A password strength evaluation system based on a large model

Abstract

This invention relates to the field of information security technology, specifically to a password strength evaluation system based on a large model. The system includes cleaning and standardizing password training data to form samples, extracting relevant fields from user information in a structured manner, calculating the number of password guesses to be evaluated using Markov and UNCM models and taking the minimum value, generating enhancement suggestions, and verifying them for security, deassociation, and standardization. If a suggestion fails, feedback iteration or rollback to a template is used to generate a new password, and finally, an evaluation report is output. This invention overcomes the black-box defects of traditional evaluations and the security illusion of large models by integrating multidimensional probabilistic models and large language models, and introducing deterministic constraint verification and dynamic feedback closed loops. It achieves a balance between measurability and interpretability in password security evaluation, and stably outputs security enhancement schemes through verification, retry, and rollback mechanisms.

Description

Technical Field

[0001] This invention relates to the field of information security technology, and in particular to a password strength evaluation system based on a large model. Background Technology

[0002] With the rapid development of internet technology and the increasing prevalence of various digital identity authentication scenarios, passwords, as the most widely used authentication method, are facing increasingly serious security problems due to frequent large-scale leaks. Accurately assessing password strength has become a key line of defense for ensuring information system security. Currently, the industry mainly uses "guessing count" as a quantitative assessment indicator of strength, which is the average number of attempts required for an attacker to guess the password using the optimal strategy. Existing assessments mainly rely on statistical methods (such as Markov models) and deep learning-based methods (such as UNCM models). These traditional probabilistic models can uniformly quantify the output into a comparable number of guesses, objectively characterizing the true security strength of the password.

[0003] However, existing password assessment methods have significant limitations. On the one hand, traditional quantitative models can only output cold, hard numbers or levels. As a black box, they lack natural language interpretation of the assessment results, making it difficult for ordinary users to understand the real vulnerabilities of passwords and obtain practical, personalized modification suggestions. On the other hand, although there have been attempts in recent years to introduce large language models to solve the above-mentioned interpretability problems, general large models are prone to creating security illusions in the field of password security. The suggested passwords they generate often seem complex, but they are easily defeated by professional guessing models. Large models lack mathematical constraints for dimensionality reduction and correlation when processing users' personal information, and due to the lack of effective verification and retry mechanisms at the system level, once a large model outputs poor suggestions, it cannot automatically correct errors, making it difficult to guarantee the absolute reliability of security suggestions.

[0004] Existing single traditional mathematical models or single large language models are not perfectly capable of fulfilling the task of highly reliable password security assistance. There is an urgent need in the field for a new password strength evaluation architecture that can integrate the rigorous measurement of traditional deterministic mathematical models with the advantages of nondeterministic semantic generation of large language models. By constructing a rigorous quantitative constraint and dynamic error correction closed loop, it can effectively overcome the technical defects of uncontrollable generation results of large models, achieve the unity of measurability and interpretability of password strength, and thus provide users with practical modification suggestions with accurate evaluation and clear explanation. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a password strength evaluation system based on a large model.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a password strength evaluation system based on a large model, the system comprising: The data preprocessing module performs cleaning, character encoding unification, and format adaptation on the password training data to obtain training samples; The user information parsing module extracts the personal information entered by the user in a structured manner to obtain structured fields related to password security analysis; The multidimensional probability measurement module constructs a Markov password probability model and an UNCM model based on the training samples; and after receiving the password to be evaluated, it calculates the number of guesses for the password to be evaluated under each model and extracts the minimum value as the quantization benchmark value. The semantic enhancement suggestion module utilizes a large language model, combining the structured fields with the password to be evaluated, to generate an initial candidate set that includes a password strength level description, vulnerability analysis results, and enhancement suggestions. The deterministic logic constraint module establishes a verification matrix and performs joint reverse verification on the enhancement suggestions in the initial candidate set. The verification matrix includes: Security enhancement constraint: The second-order guess number of the new password corresponding to the enhancement suggestion under the multi-dimensional probability measurement module, relative to the quantization benchmark value, exceeds the preset quantization threshold. Semantic deassociation constraint: The statistical correlation between the new password and the structured fields is lower than a preset threshold, so as to reduce the dependence of the new password on the user's personal information features; Structural constraints: The new password must meet the preset length limit and character composition rules; The self-optimization closed-loop control module establishes a feedback link between the semantic enhancement suggestion module and the deterministic logic constraint grid. If the enhancement suggestion fails the joint reverse verification, the quantization deviation of the verification failure is sent back to the large language model as a feedback parameter to generate enhancement suggestions again. If the enhancement suggestion generated after a preset number of iterations still fails the verification, a fallback mechanism is triggered, a preset strong password template is called to generate a fallback password, and the multidimensional probability measurement module performs a second guessing number calculation on the fallback password. The results output module generates an evaluation report based on the validated enhancement suggestions or rollback password.

[0007] As a further aspect of the present invention, the data preprocessing module performs the following processing on the password training data: Delete password samples whose length is less than the preset lower limit or greater than the preset upper limit, handle abnormal data with special character encoding, and convert password samples to UTF-8 encoding format.

[0008] As a further aspect of the present invention, when the data preprocessing module performs format adaptation for the UNCM model, it maps the password training data into a four-tuple feature sequence. The four-tuple feature sequence includes, in sequence, the email username, the email service provider, the top-level domain, and the password field, and the email service provider and the top-level domain are taken from a preset specific domain dictionary.

[0009] As a further aspect of the present invention, the Markov password probability model is a probability model constructed based on the nth-order character transition probability; the multidimensional probability measurement module determines the character transition probability based on the frequency of character combinations in the training samples, and performs weight correction on the zero-probability transition term through a smoothing algorithm.

[0010] As a further aspect of the present invention, the multidimensional probability measurement module generates a set of candidate passwords based on the Markov password probability model and the UNCM model, respectively, and sorts the set of candidate passwords in descending order according to the generation probability of each candidate password. The absolute position index of the password to be evaluated or the new password in the sorting result represents the number of guesses under the corresponding model.

[0011] As a further aspect of the present invention, the structured fields include at least one of the following: name, gender, birthday, mobile phone number, email address, and account name; The semantic enhancement suggestion module generates personal information association analysis results based on the cross-comparison and fuzzy matching relationship between the structured fields and the password to be evaluated.

[0012] As a further aspect of the present invention, the enhancement suggestion includes an optimized version of the original password and a standard enhanced version. The optimized version of the original password is generated based on the password to be evaluated by character insertion, structural reorganization, or adaptive length expansion to improve the password complexity while preserving the user's memory anchor. The character homomorphic substitution strategy is restricted or suppressed during the generation process to avoid introducing weak patterns that can be exploited by the rule-based guessing model. The standard enhanced version meets the preset length requirements and is generated using a template that includes at least two non-contiguous semantic English words, random numbers, and non-letter separators; The multidimensional probability measurement module performs parallel secondary guessing calculations on the optimized version and the standard enhanced version of the original password that have passed the joint reverse verification.

[0013] As a further aspect of the present invention, the semantic enhancement suggestion module determines the password strength level based on the mapping relationship between the quantification benchmark value and the preset strength grading rules, and generates an evaluation description according to the preset output template; When the password strength level reaches the preset security judgment level, the semantic enhancement suggestion module stops generating enhancement suggestions, the self-optimization closed-loop control module terminates the verification iteration, and the result output module directly outputs a release prompt that does not require modification.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention effectively overcomes the limitations of traditional evaluation models, which lack intuitive explanations and operational suggestions, by integrating traditional multidimensional probability measurement models with large language models and introducing deterministic logic constraint grids and dynamic self-optimization feedback loops. Simultaneously, by pre-setting matching rules, it reduces the risk of generated passwords being associated with user personal information. This not only achieves a perfect balance between rigorous measurability and deep interpretability in password security assessment, but also significantly lowers the cognitive threshold for users to construct high-strength passwords through a fully automated process of joint reverse verification, feedback retry, and fallback. It ensures that the system can stably and efficiently output mathematically verified, highly resistant personalized security password enhancement solutions with strong resistance to advanced targeted dictionary attacks in complex application scenarios. Attached Figure Description

[0015] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0017] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0018] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0019] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] See Figure 1 This invention provides a password strength evaluation system based on a large model. The specific implementation of this system is as follows: The data preprocessing module performs cleaning, character encoding unification, and format adaptation on the password training data to obtain training samples.

[0022] As a preferred implementation, the data preprocessing module performs the following processing on the password training data: Password samples shorter than a preset lower limit or longer than a preset upper limit are deleted. Abnormal data with special character encodings are handled, and all password samples are uniformly converted to UTF-8 encoding format. In practical engineering, the system pre-constructs a large-scale password training dataset covering both Chinese and English bilingual features.

[0023] For example, the English dataset selected publicly leaked passwords from Rockyou, and after removing dirty data containing garbled characters or non-ASCII extended characters, more than 2 million valid records were selected; the Chinese dataset selected publicly leaked passwords from Dodonew and other sources, and more than 16 million valid records were selected. Regarding the specific methods of data collection and acquisition, the aforementioned publicly leaked password data was collected by downloading it from open-source security vulnerability databases (such as GitHub, Kaggle, and other data hosting platforms) using existing network transmission protocols such as HTTP / FTP through the server's network communication interface (such as a gigabit Ethernet card), and persistently storing it on the server's local high-capacity physical storage device (such as a solid-state drive SSD or disk array). Use Python to write data cleaning scripts, strictly limiting password length to a preset range of 4 to 64 characters, and use "iconv" or Python's native "encode / decode" library to uniformly convert to UTF-8 format, ensuring that "@", "#", and "..." are correctly formatted. Special symbols in Chinese and English, such as “”, can be correctly recognized by the tensor quantization network during subsequent training.

[0024] Furthermore, when the data preprocessing module performs format adaptation for the UNCM model, it maps the password training data into a four-tuple feature sequence. This four-tuple feature sequence includes the email username, email service provider, top-level domain, and password field in sequence, and the email service provider and top-level domain are taken from a preset specific domain dictionary.

[0025] In specific implementation scenarios, an automated format conversion script was developed to address the specific input requirements of deep learning guessing networks such as the UNCM model. This script randomly selects parameters from a pre-defined dictionary for filling in the data: for example, the email service provider list (configs / vocabs / D0) contains typical characteristic words such as "qq, 163, gmail, outlook"; the top-level domain list (configs / vocabs / D1) contains "com, cn, net, org," etc. This transforms the original single password into a four-tuple sequence like "zhangsan163comzhangsan123." Simultaneously, a verification script iterates through and removes abnormal samples where the service provider or domain does not match the pre-defined dictionary. This pre-format adaptation step forces the neural network to actively capture and learn the high-dimensional implicit correlation features between the "network service environment context" and the "user's specific password setting habits" during subsequent autoregressive training.

[0026] The user information parsing module extracts the personal information entered by the user in a structured manner, and obtains structured fields related to password security analysis.

[0027] As a preferred implementation, the structured fields include at least one of the following: name, gender, date of birth, mobile phone number, email address, and account name; In practice, the system front-end provides corresponding input forms for users to fill in. In terms of hardware and technical means for data acquisition, users input the above information in a graphical user interface (GUI) through human-computer interaction peripherals (such as physical keyboards and touchscreens) on terminal devices (such as personal computers and smartphones). The terminal devices transmit the data to the cloud server at the system backend via wireless or wired network communication modules using secure network protocols such as HTTPS. After receiving the data, the backend extracts it using regular expressions combined with existing named entity recognition (NER) technology in natural language processing. For example, it extracts "ZhangSan", "19900101", and "13800138000". When a certain piece of information is not entered by the user or the system cannot clearly identify it (i.e., the extraction result is judged to be a mismatch or empty), the corresponding field is directly assigned the value "null" to ensure the continuity of subsequent verification logic and prevent the program from throwing a null pointer exception. The structured JSON format data is persisted in memory as input conditions for subsequent personal information correlation detection and enhanced suggestion generation.

[0028] The multidimensional probability measurement module constructs Markov password probability models and UNCM models based on training samples; and after receiving a password to be evaluated, it calculates the number of guesses for the password under each model and extracts the minimum value as the quantization benchmark value.

[0029] As a preferred implementation, the Markov password probability model is a probability model built based on n-order character transition probabilities. The multi-dimensional probability measurement module determines the character transition probabilities based on the frequency of character combinations in the training samples and corrects the weights of zero-probability transition terms through a smoothing algorithm. In a specific embodiment, the system implements a 4-order (4-gram) Markov model using Python, utilizes the collections library to fully statistically analyze the frequency of occurrence of four consecutive character combinations in the training samples, and calculates the probability state transition matrix accordingly. To address the "data sparsity" problem where rare character combinations in the real world cause probabilities to plummet to zero, the system introduces an additive smoothing formula: ; Among them, the smoothing penalty parameter Set to an empirical value of 0.01, effective character set size. Limited to 96 commonly printable characters. This formula has clear engineering implications, as it introduces small prior probabilities (i.e., smoothing penalty parameters) to all possible transition states. This effectively prevents the model from outputting zero probability when encountering unfamiliar character combinations, thus avoiding errors caused by broken computation chains. The numerator and denominator of the formula are both linear superpositions of frequencies and parameters, and the output is a dimensionless probability constant (strictly between 0 and 1), satisfying the normalization condition of the probability distribution and avoiding any dimensional conflict issues. For the UNCM model, the system is built on the PyTorch deep learning framework. To balance inference speed and generalization ability, lightweight parameters are configured for the English dataset (i.e., word embedding dimension set to 64, network layer number set to 2, and training epochs set to 2); while for the more complex Chinese Pinyin association dataset, expansion parameters are configured (word embedding dimension set to 256, hidden layer dimension set to 512, and training epochs set to 2).

[0030] Furthermore, the multidimensional probability measurement module generates candidate password sets based on the Markov password probability model and the UNCM model, respectively, and sorts the candidate password sets in descending order according to the generation probability of each candidate password. The absolute position index of the password to be evaluated or the new password in the sorting result represents the number of guesses under the corresponding model. For example, after receiving the user's password to be evaluated "admin1234", the system calls the Markov model and the UNCM model, respectively, and generates 10,000 candidate passwords through Monte Carlo sampling autoregression, and sorts them in descending order of cracking probability. Traversing the cumulative ranking array, the absolute index (ACC_RANK value) of the target password in the array is defined as the number of guesses required to crack the password. It is particularly important to note that, because different probability models focus on different aspects when capturing password vulnerabilities (Markov models are sensitive to consecutive keystrokes, while neural networks are sensitive to semantic associations), the system ultimately takes the "minimum value" of the guesses calculated by both models (i.e., the number of times it is easiest to crack the password, for example, calculating...). The result is This is set as the final quantitative benchmark value representing the true weaknesses of passwords, thereby ensuring that security assessments follow the strictest "barrel principle".

[0031] The semantic enhancement suggestion module utilizes a large language model, combined with structured fields and the password to be evaluated, to generate an initial candidate set that includes a description of the password strength level, vulnerability analysis results, and enhancement suggestions.

[0032] As a preferred implementation, the enhancement suggestions include an optimized version of the original password and a standard enhanced version; wherein, the optimized version of the original password is generated based on the password to be evaluated by character insertion, structural reorganization or adaptive length expansion. The standard enhanced version meets the preset length requirements and is generated using a template that includes at least two non-contiguous semantic English words, random numbers, and non-letter separators.

[0033] In the specific implementation, the system is used to enhance and optimize the security of user passwords. Its core is to generate candidate passwords by calling a general large language model. The large language model is integrated with the business system through an interface, as detailed below.

[0034] In this embodiment, the system establishes a communication connection with the large language model server via a RESTful API interface and based on the HTTPS protocol, using the large language model as the inference engine for password generation and optimization. Preferably, the large language model is a general-purpose language model fine-tuned with security policies, possessing text understanding and generation capabilities.

[0035] To ensure the security, stability, and controllability of the generated results, the system does not directly input the original natural language request when calling the large language model. Instead, it constructs the input content in a structured manner based on a preset prompt word construction strategy before making the call.

[0036] Specifically, the prompt word construction strategy adopts a hierarchical structure, which includes at least four parts: task role definition, input information description, generation constraint rules, and output format constraints.

[0037] The task role definition is used to define the functional role of the large language model, enabling it to perform tasks as a password security enhancement generation module, thereby avoiding the generation of content unrelated to the task.

[0038] The input information description is used to provide necessary contextual information to the large language model, including but not limited to the user's original password and user-related structured field information. Preferably, the structured field information includes username, account identifier, or other information that can be used to identify the user.

[0039] Generative constraint rules are used to limit the security policy of the generated results, and they must include at least the following constraints: (1) The generated password shall not directly contain or, through simple transformation, contain the username, original password, or other user-related fields in the input information; (2) The generated password must meet the preset length range and complexity requirements; (3) The generation process should take into account both memory and security, and can be performed by either optimized generation based on the original password or enhanced generation based on a random strategy. (4) The generated results need to have a high degree of unpredictability in order to improve the ability to resist guessing attacks.

[0040] Output format constraints are used to limit the output of the large language model to a structured data format so that the system can automatically parse and post-process it. Preferably, the output format is JSON and includes at least a security analysis field and multiple candidate password fields.

[0041] Based on the above-mentioned prompt word construction strategy, the system constructs standardized prompt content and sends it to the large language model. For example, the prompt content includes, but is not limited to, the following structure: Includes task description information to instruct the model to perform password security enhancement tasks; This includes input fields for providing the original password and user identification information; This includes generating task descriptions that require the model to generate password candidate results under different strategies; This includes descriptions of constraint rules used to limit the security and compliance of the generated results; Includes output format specifications to define the structure of the returned results.

[0042] After receiving the prompt, the large language model generates a structured result based on its internal reasoning mechanism and returns it to the system through the API interface.

[0043] After receiving the structured results, the system parses the returned content, extracts the candidate passwords and corresponding analysis information, and passes them to the subsequent rule verification module and large model judgment module for further filtering in order to obtain the target password that finally meets the security policy requirements.

[0044] Furthermore, in some embodiments, the system can also dynamically adjust the constraint parameters in the prompt words according to different business scenarios, including but not limited to the password length range, character complexity level, and generation strategy weight, thereby achieving adaptive optimization of the password generation strategy.

[0045] By employing the above methods, this embodiment can improve the diversity and usability of generated results while ensuring the security of user passwords, and effectively avoid the association between generated results and sensitive user information, thereby enhancing the overall system's security protection capabilities.

[0046] The deterministic logic constraint module establishes a verification matrix and performs joint reverse verification on the enhancement proposals in the initial candidate set. This module is the core feature of this invention that overcomes the "security illusion" of large models.

[0047] Furthermore, the verification matrix includes at least: (1) Security Enhancement Constraint: The second-order guess number of the new password corresponding to the enhancement suggestion under the multi-dimensional probability measurement module, relative to the quantization benchmark value, exceeds the preset quantization threshold. That is, the multi-dimensional probability measurement module performs the second-order guess number calculation in parallel on the optimized version of the original password and the standard enhanced version that have passed joint reverse verification. For example, if the guess number of the original password is 10,000, the system forces that the second-order estimated guess number of the generated new password must be greater than 10,000,000,000,000 (i.e., crossing the strength level corresponding to two orders of magnitude) in order to be considered to meet the security enhancement constraint. If it is determined that the enhancement does not exceed the threshold (i.e., consistency or security does not meet the standard), the password is intercepted and the retry logic of the self-optimization closed-loop control module is entered; (2) Semantic deassociation constraint: The degree of association between the new password and the structured fields must meet the preset deassociation requirements. In specific implementation, the system no longer relies on a single mathematical model to calculate the degree of association, but instead comprehensively evaluates the association between the new password and the user's structured information based on preset personal information matching rules and combined with the semantic judgment capability of the UNCM model.

[0048] Specifically, the system pre-sets multiple matching rules for detecting the relevance of personal information. These matching rules include at least one of the following: exact matching rules for names or account names, matching rules for name pinyin and its abbreviations, matching rules for birthdays and their common variations, matching rules for consecutive segments of mobile phone numbers, matching rules for email usernames, and matching rules for combinations or variations of the above fields. The system first checks the new password item by item according to these matching rules to identify whether it contains explicit features directly related to the user's personal information.

[0049] Based on rule-based matching detection, structured fields and candidate new passwords are passed as input to a large language model. Preset prompts guide the large language model to determine the potential semantic association, variant association, or combined association between the two, and output the association analysis results.

[0050] If the rule matching hits the preset risk condition, or the large language model determines that the new password has a clear association with the user's personal information, then the new password is determined to have failed the semantic deassociation constraint; conversely, if the rule matching fails and the large language model determines that there is no clear association, then the new password is determined to have met the deassociation requirement.

[0051] If any structured fields output by the user information parsing module contain missing items, the matching rule detection for the corresponding field is skipped, and the correlation determination continues based on the remaining valid fields. When all structured fields are empty, the constraint is considered passed by default. By combining rule matching with semantic determination, the reliance of password generation on user personal information features is effectively reduced, thereby reducing the risk of targeted guessing based on personal information.

[0052] (3) Structural constraints: Strictly follow regular expression matching to verify whether the new password meets the preset absolute length limit and whether it contains the four types of character features.

[0053] The self-optimization closed-loop control module establishes a feedback link between the semantic enhancement suggestion module and the deterministic logic constraint grid. If the enhancement suggestion fails the joint reverse verification, the quantization deviation of the verification failure is used as a feedback parameter and sent back to the large language model to generate enhancement suggestions again. If the enhancement suggestion generated after a preset number of iterations still fails the verification, a fallback mechanism is triggered, a preset strong password template is called to generate a fallback password, and the multidimensional probability measurement module performs a re-guessing count calculation on the fallback password.

[0054] In the specific engineering implementation, the system is set to a maximum of 3 retry iterations. If the enhancement suggestions generated by the large language model fail to be intercepted due to triggering the verification matrix, the closed-loop module will automatically construct error correction prompts and send them back for retry. The message sent back is, for example: "Error feedback: The password 'zhangsan2024!' you just generated contains the user's name feature, which does not pass the personal information matching rules, and the secondary estimated guess count is only 100,000, which does not meet the safety threshold. Please discard the user's name and year features, regenerate using completely random word roots, and ensure structural complexity."

[0055] Furthermore, if the large language model remains in a hallucinatory state and fails to output a valid password after three retries, the system will seamlessly trigger a "rollback mechanism" to ensure high service availability. The system directly blocks the output of the large language model and instead calls the server operating system's underlying high-entropy pseudo-random number generator (such as the existing / dev / urandom interface in Linux) combined with a strong password generation template (this template is enforced by regular expressions to include at least one uppercase and lowercase letter, number, and special symbol) to generate a fallback password with a high entropy value (such as "Xm7#vQ2!pL9w@z"), which is then sent back to the multi-dimensional probability measurement module for classification. This closed loop significantly improves the system's fault tolerance, ensuring that the system can still output secure and compliant results even if the server-side large model crashes or enters an infinite loop.

[0056] The results output module outputs a structured evaluation report based on the final validated enhancement suggestions or rollback password.

[0057] In one preferred implementation, the semantic enhancement suggestion module determines the password strength level based on the mapping relationship between the quantification benchmark value and the preset strength grading rules, and generates an evaluation description according to the preset output template. When the password strength level reaches the preset security judgment level, the semantic enhancement suggestion module stops generating enhancement suggestions, the self-optimization closed-loop control module terminates the verification iteration, and the result output module directly outputs a release prompt that does not require modification.

[0058] In actual project deployment, the system provides lightweight backend service interfaces based on the FastAPI framework and is rendered through the display screens of front-end devices (such as LCD and OLED displays). The output structure of the evaluation report is presented as a multi-level JSON data packet. For passwords that have reached "strong" in the initial evaluation (such as the quantization benchmark value being greater than the preset upper limit), the system performs "computing power bypass optimization"—directly skipping the subsequent large model suggestion generation stage and closed-loop verification stage, and immediately returning a green "no modification required" prompt through the result output module. This design effectively avoids unnecessary cloud-based large model computing power consumption and greatly improves the system throughput and response speed in high-concurrency scenarios.

[0059] For passwords that need optimization (i.e., branches that did not reach the "strong" level in the initial assessment), users are shown a complete in-depth health check report with "color + icon + text" in three dimensions, including specific vulnerability tracing, anti-collision prompts, and a new password modification plan with dual-version mathematical verification.

[0060] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included 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 protection of the described technical solutions.

Claims

1. A password strength evaluation system based on a large model, characterized in that, The system includes: The data preprocessing module performs cleaning, character encoding unification, and format adaptation on the password training data to obtain training samples; The user information parsing module extracts the personal information entered by the user in a structured manner to obtain structured fields related to password security analysis; The multidimensional probability measurement module constructs an UNCM model based on the training samples; and after receiving the password to be evaluated, it calculates the number of guesses for the password to be evaluated under each model, and extracts the minimum value as the quantization benchmark value. The semantic enhancement suggestion module utilizes a large language model, combining the structured fields with the password to be evaluated, to generate an initial candidate set that includes a password strength level description, vulnerability analysis results, and enhancement suggestions. The deterministic logic constraint module establishes a verification matrix and performs joint reverse verification on the enhancement suggestions in the initial candidate set. The verification matrix includes: Security enhancement constraint: The second-order guess number of the new password corresponding to the enhancement suggestion under the multi-dimensional probability measurement module, relative to the quantization benchmark value, exceeds the preset quantization threshold. Semantic deassociation constraint: The statistical correlation between the new password and the structured fields is lower than a preset threshold, so as to reduce the dependence of the new password on the user's personal information features; Structural constraints: The new password must meet the preset length limit and character composition rules; The self-optimization closed-loop control module establishes a feedback link between the semantic enhancement suggestion module and the deterministic logic constraint grid. If the enhancement suggestion fails the joint reverse verification, the quantization deviation of the verification failure is sent back to the large language model as a feedback parameter to generate enhancement suggestions again. If the enhancement suggestion generated after a preset number of iterations still fails the verification, a fallback mechanism is triggered, a preset strong password template is called to generate a fallback password, and the multidimensional probability measurement module performs a second guessing number calculation on the fallback password. The results output module generates an evaluation report based on the validated enhancement suggestions or rollback password.

2. The password strength evaluation system based on a large model according to claim 1, characterized in that: The data preprocessing module performs the following processing on the password training data: Delete password samples whose length is less than the preset lower limit or greater than the preset upper limit, handle abnormal data with special character encoding, and convert password samples to UTF-8 encoding format.

3. The password strength evaluation system based on a large model according to claim 1, characterized in that: When the data preprocessing module performs format adaptation for the UNCM model, it maps the password training data into a four-tuple feature sequence. The four-tuple feature sequence includes the email username, email service provider, top-level domain, and password field in sequence, and the email service provider and top-level domain are taken from a preset specific domain dictionary.

4. The password strength evaluation system based on a large model according to claim 1, characterized in that: The Markov password probability model is a probability model built based on the nth-order character transition probability; The multidimensional probability measurement module determines the character transition probability based on the frequency of character combinations in the training samples, and corrects the weights of zero-probability transition terms through a smoothing algorithm.

5. The password strength evaluation system based on a large model according to claim 1, characterized in that: The multidimensional probability measurement module generates a set of candidate passwords based on the Markov password probability model and the UNCM model, and sorts the set of candidate passwords in descending order according to the generation probability of each candidate password. The absolute position index of the password to be evaluated or the new password in the sorting result represents the number of guesses under the corresponding model.

6. The password strength evaluation system based on a large model according to claim 1, characterized in that: The structured fields include at least one of the following: name, gender, date of birth, mobile phone number, email address, and account name; The semantic enhancement suggestion module generates personal information association analysis results based on the cross-comparison and fuzzy matching relationship between the structured fields and the password to be evaluated.

7. The password strength evaluation system according to claim 1, characterized in that, The enhancement suggestions include an optimized version of the original password and a standard enhanced version. The optimized version of the original password is generated based on the password to be evaluated by character insertion, structural reorganization or adaptive length expansion, so as to improve the password complexity while preserving the user's memory anchor. Among them, the character homomorphic substitution strategy is restricted or suppressed during the generation process to avoid introducing weak patterns that can be exploited by the rule-based guessing model. The standard enhanced version meets the preset length requirements and is generated using a template that includes at least two non-contiguous semantic English words, random numbers, and non-letter separators; The multidimensional probability measurement module performs parallel secondary guessing calculations on the optimized version and the standard enhanced version of the original password that have passed the joint reverse verification.

8. The password strength evaluation system according to claim 1, characterized in that, The semantic enhancement suggestion module determines the password strength level based on the mapping relationship between the quantification benchmark value and the preset strength grading rules, and generates an evaluation description according to the preset output template; When the password strength level reaches the preset security judgment level, the semantic enhancement suggestion module stops generating enhancement suggestions, the self-optimization closed-loop control module terminates the verification iteration, and the result output module directly outputs a release prompt that does not require modification.

Images