Adaptive optimization method for ai-driven post-quantum cryptographic algorithm

Abstract

The application relates to the technical field of data encryption and discloses an AI driving-based adaptive optimization method of a post-quantum cryptographic algorithm. The method collects multi-source cryptographic parameter data, generates a standardized cryptographic parameter set through adaptive noise injection and parameter normalization preprocessing; a dynamic optimization network model based on multi-head attention and reinforcement learning is constructed, an adversarial evolution strategy is adopted for training, a quantum attack scene is simulated to screen an anti-quantum parameter combination; a differential parameter space mapping algorithm is used to convert into an optimal encryption protocol configuration and correct conflicts; through iterative optimization of a meta-learning framework, local parameter sensitivity and global anti-attack capability are fused, and an adaptive post-quantum cryptographic algorithm is output. The application effectively improves the performance and security of the post-quantum cryptographic algorithm, can optimize the algorithm according to different scene requirements, and provides a reliable technical scheme for coping with quantum computing threats.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data encryption, in particular to an AI-driven post-quantum cryptographic algorithm adaptive optimization method. BACKGROUND

[0002] With the rapid development of quantum computing technology, the traditional cryptographic system is facing unprecedented security threats. Quantum computers, with their powerful computing capabilities, can theoretically break the widely used encryption algorithms based on mathematical problems (such as large integer factorization and discrete logarithm problems) in a short time, which has a huge impact on the information security field. In the field of traditional cryptography, classical encryption protocols are designed without fully considering quantum attack factors, and their security relies on the computational complexity of specific mathematical problems. However, the emergence of quantum algorithms (such as Shor's algorithm) makes it possible to solve these mathematical problems on a quantum computer, leading to the possibility of rapid decryption of existing cryptographic system keys and the risk of information being stolen or tampered with.

[0003] In practical application scenarios, such as online transactions in the financial field, patient information management in the medical industry, and confidential document transmission in government departments, encryption technology is highly dependent on encryption technology to protect data security and privacy. For example, in the financial field, a large amount of financial transaction information and customer account data are transmitted and stored over the network, and once the encryption system is broken by quantum attacks, it will cause serious economic losses and social trust crisis. In the medical industry, the leakage of sensitive health information of patients not only infringes on the privacy of patients, but also may cause physical and mental harm to patients.

[0004] Currently, researchers are committed to exploring post-quantum cryptographic algorithms in the hope of finding encryption methods that are still secure and reliable in the quantum computing environment. However, post-quantum cryptographic algorithms face many challenges in practical applications. On the one hand, there are many types of post-quantum cryptographic algorithms, and different algorithms have different performances in security, computational efficiency, resource occupation, etc., making it difficult to directly determine the optimal algorithm suitable for a specific scenario. For example, some post-quantum cryptographic algorithms based on lattice theory have high security, but have high computational complexity and require high hardware performance; while algorithms based on hash functions have relatively high computational efficiency, but their security under certain attack models is controversial. On the other hand, post-quantum cryptographic algorithms need to be adapted to existing communication systems, hardware devices, etc., which involves the modification of classical encryption protocols and the optimization of hardware performance. Since different application scenarios have different requirements for encryption algorithms, such as real-time communication scenarios that require high encryption and decryption speed, and data storage scenarios that focus on long-term security, how to adaptively optimize post-quantum cryptographic algorithms according to specific scenario requirements has become a problem to be solved.

[0005] In addition, existing cryptographic algorithm optimization methods are mostly based on traditional mathematical analysis and empirical tuning, which are difficult to cope with the complex and variable security threats brought by quantum attacks. These methods are inefficient in handling high-dimensional and nonlinear cryptographic parameter spaces, and cannot fully exploit the potential relationships between parameters, resulting in an ideal balance between security and performance being difficult to achieve for the optimized algorithm. In terms of quantum attack simulation, traditional methods also have difficulty in accurately simulating real quantum attack scenarios, which may pose a security risk to the algorithm in actual application. SUMMARY

[0006] The present application aims to provide an AI-driven post-quantum cryptographic algorithm adaptive optimization method to solve the problems raised in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solution: an AI-driven post-quantum cryptographic algorithm adaptive optimization method, the method comprising:

[0008] Step S1: Collecting multi-source cryptographic parameter data, including quantum attack simulation results, classical encryption protocol parameters and hardware performance indicators, and generating a standardized cryptographic parameter set resistant to quantum interference through adaptive noise injection and parameter normalization preprocessing;

[0009] Step S2: Constructing a dynamic optimization network model based on multi-head attention and reinforcement learning, the model comprising a parameter correlation graph embedding module, an anti-quantum attack evaluation module and a multi-objective optimization decision module;

[0010] Step S3: Training the dynamic optimization network model using an adversarial evolution strategy, simulating a quantum attack scenario through a generative adversarial network, and dynamically selecting anti-quantum parameter combinations;

[0011] Step S4: Based on the differentiable parameter space mapping algorithm, converting the selected parameter combinations into continuous optimal encryption protocol configurations, and correcting configuration conflicts through robustness constraints;

[0012] Step S5: Iteratively optimizing encryption protocol configurations and network model parameters, using a meta-learning framework to integrate local parameter sensitivity and global attack resistance, and outputting an adaptive post-quantum cryptographic algorithm.

[0013] Preferably, in the step S1, the adaptive noise injection and parameter normalization preprocessing comprises the following operations:

[0014] Injecting adversarial noise through quantum attack pattern analysis to simulate parameter perturbation in a post-quantum scenario;

[0015] Using feature decomposition-based parameter grouping normalization to eliminate the dimensional differences between different encryption protocol parameters;

[0016] The sparse parameter data is filled by graph diffusion regularization using the topological structure prior of the parameter correlation graph.

[0017] Preferably, in the step S2, the parameter correlation graph embedding module captures the nonlinear dependency between parameters through a multi-head attention mechanism; the quantum attack resistant evaluation module quantifies the quantum resistance of the parameter combination using a structure combining residual connection and gated recurrent unit; and the multi-objective optimization decision module balances the encryption efficiency and security indicators through a dynamic weight distribution strategy.

[0018] Preferably, in the step S3, the generator of the generative adversarial network takes noise parameters as input and outputs candidate quantum-resistant parameter combinations; the discriminator adopts a multi-level cascade structure to evaluate the protocol compatibility, side-channel attack resistance, and computational resource occupation of the parameter combination; and a sparse regularization term is introduced in the adversarial training to improve the parameter search efficiency.

[0019] Preferably, in the step S4, the differentiable parameter space mapping algorithm includes the following steps:

[0020] Initialize a high-dimensional configuration space based on parameter sensitivity;

[0021] Optimize the continuous mapping relationship of parameter configuration by differentiable projection gradient descent method;

[0022] Eliminate conflicting configurations based on stability detection algorithm, and repair protocol vulnerabilities using robustness constraints.

[0023] Preferably, in the step S5, the meta-learning framework optimizes the local parameter update strategy and global attack-resistant target through an implicit parameter sharing mechanism, and the constraint conditions include parameter update step adaptive adjustment, protocol version compatibility verification, and hardware resource dynamic allocation.

[0024] Preferably, the dynamic optimization network model uses a composite loss function in the training phase, including:

[0025] Adversarial evolution loss, which measures the survival probability of the parameter combination under simulated quantum attacks;

[0026] Protocol consistency loss, which calculates the Hamming distance between the candidate configuration and the standard protocol template;

[0027] Resource efficiency loss, which constrains the computational overhead and memory occupation of the configuration based on hardware performance indicators.

[0028] Preferably, the multi-head attention mechanism adopts a parameter correlation-driven weight distribution strategy, in which the attention weight is calculated by a joint function of parameter mutual information entropy and protocol dependency strength, and the specific expression is:

[0029] ​

[0030] wherein, denotes attention weight, is a parameter and mutual information entropy of, denotes protocol dependency strength, and is a trainable hyperparameter satisfying .

[0031] Preferably, the robustness constraint quantifies the severity of configuration conflicts through a formal verification tool, and builds a conflict resolution loss function to drive stability optimization in the parameter mapping process, and the specific expression is:

[0032] ;

[0033] wherein, denotes conflict resolution loss, denotes the severity of the th configuration conflict, is a weight coefficient corresponding to the conflict type, denotes the total number of conflict types.

[0034] Preferably, an adaptive task sampling strategy is introduced in the meta-learning framework, and the training task distribution is dynamically adjusted according to the parameter sensitivity, and the anti-quantum attack ability of the high vulnerability area is preferentially optimized, and the specific implementation is:

[0035] ;

[0036] wherein, denotes the sampling probability of the th training task, denotes the parameter vulnerability score, is a temperature coefficient, controlling the steepness of the task distribution, denotes the total number of tasks.

[0037] Compared with the prior art, the beneficial effects of the present application are:

[0038] In the data processing stage, by collecting multi-source cryptographic parameter data and performing adaptive noise injection and parameter normalization preprocessing, a standardized cryptographic parameter set resistant to quantum interference is generated. This process effectively improves the quality and availability of data, laying a solid foundation for subsequent algorithm optimization. For example, by analyzing the injection of anti-quantum attack noise, it simulates the real parameter disturbance in the post-quantum scenario, enabling the algorithm to resist quantum attack interference during the training process. Compared with algorithms that do not undergo such processing, the algorithm is more robust when faced with actual quantum attacks. Parameter grouping normalization based on eigenvalue decomposition eliminates the dimensional differences between different encryption protocol parameters, avoiding algorithm training bias caused by dimensional problems, enabling the algorithm to more accurately learn the internal relationships between parameters, and thus improving optimization effectiveness. The topological structure of the parameter correlation graph is used to fill the sparse parameter data with graph diffusion regularization, solving the information loss problem caused by sparse data and ensuring data integrity, providing the algorithm with more comprehensive and accurate information, which helps to generate better encryption protocol configurations.

[0039] In the model construction and training stage, the dynamic optimization network model based on multi-head attention and reinforcement learning, combined with the adversarial evolution strategy training, shows strong advantages. In existing technologies, multi-head attention mechanisms (such as Transformer) have been widely used in natural language processing (NLP) and computer vision (CV), but have not been widely used in cryptographic parameter optimization. Traditional encryption algorithm optimization relies more on heuristic search or mathematical theory derivation, rather than dynamic modeling based on deep learning. Reinforcement learning has been used for protocol parameter tuning or key management, but usually combined with static rules (such as fixed reward functions), without combining attention mechanisms or dynamic multi-objective optimization. For example, some research uses Q-learning to optimize the execution efficiency of encryption algorithms, but lacks the ability to model complex inter-parameter dependencies. The parameter correlation graph embedding module captures the nonlinear dependencies between parameters through the multi-head attention mechanism, enabling the algorithm to discover parameter relationships that traditional methods cannot find, allowing the algorithm to have a deeper understanding of the complex cryptographic parameter space. For example, when analyzing multiple encryption protocol parameters (such as key length, encryption round number, hash function type, etc.), the multi-head attention mechanism can focus on the relationships between parameters from different angles, accurately grasping their comprehensive impact on encryption effectiveness and security. The anti-quantum attack evaluation module uses a structure combining residual connections and gated recurrent units, which can accurately quantify the anti-quantum strength of parameter combinations, providing a reliable basis for selecting anti-quantum parameter combinations. The multi-objective optimization decision module balances encryption efficiency and security indicators through a dynamic weight allocation strategy, enabling the algorithm to flexibly adjust encryption strategies according to different application scenario requirements, improving encryption efficiency while ensuring security, or maximizing security under specific efficiency requirements.

[0040] In the training process, a generative adversarial network is used to simulate a quantum attack scenario, dynamically screen anti-quantum parameter combinations, and introduce a sparse regularization term to improve parameter search efficiency. This not only enables the algorithm to evolve in a simulated quantum attack environment, filtering out parameter combinations with better quantum resistance, but also reduces the time and computational resource consumption of model training. At the same time, the discriminator of the generative adversarial network uses a multi-level cascade structure to evaluate the protocol compatibility, side-channel attack resistance, and computational resource occupation of the parameter combination, ensuring that the selected parameter combination performs well in multiple key performance indicators, improving the practicality and reliability of the algorithm.

[0041] Based on the differentiable parameter space mapping algorithm, the screened parameter combinations are converted into continuous optimal encryption protocol configurations, and the configuration conflicts are corrected through robustness constraints. The initialization of the high-dimensional configuration space based on parameter sensitivity, combined with the differentiable projected gradient descent method, optimizes the continuous mapping relationship of the parameter configuration, enabling efficient search for the optimal solution in a complex parameter space. Based on the stability detection algorithm, conflicting configurations are removed, and robustness constraints are used to repair protocol vulnerabilities, ensuring the accuracy and stability of the generated encryption protocol configuration, and avoiding encryption failures or security vulnerabilities caused by configuration conflicts.

[0042] In the iterative optimization process, the meta-learning framework is used to integrate local parameter sensitivity and global attack resistance to output an adaptive post-quantum cryptography algorithm. The meta-learning framework optimizes the local parameter update strategy and global attack resistance target through an implicit parameter sharing mechanism, while meeting the constraints of adaptive parameter update step size, protocol version compatibility verification, and dynamic allocation of hardware resources. This enables the algorithm to adaptively optimize in different application scenarios and hardware environments, improving the algorithm's versatility and making full use of hardware resources to further enhance encryption efficiency and security. The adaptive task sampling strategy is introduced to dynamically adjust the training task distribution based on parameter sensitivity, prioritizing optimization of the anti-quantum attack capability in high vulnerability areas, which can specifically improve the security of the algorithm in critical areas and effectively reduce the overall security risk of the system. Compared with general frameworks such as MAML / Reptile, the adaptive task sampling + hardware resource constraint implementation of domain-specific optimization is more suitable for the rapid iteration needs of post-quantum cryptography. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The working principle diagram of the AI-driven post-quantum cryptography algorithm adaptive optimization method described in the present application.

[0044] Figure 2 The working principle diagram of the dynamic optimization network model.

[0045] Figure 3 The working principle diagram of the differentiable parameter space mapping algorithm.

[0046] Figure 4 Optimization diagram for meta-learning framework. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0048] Please refer to Figures 1-4 The present application provides an AI-driven post-quantum cryptographic algorithm adaptive optimization method, and the overall implementation scheme is as follows:

[0049] Step S1: Multi-source cryptographic parameter data acquisition and preprocessing: widely collect multi-source cryptographic parameter data, which includes quantum attack simulation results, classical encryption protocol parameters and hardware performance indicators. In order to make the data have anti-quantum interference ability and facilitate subsequent processing, adaptive noise injection and parameter normalization preprocessing are required. By deeply analyzing the quantum attack mode, anti-noise is injected into the data to simulate the possible parameter disturbance in the post-quantum scenario. A parameter grouping normalization method based on feature decomposition is adopted to effectively eliminate the dimensional difference between different encryption protocol parameters, ensuring that each parameter is analyzed on the same scale. At the same time, using the topological structure prior knowledge of the parameter correlation graph, the sparse parameter data is filled by graph diffusion regularization, making the data more complete and accurate. After these processes, a standardized cryptographic parameter set resistant to quantum interference is generated;

[0050] Step S2: Construct a dynamic optimization network model: construct a dynamic optimization network model based on multi-head attention and reinforcement learning, which is composed of a parameter correlation graph embedding module, an anti-quantum attack evaluation module and a multi-objective optimization decision module. The parameter correlation graph embedding module uses the multi-head attention mechanism to effectively capture the complex nonlinear dependence between parameters. The anti-quantum attack evaluation module adopts the structure of residual connection combined with the gated recurrent unit to accurately quantify the anti-quantum strength of parameter combination. The multi-objective optimization decision module seeks a balance between encryption efficiency and security indicators through a dynamic weight allocation strategy;

[0051] Step S3: Training the dynamic optimization network model: the dynamic optimization network model is trained using an adversarial evolution strategy. The quantum attack scenario is simulated using a generative adversarial network. The generator takes noise parameters as input and outputs candidate quantum-resistant parameter combinations. The discriminator uses a multi-level cascade structure to evaluate the parameter combinations in terms of protocol compatibility, resistance to side-channel attacks, and computational resource usage. During the adversarial training process, a sparse regularization term is introduced to improve the efficiency of parameter search and dynamically select parameter combinations with excellent performance.

[0052] Step S4: Conversion and correction of encryption protocol configuration: based on the differentiable parameter space mapping algorithm, the selected parameter combinations are converted into continuous optimal encryption protocol configurations. First, a high-dimensional configuration space based on parameter sensitivity is initialized, and then the differentiable projection gradient descent method is used to optimize the continuous mapping relationship of the parameter configuration. In this process, conflicting configurations are removed based on the stability detection algorithm, and vulnerabilities in the protocol are repaired using robustness constraints to ensure the accuracy and stability of the encryption protocol configuration.

[0053] Step S5: Iterative optimization and output algorithm: the encryption protocol configuration and network model parameters are optimized through iteration. The meta-learning framework is used to integrate local parameter sensitivity and global attack resistance. The meta-learning framework optimizes the local parameter update strategy and global attack resistance target through implicit parameter sharing, while satisfying constraints such as adaptive parameter update step size, protocol version compatibility verification, and dynamic allocation of hardware resources. After continuous iteration and optimization, the adaptive post-quantum cryptographic algorithm is finally output.

[0054] The following embodiments 1 to 5 will further illustrate the implementation of the present application.

[0055] Embodiment 1:

[0056] In step S1, adaptive noise injection and parameter normalization preprocessing are key steps to improve data quality and enhance algorithm resistance to quantum interference.

[0057] For the operation of injecting anti-quantum noise based on quantum attack mode analysis, in actual application scenarios, first, various quantum attack methods need to be thoroughly studied and classified. Quantum attacks have various forms, such as quantum computers using quantum bits in parallel to interfere with key generation, encryption, and decryption processes in classical encryption protocols to varying degrees. By analyzing the characteristics and rules of these attack modes, the injection method and intensity of noise are determined. For example, for quantum key attacks, simulate the interference caused by attackers using quantum measurement techniques to obtain key-related information, inject specific types of noise into parameter data during key generation, and make it closer to the real post-quantum attack environment.

[0058] When the parameter grouping normalization based on eigenvalue decomposition is adopted, taking a data set containing multiple encryption protocol parameters as an example, it is assumed that there are three different encryption protocol parameters A, B and C in the data set, the value range of A parameter is 0-1000, the value range of B parameter is 0-1, and the value range of C parameter is 10-100. Due to the large dimension difference, direct analysis of these parameters will lead to inaccurate results. Based on the eigenvalue decomposition method, the matrix composed of these parameters is first decomposed into eigenvectors and eigenvalues. According to the size of the eigenvalue, the parameters are grouped, and the parameters with high correlation are grouped in the same group. Then, each group of parameters is normalized, for example, the maximum-minimum normalization method is used to map the value range of each group of parameters to 0-1, which eliminates the dimension difference between different encryption protocol parameters, making the subsequent analysis more scientific and accurate.

[0059] When the sparse parameter data is filled by graph diffusion regularization using the topological structure of the parameter correlation graph, a parameter correlation graph is constructed. It is assumed that there are five parameters P1, P2, P3, P4 and P5, and a graph structure is constructed according to their internal relationship. If P1 is associated with P2 and P3, and P2 is associated with P4, etc. When some parameter data is sparse, for example, the data of P3 is missing, the data of P3 is filled by a graph diffusion algorithm according to the data of other parameters associated with P3 (such as P1) using the topological structure of the graph. In the filling process, the correlation strength between parameters is considered, and the parameters with high correlation strength have greater influence on the filling result, so as to ensure that the filled data can better reflect the internal relationship between parameters and improve the integrity and reliability of the data.

[0060] Embodiment 2:

[0061] In step S2, the parameter correlation graph embedding module, the anti-quantum attack evaluation module and the multi-objective optimization decision module cooperate with each other to provide key support for the optimization of the algorithm.

[0062] The parameter correlation graph embedding module captures the nonlinear dependency between parameters by using a multi-head attention mechanism. In actual application, it is assumed that there are multiple encryption protocol parameters, such as key length, encryption round number, hash function type, etc. These parameters are not in a simple linear relationship, and the multi-head attention mechanism can mine the relationship between them from different angles. When calculating the attention weight, the formula is used. Wherein, represents the attention weight, which determines the importance of focusing on different parameters; is the mutual information entropy of parameter and , the mutual information entropy is used to measure the information sharing degree between two parameters, and the greater the mutual information entropy, the closer the correlation between the two parameters. represents the protocol dependency strength, which reflects the importance of parameters in a specific encryption protocol and the mutual dependency relationship; and are trainable hyperparameters, satisfying By adjusting these two hyperparameters, the influence of mutual information entropy and protocol dependency strength on attention weights can be flexibly balanced. In the actual training process, the values of and are constantly adjusted, so that the model can more accurately capture the dependency relationship between parameters.

[0063] The anti-quantum attack evaluation module adopts a structure combining residual connection with a gated recurrent unit. Taking the evaluation of the ability of a certain parameter combination to resist quantum attacks as an example, suppose that the input parameter combination contains multiple parameter information. The presence of residual connection enables the model to better transmit information during training, avoiding the problem of gradient vanishing. The gated recurrent unit can dynamically decide which information needs to be retained and updated according to the current input and previous memory state. When processing quantum attack-related data, the gated recurrent unit will judge the anti-quantum strength of the current parameter combination according to the attack feature information received at different times and combine the memory of similar attacks before. For example, when detecting a sudden increase in quantum attack strength, the gated recurrent unit can timely adjust the internal state to more accurately evaluate the anti-attack ability of the parameter combination in this case.

[0064] The multi-objective optimization decision module balances the encryption efficiency and security indicators through a dynamic weight distribution strategy. In actual application scenarios, different applications have different requirements for encryption efficiency and security. For example, in real-time communication scenarios, the requirement for encryption efficiency is higher; while in financial data encryption scenarios, the requirement for security is more stringent. The multi-objective optimization decision module will dynamically adjust the weights of encryption efficiency and security indicators according to the needs of specific application scenarios. If the current application pays more attention to encryption efficiency, the weight of the efficiency indicator will be appropriately increased, and the encryption speed will be optimized first under the premise of ensuring a certain security; on the contrary, if more attention is paid to security, the weight of the security indicator will be increased to sacrifice a certain efficiency in exchange for higher security.

[0065] Embodiment 3:

[0066] In step S3, the generator of the adversarial network is generated to input the noise parameters and output the candidate quantum-resistant parameter combinations. In actual operation, the generator is a neural network structure that receives randomly generated noise parameters. These noise parameters can be random numbers conforming to a certain distribution (such as a normal distribution). The generator generates a series of candidate quantum-resistant parameter combinations by learning and transforming the noise parameters. For example, the generator may generate parameter combinations of different key lengths and encryption algorithm combinations to explore more optimal quantum-resistant parameter settings. The generator continuously adjusts its own parameters during the training process, so that the generated parameter combinations are closer to the real quantum-resistant parameters.

[0067] The discriminator adopts a multi-level cascade structure to evaluate the protocol compatibility, side-channel attack resistance, and computational resource occupation of the parameter combinations. Taking a dataset containing multiple parameter combinations as an example, the first layer of the discriminator may first evaluate the protocol compatibility of the parameter combinations. For example, it checks whether the generated parameter combinations conform to the specifications of a specific encryption protocol and whether there are parameter settings that conflict with the protocol. The second layer evaluates the side-channel attack resistance by analyzing the performance of the parameter combinations when facing side-channel attacks (such as power analysis attacks, electromagnetic radiation attacks, etc.) to determine their attack resistance. The third layer evaluates the computational resource occupation by considering the demand for computational resources (such as CPU and memory) during encryption and decryption processes. Through this multi-level cascade structure, the performance of the parameter combinations can be evaluated comprehensively and in detail.

[0068] A sparse regularization term is introduced in the adversarial training to improve the efficiency of parameter search. The role of the sparse regularization term is to make some parameters in the model approach 0, thereby achieving the purpose of filtering important parameters. During training, assume that there are a large number of parameters in the model, some of which have little contribution to the performance of the model or are redundant. By introducing a sparse regularization term, such as an L1 regularization term, the model will automatically adjust the parameters during training, causing some unimportant parameters to gradually become 0. This not only reduces the complexity of the model and improves the training speed, but also avoids overfitting, making the model more efficient and accurate in searching for quantum-resistant parameter combinations.

[0069] In the training phase, a composite loss function is used, where the adversarial evolution loss measures the survival probability of parameter combinations under simulated quantum attacks. Assuming that in a simulated quantum attack experiment, a set of parameter combinations are attacked multiple times, and the proportion of times that they are not broken out of the total number of attacks is calculated, this proportion is the survival probability of the parameter combination. The adversarial evolution loss measures this survival probability to encourage the model to generate parameter combinations with better quantum attack resistance. The protocol consistency loss calculates the Hamming distance between the candidate configuration and the standard protocol template. The standard protocol template is an encryption protocol configuration that meets industry standards and security requirements. The Hamming distance is used to measure the number of different bits between two configurations. For example, the candidate configuration is "10101" and the standard protocol template is "11100", then their Hamming distance is 2. By minimizing the protocol consistency loss, the candidate configuration is closer to the standard protocol template. The resource efficiency loss constrains the calculation overhead and memory usage of the configuration based on hardware performance indicators. Taking a specific hardware device as an example, its CPU processing capacity and memory size are limited. The resource efficiency loss calculates the calculation overhead and memory usage of the parameter combination when running on the hardware according to these performance indicators of the hardware, and adjusts the parameter combination to make the calculation overhead and memory usage within the range that the hardware can withstand, while ensuring the performance of the encryption algorithm.

[0070] Embodiment 4:

[0071] In step S4, the differentiable parameter space mapping algorithm first initializes a high-dimensional configuration space based on parameter sensitivity. In practical applications, different encryption protocol parameters have different effects on encryption effectiveness and security, i.e., different parameter sensitivities. For example, a small change in key length may have a large impact on encryption security, while some fine-tuning parameters in the encryption algorithm have relatively small impact on security. According to this parameter sensitivity information, a high-dimensional configuration space is constructed. In this space, each dimension represents a parameter, and the points in the space represent different parameter combinations. During initialization, an initial range is set for each parameter based on experience or prior knowledge to determine the boundaries of the configuration space.

[0072] The continuous mapping relationship of parameter configuration is optimized by the differentiable projection gradient descent method. In the optimization process, it is assumed that the goal is to find a set of parameter combinations that make the encryption protocol achieve the best under certain security and efficiency requirements. The differentiable projection gradient descent method adjusts the parameters step by step based on the gradient information of the current parameter combination on the target function. For example, the target function is a function that considers encryption efficiency and security comprehensively, and the gradient represents the rate of change of the function at the current parameter point. By moving the parameters in the opposite direction of the gradient, the value of the target function is gradually reduced, and a better parameter configuration is found. During the movement of the parameters, it is necessary to ensure that the parameters are always within the legal configuration space, which is the role of the projection operation.

[0073] Based on the stability detection algorithm, conflict configurations are eliminated, and robustness constraints are used to repair protocol vulnerabilities. The stability detection algorithm analyzes the running conditions of parameter combinations under different conditions to determine whether there is a conflict. For example, check if two parameters will produce contradictory results during encryption and decryption, or if it will cause the encryption protocol to malfunction. When a conflict configuration is detected, it is eliminated. Robustness constraints quantify the severity of configuration conflicts using formal verification tools and build a conflict resolution loss function to drive stability optimization during parameter mapping. In the formula represents the conflict resolution loss, which measures the impact of configuration conflicts on the overall encryption protocol; represents the severity of the class configuration conflict, and different types of conflicts (such as key conflicts, algorithm compatibility conflicts, etc.) have different severity values; is the weight coefficient corresponding to the conflict type, used to adjust the importance of different conflict types in the conflict resolution loss calculation; represents the total number of conflict types. By minimizing the conflict resolution loss, protocol vulnerabilities are repaired, making the parameter mapping process more stable, and the generated encryption protocol configuration more reliable.

[0074] Embodiment 5:

[0075] In step S5, the meta-learning framework and the adaptive task sampling strategy are the core elements of iterative optimization of encryption protocol configuration and network model parameters, and play a key role in improving the overall performance of the algorithm.

[0076] The meta-learning framework uses an implicit parameter sharing mechanism to cooperatively optimize the local parameter update strategy and the global anti-attack target. In actual application scenarios, there are many different types of encryption tasks, such as lightweight encryption tasks between Internet of Things devices and high-intensity encryption tasks in enterprise data centers. These tasks have obvious differences in encryption requirements. Internet of Things devices are more focused on lightweight encryption and low energy consumption due to resource constraints, while enterprise data centers have very high requirements for encryption security and data integrity.

[0077] The implicit parameter sharing mechanism of the meta-learning framework can extract the common parts of parameters between different tasks. For example, although different tasks have different requirements for key length and encryption algorithm, there may be similarities in the basic mathematical principles and some intermediate calculation steps of key generation. By sharing these common parameters, the model can use the experience accumulated by other tasks to avoid repeated learning and speed up convergence when updating local parameters for a specific task. At the same time, during the process of updating local parameters, the global anti-attack target is always optimized to ensure that the algorithm has the ability to resist quantum attacks regardless of the encryption task it faces. ​

[0078] The constraints of the meta-learning framework include adaptive adjustment of parameter update step, verification of protocol version compatibility, and dynamic allocation of hardware resources. In terms of adaptive adjustment of parameter update step, at the initial stage of training, the model has limited understanding of the optimal value of the parameters. At this time, a larger update step is used to quickly explore the parameter space and accelerate model convergence. As the training progresses, when the model gradually approaches the optimal solution, the update step is reduced to finely adjust the parameters and avoid missing the optimal value. For example, in gradient-based optimization algorithms, the learning rate is dynamically adjusted according to the change of the gradient to achieve adaptive change of the update step.

[0079] Protocol version compatibility verification is also crucial. Cryptographic protocols are in a process of continuous development and update, and new versions will fix security vulnerabilities of old versions and improve encryption performance. When iteratively optimizing the configuration of the encryption protocol, the meta-learning framework verifies whether the generated configuration is compatible with the current protocol version. For example, when a new encryption protocol version is adopted, check whether the encryption algorithm, key exchange method, etc. in the configuration comply with the specifications of the new version to ensure the stability and security of the encryption system and avoid encryption and decryption failures due to version incompatibility.

[0080] Dynamic allocation of hardware resources is to reasonably allocate hardware resources according to the actual resource status of the hardware devices. Different hardware devices, such as high-performance servers and low-power embedded devices, have significant differences in computing power, memory capacity, etc. The meta-learning framework will monitor the usage of hardware resources in real time. For computationally intensive encryption tasks, more computing resources are preferentially allocated; for tasks with high memory requirements, memory space is reasonably allocated. For example, when running encryption algorithms on embedded devices, due to their limited memory and computing power, the meta-learning framework will adjust the execution mode of the algorithm, adopt a more lightweight encryption strategy, and reasonably allocate limited resources to ensure the smooth completion of encryption tasks.

[0081] The adaptive task sampling strategy introduced in the meta-learning framework dynamically adjusts the training task distribution according to parameter sensitivity, and preferentially strengthens the quantum attack resistance of high vulnerability areas. In the formula represents the sampling probability of the th training task, which determines the likelihood of each task being selected in the training process. represents the parameter vulnerability score, which is obtained by comprehensively evaluating factors such as the sensitivity of the parameter to quantum attacks and the impact on encryption security. The higher the score, the more sensitive the task area of the parameter to quantum attacks, and the higher the vulnerability. is the temperature coefficient, which controls the steepness of the task distribution. When ​When larger, the task sampling probability is more concentrated on tasks with high vulnerability scores; when When smaller, the task sampling probability distribution is relatively uniform. represents the total number of tasks, used to normalize the sampling probability of all tasks.

[0082] In actual operation, by calculating the parameter vulnerability score of each task, the task sampling probability is adjusted according to the above formula. For example, in a system containing multiple encryption tasks, task A involves parameters that are vulnerable to quantum attacks, and its parameter vulnerability score is high; the parameters of task B are relatively stable, and the score is low. The adaptive task sampling strategy will make task A have a higher probability of being selected for training, giving priority to improving its resistance to quantum attacks, thereby achieving key optimization of high vulnerability areas in the entire encryption system and enhancing the overall security of the system.

[0083] It should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device.

[0084] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An AI-driven adaptive optimization method for post-quantum cryptography algorithms, characterized in that, Includes the following steps: Step S1: Collect multi-source cryptographic parameter data, including quantum attack simulation results, classical encryption protocol parameters and hardware performance indicators. Through adaptive noise injection and parameter normalization preprocessing, generate a standardized cryptographic parameter set resistant to quantum interference. Step S2: Construct a dynamic optimization network model based on multi-head attention and reinforcement learning. The model includes a parameter correlation graph embedding module, a quantum attack resistant evaluation module, and a multi-objective optimization decision module. The parameter correlation graph embedding module captures the nonlinear dependencies between parameters through a multi-head attention mechanism; the quantum attack resistance evaluation module adopts a structure combining residual connections and gated recurrent units to quantify the quantum resistance strength of parameter combinations; the multi-objective optimization decision module balances encryption efficiency and security indicators through a dynamic weight allocation strategy. Step S3: Train the dynamically optimized network model using an adversarial evolution strategy, simulate quantum attack scenarios by generating adversarial networks, and dynamically select anti-quantum parameter combinations; Step S4: Based on the differentiable parameter space mapping algorithm, the filtered parameter combinations are converted into continuous optimal encryption protocol configurations, and configuration conflicts are corrected through robustness constraints. Step S5: Iteratively optimize the encryption protocol configuration and network model parameters, and use the meta-learning framework to integrate local parameter sensitivity and global anti-attack capability to output an adaptive post-quantum cryptography algorithm; In step S1, the adaptive noise injection and parameter normalization preprocessing includes the following operations: By analyzing the injection of adversarial noise through quantum attack modes, parameter perturbations in a post-quantum scenario are simulated. A parameter grouping normalization based on eigenvalue decomposition is adopted to eliminate the dimensional differences between parameters of different encryption protocols; By leveraging the prior topological structure of the parameter association graph, graph diffusion regularization is applied to fill sparse parameter data. In step S4, the differentiable parameter space mapping algorithm includes the following steps: Initialize a parameter-sensitive high-dimensional configuration space; The continuous mapping relationship of parameter configuration is optimized by differentiable projective gradient descent method; Conflicting configurations are eliminated based on stability detection algorithms, and robustness constraints are used to fix protocol vulnerabilities. In step S5, the meta-learning framework jointly optimizes the local parameter update strategy and the global anti-attack target through an implicit parameter sharing mechanism. The constraints include adaptive adjustment of parameter update step size, protocol version compatibility verification, and dynamic allocation of hardware resources.

2. The method according to claim 1, characterized in that: In step S3, the generator of the generative adversarial network takes noise parameters as input and outputs candidate quantum-resistant parameter combinations. The discriminator adopts a multi-layer cascaded structure to evaluate the protocol compatibility, resistance to side-channel attacks, and computational resource consumption of parameter combinations. A sparse regularization term is introduced in the adversarial training to improve the efficiency of parameter search.

3. The method according to claim 1, characterized in that: The dynamically optimized network model employs a composite loss function during the training phase, including: Adversity to evolutionary loss measures the survival probability of parameter combinations under simulated quantum attacks; Protocol consistency loss is calculated by determining the Hamming distance between the candidate configuration and the standard protocol template. Resource efficiency loss refers to the computational overhead and memory usage constrained by hardware performance metrics.

4. The method according to claim 1, characterized in that: The multi-head attention mechanism employs a parameter-dependent weight allocation strategy, where the attention weights are calculated using a joint function of parameter mutual information entropy and protocol dependency strength, specifically expressed as: ； in, Indicates attention weights, For parameters and Mutual information entropy, Indicates the strength of protocol dependency. and For trainable hyperparameters, satisfying .

5. The method according to claim 1, characterized in that: The robustness constraints are quantified using formal verification tools to determine the severity of configuration conflicts, and a conflict resolution loss function is constructed to drive stability optimization during the parameter mapping process. The specific expression is as follows: ； in, This indicates that the conflict has been resolved and losses have been mitigated. Indicates the first The severity of class configuration conflicts, These are the weighting coefficients corresponding to the conflict types. This indicates the total number of conflict types.

6. The method according to claim 1, characterized in that: The meta-learning framework introduces an adaptive task sampling strategy, which dynamically adjusts the training task distribution based on parameter sensitivity, prioritizing the optimization of the quantum attack resistance of highly vulnerable regions. Specifically, this is implemented as follows: ； in, Indicates the first The sampling probability of each training task. Indicates the vulnerability score of the parameter. The temperature coefficient controls the steepness of the task distribution. This indicates the total number of tasks.

Images