Directional password guessing method and system based on large language model

Abstract

The invention relates to the technical field of electric data processing, in particular to a directional password guessing method and system based on a large language model, and the method comprises the following steps: obtaining the personal information and / or old password of a user; rewriting the personal information and / or the old password according to a set format; and inputting the rewritten personal information and / or the old password into the trained glm-4-9b model, and outputting to obtain a guess password. According to the method, the large language model is applied to a directional password guessing scene, a highly targeted password is generated by using personal information and an old password in an end-to-end manner, and an additional dictionary database is not required to be introduced and complex probability calculation is not required to be designed like personal information tags set by experience in a conventional probability model; the method can help to quickly recover the password hash file, reveal the potential security problem of the user password, and help to provide reference for the password security policy of a network service party.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic data processing, and in particular to a directional password guessing method and system based on a large language model. Background Art

[0002] In order to protect the privacy of users, users are usually required to set passwords. Password guessing refers to guessing the user's password through technical means. Password guessing has positive significance. For example, it can detect the strength of the user's password to prompt the user to set a more secure password. Therefore, password guessing has always been a topic that scholars have worked hard to study.

[0003] For example, the Chinese invention with application number 202210073083.1 discloses a directional password guessing method called "a directional online password guessing method based on a pointer generation network". Although it can use personal information to generate targeted guessed passwords, the use of personal information is still based on personal information labels set by human experience, and lacks effective generalization in real scenarios. Summary of the invention

[0004] The purpose of the present invention is to provide a directional password guessing method and system based on a large language model, so as to improve the problem that password guessing based on personal information tags in the prior art has certain limitations.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides a directional password guessing method based on a large language model, comprising the following steps:

[0007] Obtaining users’ personal information and / or old passwords;

[0008] Rewrite the personal information and / or old password according to the set format;

[0009] The rewritten personal information and / or old password is input into the trained glm-4-9b model, and the guessed password is output.

[0010] In a more optimized solution, the setting format is alpaca format.

[0011] In the above scheme, the original data set is rewritten in alpaca format, which not only makes the rewritten original data set more conducive to the learning of the glm-4-9b model and improves its recognition ability, but also the alpaca format is widely used and easy to implement.

[0012] In one possible implementation scheme, the glm-4-9b model is obtained by training through the following steps:

[0013] Obtain a dataset of personal information and password data that is publicly leaked on the Internet and contains email information, and clean it to remove incorrectly formatted data entries;

[0014] Using the email address as an identifier, match the personal information dataset and password dataset of the same user, and associate and match the matched personal information dataset and password dataset to construct the original training dataset;

[0015] Rewrite the original training data set using the alpaca format to obtain a training data set;

[0016] The training data set is used to train the glm-4-9b model.

[0017] In the above scheme, personal information and password information including email addresses are collected first. Since email addresses are easier to obtain, the email address is used as an identifier to associate and match the personal information and password data of the same user to help the GLM-4-9B model learn. Rewriting the data set in the Alpaca format is more helpful for the GLM-4-9B model to recognize passwords and improve the accuracy of the model's password guessing results. Therefore, the GLM-4-9B model trained by the above method has strong recognition ability, is suitable for password guessing, and has high accuracy.

[0018] In one possible implementation scheme, the removal of incorrectly formatted data entries includes: removing entries with incorrect email addresses and password formats; removing entries containing characters exceeding printable ASCII codes; and removing duplicate entries.

[0019] In the above scheme, by removing the above three types of data entries, not only can the erroneous data be reduced, but also duplicate data can be avoided, data redundancy can be reduced, and it is helpful for the large language model to learn correctly.

[0020] The process of using the mailbox as an identifier to match the personal information dataset and password dataset of the same user, and associating and matching the matched personal information dataset and password dataset, includes:

[0021] First, based on personal information, match through email addresses to complete the element information of the personal information dataset as much as possible. The element information includes: name, birthday, phone number, user name, ID number, and email address;

[0022] Next, based on the password dataset, matching is performed by email address, trying to match multiple different passwords of the same user in the password dataset;

[0023] Finally, based on the matched personal information dataset and password dataset, the personal information and historical passwords of the same user are associated and matched through the email matching method to obtain the original training dataset. A complete original training data set contains: {personal information: (name, birthday, phone number, user name, ID number, email address), historical passwords: (old password 1, old password 2, old password 3, ...)}.

[0024] The alpaca format is used to rewrite the original training data set, including: the instruction field guides the LLM to understand the directional password guessing task through natural language description; the input field is set to the user's personal identifiable information and an old password, and the name of the Chinese user is expressed in the form of Chinese pinyin, with the surname and the first name separated by a space; the output field randomly selects an old password from the historical password data in the original training data set as the target password.

[0025] In a second aspect, an embodiment of the present invention provides a directional password guessing system based on a large language model, comprising:

[0026] Data collection module, used to obtain users' personal information and / or old passwords;

[0027] A data rewriting module, used to rewrite the personal information and / or old password according to a set format;

[0028] Model training module, used to train the glm-4-9b model using the training data set;

[0029] The password guessing module is used to input the rewritten personal information and / or the old password into the trained glm-4-9b model and output the guessed password.

[0030] In a third aspect, the present invention provides a computer program product, comprising computer-readable instructions, characterized in that the computer-readable instructions, when executed by a processor, implement the steps of the directional password guessing method based on a large language model of the present invention.

[0031] In a fourth aspect, the present invention provides a computer-readable storage medium comprising computer-readable instructions, characterized in that the computer-readable instructions, when executed by a processor, implement the steps of the directional password guessing method based on a large language model of the present invention.

[0032] In a fifth aspect, the present invention provides an electronic device, comprising: a memory storing program instructions; a processor connected to the memory, executing the program instructions in the memory, and implementing the steps in the directional password guessing method based on a large language model of the present invention.

[0033] Compared with the prior art, the present invention has the following technical advantages:

[0034] By applying a large language model to targeted password guessing scenarios, personal information and old passwords are used in an end-to-end manner to generate highly targeted passwords. This eliminates the need for empirically set personal information labels, the introduction of additional dictionary libraries, and the design of complex probability calculations as in previous probabilistic models. This can help quickly recover password hash files, reveal potential security issues with user passwords, and provide a reference for password security policies of network service providers.

[0035] Directed password guessing can generally be divided into directed guessing based on personal information and directed guessing based on old passwords. The former requires a manually set personal information structure, while the latter requires editing and transformation. The present invention obtains the target password directly from personal information and old passwords, does not require setting personal information tags, and does not require editing and transformation. It is highly targeted and adaptable, and has high efficiency.

[0036] By fine-tuning a smaller-scale large language model and applying it to the downstream task of targeted password guessing, we can generate highly targeted guesses with the help of the large language model's prior knowledge, excellent natural language understanding and generation capabilities. This breaks through the limitations of previous targeted password guessing models that rely on experience to set fixed types of personal information labels and use external dictionaries to assist in generating targeted guesses, resulting in insufficient generalization. This solves the problem that previous targeted password guessing methods designed guessing methods based on personal information and old passwords separately, making it difficult to consider both personal information and old passwords at the same time when generating guesses.

[0037] For other advantages of the present invention, please refer to the relevant description in the embodiment section. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 The flowchart of the directional password guessing method based on the large language model is given as an example in the embodiment of the present invention.

[0040] Figure 2 It is a training and reasoning flow chart of the glm-4-9b model in an embodiment of the present invention.

[0041] Figure 3 The figure is a block diagram of a directional password guessing system based on a large language model in an embodiment of the present invention.

[0042] Figure 4A block diagram of the components of an electronic device. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0044] See also Figure 1 In this embodiment, a directional password guessing method based on a large language model is provided, comprising the following steps:

[0045] S10, obtaining a personal information dataset containing email information and a password data dataset that are publicly leaked on the Internet (not actively disclosed by the user but leaked on the Internet, which can also be understood as being available to the public on the Internet) as the original dataset, and cleaning the original dataset to remove incorrectly formatted data entries. Data cleaning mainly includes: 1) removing entries with incorrect email and password formats, 2) removing entries containing characters exceeding printable ASCII codes, and 3) removing duplicate entries.

[0046] Since the publicly leaked data sets in the real network are incomplete, some data sets have missing personal information, and some have only one password. These publicly leaked data sets usually contain email addresses. Therefore, this article uses email addresses as user identifiers and uses email addresses to bring together the information of the same user scattered in different leaked data sets in order to complete personal information and match multiple passwords of the same user. Therefore, when collecting raw data, the collected personal information and password data need to include email address information.

[0047] S20, using the email address as an identifier, matching the personal information dataset and the password dataset of the same user, and associating and matching the matched personal information dataset and password dataset to construct an original training dataset.

[0048] When this step is specifically executed, first, based on personal information, match through email addresses to complete the element information of the personal information data set as much as possible, mainly including 6 elements: name, birthday, phone number, user name, ID number, and email address. For example, the personal information of user u is scattered in two data sets A and B. The personal information in data set A includes email address u, name u, birthday u, and the personal information in data set B includes email address u, phone number u, and user name u. By matching the email address u in the two data sets, u's personal information can be completed, that is, email address u, name u, birthday u, phone number u, and user name u.

[0049] Next, based on the password dataset, matching is performed by email address, trying to match multiple different passwords of the same user in the password dataset as much as possible.

[0050] Finally, based on the above matched personal information dataset and password dataset, the personal information and historical passwords of the same user are associated and matched through the email matching method to obtain the final original training dataset.

[0051] A complete piece of original training data contains: {personal information: (name, birthday, phone number, username, ID number, email address), historical passwords: (old password 1, old password 2, old password 3, ...)}.

[0052] If the entire data processing flow is divided into multiple stages, the stages in which the above steps S10 and S20 are located can be collectively referred to as the data preprocessing stage, the subsequent steps S30 and S40 can be collectively referred to as the model fine-tuning stage, and step S50 can be referred to as the generation stage.

[0053] S30, based on the original training data set, rewrite it using the alpaca format to obtain the training data set.

[0054] Specifically, the alpaca format includes three fields: instruction, input, and output. The instruction field guides the LLM (Large Language Model) to understand the directional password guessing task through natural language description. The input field is set to the PII (Personally Identifiable Information) of a specific user and an old password. For Chinese users, the name is expressed in Chinese pinyin, with a space between the surname and the first name. The output field, that is, the label of supervised learning, randomly selects an old password from the historical password data in the original training data set as the target password. In the end, each training data set is in the shape of: Instruction: "Using someone's personally identifiable information and old passwords, guess the most likely new password for this person." Input: Name is "zhang san", birthday is 19xx0x0x, phone number is13012345678, id number is443068219xx0x0xxxxx, email is scl_abccdef@163.com,username is abcdef,oldpasswords are scl123456,scl_abcdefo.output:sclabcdef.

[0055] In this step, the original training data set is fine-tuned using the alpaca format to construct a training data set, so that when the LLM is trained based on the training data set, it can have stronger recognition capabilities in password recognition application scenarios in a more targeted manner.

[0056] S40, training the glm-4-9b model using the training data set. Because the training process is also a process of fine-tuning the structure and parameters of the glm-4-9b model, training the glm-4-9b model can also be called fine-tuning the glm-4-9b model.

[0057] The pre-trained glm-4-9b model is used as the baseline model for fine-tuning, and autoregressive blank filling is used as the training objective. During the entire fine-tuning process, the glm-4-9b model learns how to adjust the user's PII and the old password to obtain the user's new password.

[0058] Specifically, given an input text X = [x1,...,x n ], sampling multiple text fragments {s1,..,s m}, each fragment s i corresponds to a series of consecutive marks in X Each fragment is replaced by a single [MASK] token, forming the broken text X corrupt The model predicts the missing tokens in these segments autoregressively, which means that when predicting the missing tokens in a segment, the model has access to the corrupted text and previously predicted segments. In order to fully capture the interdependencies between different segments, the order of the segments is randomly permuted, similar to a ranking language model. Formally, let Z m is the set of all possible permutations [1,2,...,m] of index sequences of length m, s Z＜i for GLM defines the pre-training objectives as follows:

[0059]

[0060] Generate the tokens in each blank in order from left to right, that is, generate fragment s i The probability of is decomposed into:

[0061]

[0062] GLM training randomly removes some consecutive words from the text, and then trains the model to restore these words one by one in a certain order. This method combines the advantages of autoencoding and autoregression pre-training methods, and shows excellent performance in natural language understanding and text generation tasks, and has excellent generalization ability for different downstream tasks.

[0063] See also Figure 2 , the specific training steps are as follows:

[0064] 1) Loading training data set and pre-trained model: In the first step of the training process, the training data set and pre-trained model required for the target task are loaded. The training data set includes the user's personal information (PII) and historical passwords, which can be recorded as Where P i represents the password of the i-th user, I i Represents the personal information related to the password. In order to ensure the diversity and representativeness of the data, each instance in the training data set (P i ,I i ) are all screened and cleaned. Next, load the glm-4-9b model, denoted as M pretrainedThe model is pre-trained on a large-scale corpus, has strong generative capabilities, and has learned a large number of language rules. By loading the model's pre-trained parameters θ pretrained , providing a basis for subsequent fine-tuning.

[0065] 2) Before entering the training, the input data needs to be preprocessed, that is, the text data (including passwords and personal information) is converted into vector representation through the glm-4-9b tokenizer. Assume that the input text is a sequence X consisting of multiple words = [x1, x2, ..., x n ], the tokenizer converts it into a fixed-dimensional vector sequence V = [v1, v2, ..., v n ], where each v i is a high-dimensional embedding vector. In this process, the tokenizer assigns a unique identifier to each word or character and maps it to the embedding space. In this way, the text data can be understood by the model and processed in subsequent training.

[0066] 3) In order to speed up the training process and reduce memory overhead, the LoRA method is used here to fine-tune the model. Specifically, LoRA reduces the amount of parameter updates by introducing low-rank matrix decomposition in some weight matrices of the model. For each layer weight matrix W of glm-4-9b, LoRA introduces low-rank matrices A and B (where r<<d, r and d are the ranks of the low-rank matrices A and B respectively), and it can be expressed as:

[0067] W′=W+ΔW

[0068] Among them, ΔW represents the adaptive update introduced by LoRA during fine-tuning. Through this low-rank update method, the number of model parameters and memory consumption are significantly reduced while maintaining the good performance of the model.

[0069] 4) Train the model and update parameters

[0070] During the training phase, the goal of the model is to minimize the loss function To optimize the parameters. Assume that in each step of training, the input is the vectorized data V and the output is the password predicted by the model The model updates its parameters by calculating the difference between the predicted password and the true password. The loss function can be expressed as:

[0071]

[0072] Among them, f(M θ (V i )) represents the prediction result of the model under parameter θ, Pi is the real password, Loss is the cross entropy loss, and N is the number of samples. The gradient is calculated by the back propagation algorithm based on the loss value, and the AdamW optimizer is used to determine the amplitude of the parameter update. Due to the use of the LoRA method, parameter updates only occur on low-rank matrices A and B, and other weight matrices W remain unchanged, thereby reducing computational and memory overhead.

[0073] 5) Combine fine-tuning parameters with pre-training parameters

[0074] After training is completed, the fine-tuned LoRA parameters ΔW=AB will be merged with the pre-trained parameters W of the glm-4-9b model to obtain the final model weight W final :

[0075] W final =W+ΔW

[0076] The merging operation ensures that the knowledge gained during the fine-tuning process is effectively integrated into the model, allowing the model to fully demonstrate its capabilities in password guessing tasks while retaining a certain degree of original language understanding and generation capabilities.

[0077] 6) Finally, save the trained glm-4-9b model M final , the model includes the combined weights W final and other necessary model parameters. The saving process persists the state of the model for subsequent use or deployment. This step ensures that the model can be reused for inference tasks or further training.

[0078] During the training process, the training hyperparameters are set as follows: 1 epoch in total, batch_size is 8, gradient_accumulation is 32, torch_dtype is torch.bfloat16, AdamW optimizer is used, and the learning rate is 5e-4. The parameters of LoRA are: lora_r is 8, lora_alpha is 32, and lora_dropout is 0.1. By setting these parameters, the accuracy of the model in guessing passwords can be improved.

[0079] S50, collect personal information and / or old passwords of target users (target users refer to users whose passwords are to be guessed), rewrite the collected data in Alpaca format, and then input the rewritten data into the trained GLM-4-9B model to output the guessed password.

[0080] Specifically, load the fine-tuned glm-4-9b model. The model input cannot be empty. Input the directional information in alpaca format. If the target user's personal information is known, enter the target user's personal information. The personal information can be a complete set of 6 elements or a default one. If the user's old password is known, enter the user's old password. The old passwords can be multiple. If both the personal information and the old password are known, all of them are used as input.

[0081] That is to say, the above method can guess the user's new password based on personal information and / or historical passwords. It is not limited to personal information and can also be historical passwords. Therefore, it can be adapted to more application scenarios.

[0082] See also Figure 2 Considering the low-tolerance natural language task of password guessing, this step uses the beam search method to generate multiple targeted candidate passwords, namely P(pw geuss |PII,pw old ) The top q guesses with the highest probability are selected. Specifically, the q paths with the highest probability generated by the model (i.e., beam width, which represents the guessing budget q in the password guessing task) are selected, and these paths are input iteratively until an end marker is encountered, indicating that a guess has been generated. The total number of iterations does not exceed the search depth. Finally, q candidate passwords are obtained, which are sorted in descending order of probability.

[0083] In order to speed up the generation of passwords, the traditional Attention calculation is replaced by Paged Attention. Paged Attention draws on the idea of ​​managing virtual memory through Page in the operating system: KVCache (used to store the results of previous calculations to reduce repeated calculations and improve the calculation efficiency of Attention) is divided into fixed-size blocks. These blocks do not need to be stored in continuous memory and are managed by a unified memory allocator. Requests apply for memory on demand, and there is no need to reserve memory of max_seq_len size in advance. After the request is completed, its own blocks are released, and the system only needs to allocate small blocks. Paged Attention can share cache blocks. In the beam search scenario, multiple results need to be generated for the same prompt. These sub-requests can share the same batch of prompt caches. Page Attention can increase the throughput of beam search by more than 10 times, thereby increasing the speed of password generation.

[0084] The above method fine-tunes a smaller-scale large language model and applies it to the downstream task of targeted password guessing. It uses the prior knowledge, excellent natural language understanding and generation capabilities of the large language model to generate highly targeted guesses. It breaks through the limitations of previous targeted password guessing models that rely on experience to set fixed types of personal information labels and use external dictionaries to assist in generating targeted guesses, resulting in insufficient generalization. It also solves the problem that previous targeted password guessing methods designed guessing methods based on personal information and old passwords separately, making it difficult to consider both personal information and old passwords at the same time when generating guesses.

[0085] See also Figure 3 In this embodiment, a directional password guessing system based on a large language model is also provided, including:

[0086] Data collection module, used to obtain users' personal information and / or old passwords;

[0087] A data rewriting module, used to rewrite the personal information and / or old password according to a set format;

[0088] Model training module, used to train the glm-4-9b model using the training data set;

[0089] The password guessing module is used to input the rewritten personal information and / or the old password into the trained glm-4-9b model and output the guessed password.

[0090] For more detailed processing operations of the above modules, please refer to the relevant description in the above method, which will not be repeated here.

[0091] As shown in FIG. 4 , this embodiment also provides an electronic device, which may include a processor 41 and a memory 42, wherein the memory 42 is coupled to the processor 41. It is worth noting that this figure is exemplary, and other types of structures may be used to supplement or replace this structure to achieve data extraction, report generation, communication or other functions.

[0092] like Figure 4 As shown, the electronic device may further include: an input unit 43, a display unit 44 and a power supply 45. It is worth noting that the electronic device does not necessarily have to include Figure 4 In addition, electronic devices may also include Figure 4 For components not shown, reference may be made to the prior art.

[0093] The processor 41 is sometimes also called a controller or an operation control, and may include a microprocessor or other processor devices and / or logic devices. The processor 41 receives inputs and controls the operations of various components of the electronic device.

[0094] The memory 42 may be, for example, one or more of a cache, a flash memory, a hard drive, a removable medium, a volatile memory, a non-volatile memory or other suitable devices, and may store information such as configuration information of the processor 41 and instructions executed by the processor 41. The processor 41 may execute the program stored in the memory 42 to implement information storage or processing. In one embodiment, the memory 42 also includes a buffer memory, i.e., a buffer, to store intermediate information.

[0095] An embodiment of the present invention further provides a computer program product, including computer-readable instructions. When the computer-readable instructions are executed in an electronic device, the program product enables the electronic device to execute the operation steps included in the method of the present invention.

[0096] An embodiment of the present invention further provides a storage medium storing computer-readable instructions, wherein the computer-readable instructions enable an electronic device to execute the operation steps included in the method of the present invention.

[0097] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0098] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0099] The above-described embodiments are only specific implementations of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications, replacements and improvements within the technical scope disclosed by the present invention, and these modifications, replacements and improvements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A directional password guessing method based on a large language model, characterized in that: The following steps are involved: Obtaining users’ personal information and / or old passwords; Rewrite the personal information and / or old password according to the set format; The rewritten personal information and / or old password is input into the trained glm-4-9b model, and the guessed password is output.

2. The directional password guessing method based on a large language model according to claim 1, characterized in that: The setting format is alpaca format.

3. The directional password guessing method based on a large language model according to claim 1, characterized in that: The glm-4-9b model is trained by the following steps: Obtain a dataset of personal information and password data that is publicly leaked on the Internet and contains email information, and clean it to remove incorrectly formatted data entries; Using the email address as an identifier, match the personal information dataset and password dataset of the same user, and associate and match the matched personal information dataset and password dataset to construct the original training dataset; Rewrite the original training data set using the alpaca format to obtain a training data set; The training data set is used to train the glm-4-9b model.

4. The directional password guessing method based on a large language model according to claim 3 is characterized in that: The removing of incorrectly formatted data entries includes: removing entries with incorrect email and password formats; removing entries containing characters exceeding printable ASCII codes; and removing duplicate entries.

5. The directional password guessing method based on a large language model according to claim 3 is characterized in that: The process of using the mailbox as an identifier to match the personal information dataset and password dataset of the same user, and associating and matching the matched personal information dataset and password dataset, includes: First, based on personal information, match through email addresses to complete the element information of the personal information data set as much as possible. The element information includes: name, birthday, phone number, user name, ID number, and email address; Next, based on the password dataset, matching is performed by email address, trying to match multiple different passwords of the same user in the password dataset; Finally, based on the matched personal information dataset and password dataset, the personal information and historical passwords of the same user are associated and matched through the email matching method to obtain the original training dataset. A complete original training data set contains: {personal information: (name, birthday, phone number, user name, ID number, email address), historical passwords: (old password 1, old password 2, old password 3, ...)}.

6. The directional password guessing method based on a large language model according to claim 3 is characterized in that: The alpaca format is used to rewrite the original training data set, including: the instruction field guides the LLM to understand the directional password guessing task through natural language description; the input field is set to the user's personal identifiable information and an old password, and the name of the Chinese user is expressed in the form of Chinese pinyin, with the surname and the first name separated by a space; the output field randomly selects an old password from the historical password data in the original training data set as the target password.

7. A directional password guessing system based on a large language model, characterized in that: include: Data collection module, used to obtain users' personal information and / or old passwords; A data rewriting module, used to rewrite the personal information and / or old password according to a set format; Model training module, used to train the glm-4-9b model using the training data set; The password guessing module is used to input the rewritten personal information and / or the old password into the trained glm-4-9b model and output the guessed password.

8. A computer program product comprising computer readable instructions, characterized in that: When the computer-readable instructions are executed by a processor, the steps in the directional password guessing method based on a large language model described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium comprising computer-readable instructions, characterized in that: When the computer-readable instructions are executed by a processor, the steps in the directional password guessing method based on a large language model described in any one of claims 1 to 6 are implemented.

10. An electronic device, characterized in that: include: Memory, which stores program instructions; A processor is connected to the memory and executes program instructions in the memory to implement the steps in the directional password guessing method based on a large language model as described in any one of claims 1 to 6.

Images