through the characteristics of the dataset. It can be directly adapted to various lightweight block ciphers and has strong versatility. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0050] Figure 1 This is a flowchart illustrating the fully automated high-quality input differential filtering method for differential neural network analysis according to an embodiment of the present invention. Detailed Implementation

[0051] The exemplary solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art.

[0052] This invention, based on PCA eigenvalue quantization analysis and a difference combination strategy, designs a fully automated high-quality input difference filtering method for difference neural network analysis, such as... Figure 1 As shown, it includes the following steps:

[0053] Step S1: Determine the target object, input parameters, and dataset generation function.

[0054] The target is a lightweight block cipher E with a block size of bs and a key size of ks.

[0055] Input parameters include: the number of samples in the dataset n (e.g., 10). 5 The number of input differences to be filtered, m, and the number of single-bit input difference candidates to be used to generate multi-bit input differences, k (default k=5).

[0056] The dataset generation function `GenDataset(Δ,n,bs)` generates a corresponding ciphertext pair dataset based on a given input difference `Δ`. Specifically, `GenDataset` generates a massive number of plaintext pairs (P1, P2) satisfying `P1⊕P1=Δ` using the input difference `Δ`. These plaintext pairs are then encrypted using the target algorithm for `r` rounds to obtain the corresponding ciphertext pairs (C1, C2). The set of all (C1, C2) constitutes the ciphertext pair dataset, serving as the positive sample dataset. The negative sample dataset can be obtained by encrypting it using a random permutation function. The positive and negative sample datasets are concatenated to train a neural network discriminator, enabling the neural network discriminator to distinguish between non-random ciphertext pair distributions and random ciphertext pair distributions caused by a specific input difference.

[0057] Step S2, Single-bit difference evaluation: Traverse all single-bit input differences, call the dataset generation function GenDataset to generate the corresponding positive sample ciphertext pair dataset, perform feature extraction and quantization analysis on the positive sample ciphertext pair dataset, and obtain the quality evaluation value of each single-bit input difference.

[0058] Iterate through all single-bit input differences, and for each single-bit difference Δ single Perform the following operations, where the total number of single-bit input differences is bs, and in each single-bit input difference there is only 1 bit that is 1, and all the other (bs-1) bits are 0. For example, Δ0 has only the 0th bit that is 1, and the rest are 0. The binary representation is 0000 0001.

[0059] Step S201, Positive Sample Dataset Generation: For each single-bit input difference Δ single Call the dataset generation function GenDataset to generate each Δ single The corresponding dataset D contains n positive sample ciphertext pairs. single .

[0060] Step S202, Feature Extraction and Quantitative Analysis:

[0061] a. Calculate dataset D single The covariance matrix A;

[0062] b. Perform PCA eigenvalue decomposition on the covariance matrix A to obtain the eigenvalue set. Where q = 2 × bs, q corresponds to the dimension of the ciphertext pair, and bs represents the block size of the block cipher;

[0063] c. Calculate the mean of the eigenvalue set Λ: ;

[0064] d. Calculate the variance of the eigenvalue set Λ: The larger the variance, the stronger the non-randomness of the dataset, and the higher the quality of the corresponding input difference.

[0065] Step S203, Result Storage: Store the single-bit input differential Δ single Its corresponding eigenvalue variance Associated storage to the evaluation set Δ set .

[0066] Step S3, Multi-bit Differential Generation: Based on the quality assessment value of the single-bit input differential, high-quality single-bit input differentials are selected to form a candidate set, and high-potential multi-bit input differentials are generated through XOR combination. This step specifically includes the following sub-steps:

[0067] Step S301, Single-bit input differential filtering: The evaluation set Δset Single-bit input differential in eigenvalue variance Sort the data in descending order and select the top k (default k=5) high-quality single-bit input differences to form the candidate set Δ. top-single .

[0068] Step S302: For the candidate set Δ top-single The elements in the array are XORed together to generate 2-bit and 3-bit differences:

[0069] a. 2-bit differential generation: for Δ top-single The single-bit input difference is XORed and combined to obtain a 2-bit difference. ;

[0070] b. 3-bit differential generation: for Δ top-single The single-bit input difference is XORed and combined to obtain a 3-bit difference. .

[0071] Step S303, Multi-bit set construction: Store the generated 2-bit and 3-bit differences into a unified multi-bit difference set Δ multi .

[0072] Step S4, Unified Evaluation and Optimal Difference Selection: Quantitatively evaluate the generated multi-bit input differences to obtain their quality evaluation values. Integrate the quality evaluation values ​​of all single-bit and multi-bit input differences and sort them according to these values. Based on the sorting, select a preset number of optimal input differences. This step specifically includes the following sub-steps:

[0073] Step S401, Multi-bit Input Differential Evaluation: Traverse the multi-bit differential set Δ multi Each multi-bit input differential Δ m The dataset generation function GenDataset is called to generate its corresponding difference dataset D. m Repeat the steps of covariance matrix calculation, eigenvalue decomposition, and variance calculation in step S202 to obtain Δ. m eigenvalue variance , will (Δ m , ) Associated storage to Δ set This enables the integration of differential evaluation results for single-bit and multi-bit inputs.

[0074] Step S402, Global Sorting and Optimal Selection: Δ set All input differences are expressed according to their eigenvalue variance σ 2 Sort the samples in descending order and select the top m differences as the optimal input differences.

[0075] Corresponding to the above method, embodiments of the present invention also disclose a fully automated high-quality input differential filtering system for differential neural network analysis, comprising:

[0076] The parameter and function definition module is used when the target object is a block cipher; the input parameters include the number of dataset samples n, the number of input differences to be filtered m, and the number of single-bit input difference candidates k used to generate multi-bit input differences; the dataset generation function GenDataset is used to generate the corresponding ciphertext pair dataset based on the given input differences;

[0077] The single-bit difference evaluation module is used to traverse all single-bit input differences, call the dataset generation function GenDataset to generate the corresponding positive sample ciphertext pair dataset, perform feature extraction and quantization analysis on the positive sample ciphertext pair dataset, and obtain the quality evaluation value of each single-bit input difference.

[0078] The multi-bit difference generation module is used to select high-quality single-bit input differences based on the quality evaluation value of single-bit input differences to form a candidate set, and generate multi-bit input differences through XOR combination.

[0079] The optimal difference selection module is used to quantize and evaluate the generated multi-bit input differences to obtain their quality evaluation values, integrate the quality evaluation values ​​of all single-bit and multi-bit input differences, sort them according to the values, and select a preset number of optimal input differences based on the sorting.

[0080] The effectiveness and superiority of the present invention will be verified below based on experimental data.

[0081] Experimental platform: Intel(R) Core(TM) i9-14900K CPU @ 3.20GHz + NVIDIA GeForce RTX 4080 SUPER, training set size 10 7 Validation set 10 6 .

[0082] Time efficiency and result reliability experiments were conducted on 5-wheel SPECK-32 / 64, 8-wheel SIMON-32 / 64 and 8-wheel SIMECK-32 / 64 respectively, and the results are shown in Table 1 and Table 2.

[0083]

[0084]

[0085] The following conclusions can be drawn from Tables 1 and 2:

[0086] 1. Significantly improved efficiency: In single-input differential filtering scenarios, the runtime is only 5% of GIDE's; in multi-input differential filtering scenarios, the runtime is only 0.1% of GIDE's, and in multi-Hamm weight scenarios, there is no need to run repeatedly. The efficiency advantage becomes more obvious as the task scale increases.

[0087] 2. High reliability: Experimental verification on three cryptographic algorithms, SPECK-32 / 64, SIMON-32 / 64 and SIMECK32 / 64, shows that the selected optimal input difference is optimal and significantly better than existing methods.

[0088] In addition to the above conclusions, the present invention also has the following advantages:

[0089] 3. Low knowledge dependency: It does not require cryptanalysts to master complex differential cryptanalysis theories or algorithm customization modeling capabilities. It can be started by simply inputting basic parameters, which lowers the technical application threshold.

[0090] 4. Fully automated: From dataset generation and feature quantization to result selection, there is no human intervention throughout the process, which avoids result deviations caused by parameter tuning and improves the stability and ease of use of the method.

[0091] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A fully automated high-quality input differential filtering method for differential neural network analysis, characterized in that, Includes the following steps: Determine the target object, input parameters, and dataset generation function: The target object is a block cipher; the input parameters include the number of dataset samples n, the number of input differences to be filtered m, and the number of single-bit input difference candidates k used to generate multi-bit input differences; The GenDataset function is used to generate a corresponding ciphertext pair dataset based on a given input difference. Single-bit difference evaluation: Traverse all single-bit input differences, call the dataset generation function GenDataset to generate the corresponding positive sample ciphertext pair dataset, perform feature extraction and quantization analysis on the positive sample ciphertext pair dataset, and obtain the quality evaluation value of each single-bit input difference; Multi-bit differential generation: Based on the quality assessment value of single-bit input differentials, high-quality single-bit input differentials are selected to form a candidate set, and multi-bit input differentials are generated by XOR combination. Unified evaluation and optimal difference selection: The generated multi-bit input differences are quantitatively evaluated to obtain their quality evaluation values. The quality evaluation values ​​of all single-bit and multi-bit input differences are integrated and sorted according to the values. Based on the sorting, a preset number of optimal input differences are selected.

2. The fully automated high-quality input differential filtering method for differential neural network analysis according to claim 1, characterized in that, The block cipher has a block size of bs and a key size of ks; the total number of single-bit input differences is bs, and each single-bit input difference has only one bit that is 1, while all other (bs-1) bits are 0.

3. The fully automated high-quality input differential filtering method for differential neural network analysis according to claim 1, characterized in that, The GenDataset function generates a ciphertext pair dataset as follows: it generates a massive number of plaintext pairs (P1, P2) using the input difference Δ, satisfying P1⊕P1=Δ; it encrypts these plaintext pairs through r rounds to obtain the corresponding ciphertext pairs (C1, C2), and the set of all (C1, C2) constitutes the ciphertext pair dataset.

4. The fully automated high-quality input differential filtering method for differential neural network analysis according to claim 1, characterized in that, The single-bit differential evaluation specifically includes: For each single-bit input differential Δ single Call the dataset generation function GenDataset to generate each Δ single The corresponding dataset D contains n positive sample ciphertext pairs. single ; Calculate dataset D single The covariance matrix A; Perform PCA eigenvalue decomposition on the covariance matrix A to obtain the eigenvalue set. Where q = 2 × bs, q corresponds to the dimension of the ciphertext pair, and bs represents the block size of the block cipher; Calculate the variance of the eigenvalue set Λ ; Single-bit input differential Δ single Its corresponding eigenvalue variance Associated storage to the evaluation set Δ set .

5. The fully automated high-quality input differential filtering method for differential neural network analysis according to claim 4, characterized in that, The variance of the eigenvalue set Λ The calculation formula is as follows: Where, μ Λ The variance is the mean of the feature set Λ. The larger the variance, the stronger the non-randomness of the dataset, and the higher the quality of the corresponding input difference.

6. The fully automated high-quality input differential filtering method for differential neural network analysis according to claim 4, characterized in that, The multi-bit differential generation specifically includes: The evaluation set Δ set Single-bit input differential in eigenvalue variance Arrange the data in descending order and select the top k high-quality single-bit input differences to form the candidate set Δ. top-single ; For candidate set Δ top-single The elements in the array are XORed together to generate 2-bit and 3-bit differences; The generated 2-bit and 3-bit differences are stored in a multi-bit difference set Δ. multi .

7. The fully automated high-quality input differential filtering method for differential neural network analysis according to claim 6, characterized in that, The formula for generating the 2-bit difference is as follows: The formula for generating the 3-bit difference is as follows: .

8. The fully automated high-quality input differential filtering method for differential neural network analysis according to claim 6, characterized in that, The unified evaluation and optimal difference selection specifically include: Traversing the multi-bit difference set Δ multi Each multi-bit input differential Δ m Repeat the single-bit differential evaluation step to calculate Δ. m eigenvalue variance and input multi-bit differential Δ m Its eigenvalue variance Store to evaluation set Δ set ; The evaluation set Δ set All input differences are sorted in descending order of their eigenvalue variances, and the top m differences are selected as the optimal input difference output.

9. A fully automated high-quality input differential filtering system for differential neural network analysis, characterized in that, include: The parameter and function definition module is used when the target object is a block cipher; the input parameters include the number of dataset samples n, the number of input differences to be filtered m, and the number of single-bit input difference candidates k used to generate multi-bit input differences; The GenDataset function is used to generate a corresponding ciphertext pair dataset based on a given input difference. The single-bit difference evaluation module is used to traverse all single-bit input differences, call the dataset generation function GenDataset to generate the corresponding positive sample ciphertext pair dataset, perform feature extraction and quantization analysis on the positive sample ciphertext pair dataset, and obtain the quality evaluation value of each single-bit input difference. The multi-bit difference generation module is used to select high-quality single-bit input differences based on the quality evaluation value of single-bit input differences to form a candidate set, and generate multi-bit input differences through XOR combination. The optimal difference selection module is used to quantize and evaluate the generated multi-bit input differences to obtain their quality evaluation values, integrate the quality evaluation values ​​of all single-bit and multi-bit input differences, sort them according to the values, and select a preset number of optimal input differences based on the sorting.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images