A hierarchical and decentralized management method for enterprise data based on large language model RAG

Abstract

The present invention discloses a method for hierarchical and decentralized management of enterprise data based on a large language model (RAG), which relates to the technical field of data security and access control. The method comprises the following steps: obtaining a voice audio stream and a lip shape three-dimensional trajectory point set, loading permission baseline parameters, and generating a permission token, wherein the permission token comprises a quantum watermark seed, a role identifier, a permission token expiration stamp, and a permission level; obtaining permission data blocks through a retrieval enhancement generation framework based on the permission level and role identifier, dividing the hierarchies according to the permission baseline parameters, generating retrieval keywords based on the permission token, establishing a hierarchical index, matching permission data blocks, reordering the permission data blocks after matching the role identifiers, generating a response text, converting the quantum watermark seed into a binary stream, and embedding the stream into the response text to obtain a watermark response text; the present invention solves the problem of lack of dynamic risk perception in permission decision-making through multimodal biometric fusion and Shannon entropy value calculation.

Description

Technical Field

[0001] The present invention relates to the technical field of data security and access control, and in particular to a method for hierarchical and decentralized management of enterprise data based on a large language model (RAG). Background Art

[0002] As enterprises accelerate their digital transformation, the value of data assets is becoming increasingly prominent. Enterprises' demands for data security and permissions management are gradually shifting from simple, static controls to dynamic, intelligent, and refined governance. Enterprise data management primarily relies on technical frameworks such as role-based access control or attribute-based encryption (ABE), enabling hierarchical and categorized access to data through pre-defined permission rules. With the rise of large language models and search-enhanced generative technologies, enterprises are exploring the integration of external knowledge bases with generative models to improve data query efficiency and accuracy.

[0003] However, existing RAG applications mostly focus on knowledge retrieval and generation optimization, and few studies have deeply integrated it with dynamic permission management. As a result, the security boundaries of data access still rely on traditional static rules. There are two core pain points in permission management that need to be overcome: First, permission decisions lack dynamic risk perception. Existing methods usually grant fixed permissions based on user identity or role, ignoring the semantic relevance and risk dynamics of data access behavior. Second, anti-counterfeiting and traceability capabilities are weak. Although existing methods achieve anti-tampering protection through digital signatures or watermarking technology, their ability to combat new attack methods such as deep fakes is limited. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method for hierarchical and decentralized management of enterprise data based on a large language model RAG to solve the problems of lack of dynamic risk perception and weak anti-counterfeiting ability in authority decision-making.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for hierarchical and decentralized management of enterprise data based on a large language model RAG, which comprises:

[0008] Acquire the voice audio stream and lip shape three-dimensional trajectory point set, load the permission baseline parameters, and generate a permission token, which includes a quantum watermark seed, a role identifier, a permission token expiration stamp, and a permission level;

[0009] Based on the permission level and role identification, the permission data block is obtained through the retrieval enhancement generation framework. The hierarchy is divided according to the permission baseline parameters. The retrieval keyword is generated based on the permission token, and a hierarchical index is established to match the permission data block. The permission data block is reordered after role identification matching, and a response text is generated. The quantum watermark seed is converted into a binary stream and embedded into the response text to obtain the watermark response text.

[0010] Construct a causal graph model, calculate the abnormal probability of illegal data access, trigger deep fake identification based on the abnormal probability of illegal data access, generate the fake probability, mark the risk type, and record the abnormal detection report;

[0011] Update permission baseline parameters based on anomaly detection reports.

[0012] As a preferred solution of the enterprise data hierarchical and decentralized management method based on the large language model RAG of the present invention, wherein: the steps of obtaining the voice audio stream and the lip shape three-dimensional trajectory point set, loading the authority baseline parameters, and generating the authority token are as follows:

[0013] Extract the voiceprint embedding vector and motion vector from the speech audio stream and lip shape three-dimensional trajectory point set, fuse them into a fused feature vector, and calculate the Shannon entropy value of the fused feature vector;

[0014] According to the Shannon entropy value of the fused feature vector and the voiceprint embedding vector, a linear transformation layer of the attention mechanism is used to generate the voiceprint query vector and the voiceprint key vector;

[0015] Calculate the dot product of the voiceprint query vector and the voiceprint key vector to obtain the similarity score between the voiceprint query vector and the voiceprint key vector. Use the Softmax function to normalize the similarity score between the voiceprint query vector and the voiceprint key vector to generate the voiceprint credibility weight.

[0016] According to the Shannon entropy value of the fused feature vector and the motion vector, a linear transformation layer of the attention mechanism is used to generate a motion query vector and a motion key vector. The similarity score between the motion query vector and the motion key vector is normalized using the Sigmoid function to generate a semantic risk weight.

[0017] The dynamic permission value is calculated using the Sigmoid function based on the Shannon entropy value of the fused feature vector, the voiceprint credibility weight, and the semantic risk weight.

[0018] Load permission baseline parameters;

[0019] The authority baseline parameters include authority level mapping, weight baseline value, voiceprint similarity threshold, lip movement synchronization threshold, cross-level access count threshold and quantum watermark generation key;

[0020] Generates a permission token based on the dynamic permission value and permission baseline parameters.

[0021] As a preferred solution of the enterprise data hierarchical and decentralized management method based on the large language model RAG of the present invention, wherein: the generating of the authority token according to the dynamic authority value and the authority baseline parameter comprises the following steps: generating a quantum watermark seed using a chaotic mapping algorithm based on the dynamic authority value and the authority baseline parameter;

[0022] According to the permission baseline parameters, the dynamic permission value is converted into the permission level and permission token expiration stamp;

[0023] Encode the quantum watermark seed, permission level and permission token expiration stamp into the permission token.

[0024] As a preferred solution of the enterprise data layered and decentralized management method based on the large language model RAG of the present invention, the specific steps of constructing the causal graph model and calculating the abnormal probability of illegal data access are as follows:

[0025] Define permission token status nodes, abnormal access result nodes, data access mode nodes, and deep fake feature nodes;

[0026] Establish directed edges from the permission token status node to the abnormal access result node, the data access pattern node to the abnormal access result node, and the deep fake feature node to the permission token status node;

[0027] Generate binary codes, unit time frequencies, and normalized deviation values ​​based on the data block access frequency and cross-level access counts in the data access pattern node, deep fake feature nodes, and permission tokens;

[0028] Generate a conditional probability table based on binary coding, unit time frequency and normalized deviation value;

[0029] According to the conditional probability table, the abnormal probability value of illegal data access is calculated using the Bayesian formula.

[0030] As a preferred solution of the enterprise data layered and decentralized management method based on the large language model RAG of the present invention, the specific steps of generating the forgery probability are as follows:

[0031] Based on the voiceprint embedding vector, a sliding window is used to generate the voiceprint forgery probability value;

[0032] According to the speech audio stream and the lip shape three-dimensional trajectory point set, the dynamic time warping algorithm is used to generate the lip movement abnormality probability value.

[0033] As a preferred solution of the enterprise data hierarchical and decentralized management method based on the large language model RAG of the present invention, the risk type marking includes: marking voiceprint forgery attack, lip movement forgery attack and composite forgery attack according to the voiceprint forgery probability value and lip movement abnormality probability value; marking authority token abuse attack based on the number of cross-level accesses in the data access pattern node; marking watermark tampering attack based on the quantum watermark seed;

[0034] The recording of the anomaly detection report refers to recording the abnormal probability value of illegal data access, the voiceprint forgery probability value, the lip movement abnormal probability value, the risk type, the attack timestamp and the quantum watermark seed of the authority token as an anomaly detection report.

[0035] As a preferred solution of the enterprise data hierarchical and decentralized management method based on the large language model RAG of the present invention, wherein: the updating of the authority baseline parameter based on the anomaly detection report comprises the following steps: statistically analyzing the attack frequency of the risk type based on the anomaly detection report;

[0036] Adjust the cross-level access count threshold and voiceprint similarity threshold in the permission baseline parameters based on the attack frequency of the risk type;

[0037] Based on the anomaly detection report, the number of consecutive triggering of watermark tampering attacks is counted, and the quantum watermark generation key in the permission baseline parameters is replaced.

[0038] As a preferred solution of the enterprise data hierarchical and decentralized management method based on the large language model RAG described in the present invention, wherein: the reordering of the permission data blocks after role identification matching refers to defining priority sorting rules, and using the priority sorting rules to reorder the permission data blocks after role identification matching.

[0039] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the method for hierarchical and decentralized management of enterprise data based on the large language model RAG as described in the first aspect of the present invention.

[0040] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for hierarchical and decentralized management of enterprise data based on the large language model RAG as described in the first aspect of the present invention.

[0041] The beneficial effects of the present invention are as follows: through multimodal biometric fusion and Shannon entropy calculation, the uncertainty of user behavior characteristics can be captured in real time, the problem of lack of dynamic risk perception in permission decision-making is solved, the ability to identify disguised attacks is improved, the uniqueness and unpredictability of permission tokens are ensured, and the anti-counterfeiting ability is improved; heterogeneous data such as permission status, access behavior, and biometric anomalies are incorporated into a unified causal graph model, which solves the problem of isolated analysis of risk factors, realizes accurate modeling of complex attacks, and can distinguish deep fake attacks. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. 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 any creative work.

[0043] Figure 1 The figure is a flowchart of the hierarchical and decentralized management method of enterprise data based on the large language model RAG.

[0044] Figure 2 Schematic diagram of the authorization token generation process.

[0045] Figure 3 Schematic diagram of causal graph model construction and abnormal probability calculation of illegal data access.

[0046] Figure 4 Schematic diagram of permission data block reordering and watermark embedding. DETAILED DESCRIPTION

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0048] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0049] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0050] Reference Figures 1 to 4, as one embodiment of the present invention, provides a method for hierarchical and decentralized management of enterprise data based on a large language model RAG, comprising the following steps:

[0051] S1: Obtain the voice audio stream and lip shape three-dimensional trajectory point set, load the permission baseline parameters, and generate the permission token. The permission token includes the quantum watermark seed, role identifier, permission token expiration stamp, and permission level.

[0052] The specific steps are as follows:

[0053] A microphone and depth camera controlled by synchronous pulse signals are used to obtain time-aligned original sound data and lip three-dimensional point cloud sequences. The effective speech segments of the original sound data are detected by short-time energy and zero-crossing rate, and spectral subtraction is used to eliminate the ambient noise of the original sound data within the effective speech segments to obtain a speech audio stream. Radius filtering is used to remove outliers in the lip three-dimensional point cloud sequence. The lip key points are defined based on the anatomical features of the lips and the facial action coding unit standard, and the three-dimensional motion trajectory of the lip key points is tracked to generate a lip shape three-dimensional trajectory point set.

[0054] The speech audio stream is input into the voiceprint recognition model, and the speech audio stream is framed according to fixed frame length and frame shift. The Hamming window is used to suppress spectral leakage. The short-time Fourier transform is performed on each frame of the speech audio stream to obtain the spectrum. The environmental noise is eliminated through noise estimation and spectral subtraction. The denoised spectrum is subjected to Mel filter bank and cepstral analysis to extract Mel-frequency cepstral coefficient features. The Mel-frequency cepstral coefficient features are aggregated using convolutional layers and attention pooling layers to obtain the voiceprint embedding vector; the three-dimensional coordinate difference between two adjacent frames of each lip key point is calculated to obtain the instantaneous displacement of the lip movement, and the instantaneous displacement of the lip movement is spliced ​​into a continuous numerical sequence to form an initial motion vector. The initial motion vectors of adjacent time windows are superimposed to expand the temporal representation capability of the initial motion vector. The fully connected layer of the neural network model is used to perform linear transformation and nonlinear activation on the expanded initial motion vector to obtain a motion vector; the voiceprint embedding vector and the motion vector are spliced ​​into a joint vector, and the joint vector is linearly transformed to generate a fusion feature vector.

[0055] The fused feature vector is divided into intervals using the equal-width interval partitioning method. The frequency of occurrence of all dimension values ​​of the fused feature vector in each equal-width interval is counted to form an interval probability distribution. The Shannon entropy value of the fused feature vector is calculated based on the interval probability. The expression is as follows:

[0056] ;

[0057] in, is the Shannon entropy of the fused feature vector, is the total number of equal-width intervals, is an equal-width interval index, For the The probability of an interval of equal width.

[0058] It should also be noted that the fusion of voiceprint embedding vectors and motion vectors, along with Shannon entropy calculation, enables a quantitative assessment of the dynamic complexity of biometric features. By superimposing the time-series expansion of the difference between the three-dimensional lip trajectory point set and the three-dimensional coordinates of the voice audio stream, the motion vector accurately represents the correlation between the user's physiological characteristics and behavioral patterns. Linearly transforming the concatenated joint vector of the voiceprint embedding vector and the motion vector creates a fused feature vector that not only preserves the semantic characteristics of the speech content but also captures the spatiotemporal consistency between lip movements and speech. By using equal-width interval partitioning and Shannon entropy calculation, the uncertainty of user biometric features is converted into a quantifiable indicator, providing an objective basis for the subsequent generation of dynamic permission values. Compared to existing static biometric recognition methods, this method can detect abnormal fluctuations in user behavioral characteristics in real time, such as sudden entropy changes caused by the desynchronization of lip movements and speech in deepfake attacks. This improves the detection sensitivity of new attacks at the feature level.

[0059] The voiceprint embedding vector is input into the linear transformation layer of the attention mechanism to generate a voiceprint query vector. The Shannon entropy value of the fused feature vector is concatenated with the voiceprint embedding vector and then input into the linear transformation layer of the attention mechanism to generate a voiceprint key vector. The dot product of the voiceprint query vector and the voiceprint key vector is calculated to obtain the similarity score between the voiceprint query vector and the voiceprint key vector. The similarity score between the voiceprint query vector and the voiceprint key vector is normalized using the Softmax function to generate the voiceprint credibility weight.

[0060] The motion vector is input into the linear transformation layer of the attention mechanism to generate a motion query vector; the Shannon entropy value of the fused feature vector is concatenated with the motion vector and input into the linear transformation layer of the attention mechanism to generate a motion key vector; the dot product of the motion query vector and the motion key vector is calculated to obtain the similarity score between the motion query vector and the motion key vector, and the Sigmoid function is used to normalize the similarity score between the motion query vector and the motion key vector to generate a semantic risk weight.

[0061] Read the permission baseline parameters from the storage medium, including the permission level mapping relationship, weight baseline value, voiceprint similarity threshold, lip movement synchronization threshold, cross-level access count threshold and quantum watermark generation key; calculate the SHA-256 hash value of the permission baseline parameters and compare it with the pre-stored hash value. If they are inconsistent, it is determined that the permission baseline parameters have been tampered with and the default baseline parameters are loaded.

[0062] The voiceprint credibility weight ratio, semantic risk weight ratio and Shannon entropy value ratio are extracted from the permission baseline parameters; the voiceprint credibility weight ratio, semantic risk weight ratio and Shannon entropy value ratio are used to weighted sum the voiceprint credibility weight, semantic risk weight and Shannon entropy value to obtain the initial weight value; the Sigmoid function is used to compress the initial weight value to the range of 0 to 1 to obtain the dynamic permission value.

[0063] The dynamic authority value and the quantum watermark generation key are concatenated and used as the initial conditions of the chaotic mapping algorithm. The chaotic mapping algorithm is used to generate an unpredictable pseudo-random sequence: the concatenated dynamic authority value and the quantum watermark generation key are converted into floating-point numbers through SHA-256 hash, and the floating-point number is used as the initial value of the Logistic chaotic mapping to iteratively generate a pseudo-random sequence; a fixed length is truncated as the quantum watermark seed.

[0064] Convert the dynamic permission value into the permission level according to the permission level mapping relationship in the permission baseline parameter; extract the time base in the permission level mapping relationship, multiply the dynamic permission value by the time base to obtain the valid duration; take the current time as the permission token start time, and the valid duration as the duration, add the permission token start time to the valid duration to obtain the permission token expiration timestamp, and encapsulate the permission token start timestamp and the permission token expiration timestamp into the permission token expiration stamp;

[0065] The role identifier is extracted from the user authentication information, and the role identifier is bound to the permission level (for example, the role "auditor" only allows permission level ≤ 2). The quantum watermark seed, permission token time stamp and permission level are encoded into the permission token.

[0066] It should also be noted that the security features of permission tokens are enhanced through the coordinated application of the attention mechanism and chaotic mapping. Semantic risk weighting uses a sigmoid function to process motion vector similarity scores, distinguishing between normal lip movements and abnormal mechanical movements. In particular, the use of Shannon entropy as an input parameter in the weight calculation ensures that the generation of dynamic permission values ​​reflects the security posture of the current session environment. When using a chaotic mapping algorithm to generate quantum watermark seeds, the initial condition includes the concatenation of the dynamic permission value and the quantum watermark generation key, making the watermark seed session-unique and unpredictable, avoiding the pattern repetition issue that can occur with existing pseudo-random number generators. The dual timestamp mechanism (start time and expiration time) of the permission token validity stamp, combined with multi-dimensional binding rules for role identifiers, enables fine-grained time-sensitive permission control and prevents unauthorized reuse of permission tokens.

[0067] S2: Based on the permission level and role identification, the permission data block is obtained through the retrieval enhancement generation framework, the hierarchy is divided according to the permission baseline parameters, the retrieval keyword is generated based on the permission token, and a hierarchical index is established to match the permission data block. The permission data block is reordered and a response text is generated. The quantum watermark seed is converted into a binary stream and embedded into the response text to obtain the watermark response text.

[0068] The specific steps are as follows:

[0069] According to the permission level in the permission token, query the mapping relationship between the permission level and the permission data block in the permission baseline parameter to obtain the permission data block identifier corresponding to the permission level;

[0070] Extracting a permission data block from a hierarchical encryption database using a permission data block identifier, wherein the format of the permission data block is a set of key-value pairs;

[0071] The permission data blocks are divided into public layer, internal layer and confidential layer according to the sensitivity classification rules in the permission baseline parameters. Each layer of data blocks is independently encrypted and stored and marked with a level label.

[0072] Attach metadata to each permission data block, including permission level requirements and a list of role identifiers that are allowed access.

[0073] Generate search keywords based on the permission level and role identifier in the permission token, use Elasticsearch (a distributed search and analysis engine) to establish a hierarchical index, including a public layer index, an internal layer index, and a confidential layer index. Match the permission data blocks according to the search keywords in the corresponding hierarchical index, and return a set of permission data blocks whose permission levels and role identifiers meet the matching conditions.

[0074] Define priority sorting rules: the first priority is the permission data block with the closest expiration stamp to the current time; the second priority is that the permission data block with a complete role identifier is prioritized over the permission data block with a partial match; the third priority is to sort by confidentiality layer > internal layer > public layer;

[0075] The permission data blocks are reordered after matching the role identifiers based on the priority sorting rules. Removing the role identifier no longer allows access to the permission data blocks in the list.

[0076] The permission data block is input into the T5 model to generate an extended context, and the extended context and the permission data block are spliced ​​into enhanced text; the enhanced text is input into the GPT-3 model to generate a response text in natural language format.

[0077] The quantum watermark seed is converted into an ASCII code stream using Base64 encoding. Non-continuous space positions are randomly selected in the response text, and the ASCII code stream is replaced with Unicode zero-width space characters to generate the watermark response text.

[0078] It should also be noted that the combination of a role-identity matching reordering mechanism and watermark embedding technology achieves a balance between data security and access efficiency. By defining a composite sorting rule that includes time stamp priority, role full match priority, and data sensitivity level, the retrieval results of permission data blocks not only meet the principle of least privilege but also dynamically optimize the data presentation order based on the business scenario. For example, when auditing role access, the most recently generated confidential layer data blocks are prioritized, while ordinary business roles prioritize matching public layer data. The cascaded generation architecture of the T5 model and the GPT-3 model can transform structured permission data into response text that conforms to natural language interaction habits. The retrieval-enhanced generation framework ensures the accuracy and compliance of the output content. The quantum watermark seed embeds zero-width space characters in the response text, which not only makes the watermark invisible but also leverages the characteristics of Unicode encoding to ensure the persistent retention of watermark information during text copying and transcoding, providing reliable data for subsequent traceability audits.

[0079] S3: Build a causal graph model, calculate the abnormal probability of illegal data access, trigger deep fake identification based on the abnormal probability of illegal data access, generate the fake probability, mark the risk type, and record the abnormal detection report.

[0080] The specific steps are as follows:

[0081] The state where the permission token validity stamp has not expired and the quantum watermark seed has not been tampered with is defined as the permission token valid state, the state where the permission token validity stamp exceeds the permission token expiration timestamp or the hash value of the current quantum watermark seed is inconsistent with the quantum watermark seed in the watermark response text is defined as the permission token expired state, and the permission token valid state and the permission token expired state are defined as permission token state nodes;

[0082] The number of requests for the same permission data block per unit time is defined as the data block access frequency. The number of accesses to data blocks of different permission levels in a single session is defined as the cross-level access count. The data block access frequency and cross-level access count are defined as data access pattern nodes.

[0083] The cosine similarity between the voiceprint embedding vector and the voiceprint embedding vector generated by user registration is defined as voiceprint similarity, the time alignment error between the motion vector and the voiceprint embedding vector is defined as lip movement synchronization, and the voiceprint similarity and lip movement synchronization are defined as deep fake feature nodes;

[0084] The attacker's behavior of obtaining unauthorized data by forging an identity or permission token is defined as a successful illegal access, the attack behavior being detected and blocked is defined as a failed illegal access, and illegal access success and illegal access failure are defined as abnormal access result nodes.

[0085] Extract permission token leakage events, deep fake feature anomaly events, illegal access success events, illegal access failure events, and high-frequency cross-level access events from historical anomaly detection reports. Count the co-occurrence times of permission token expired status events and illegal access success, the co-occurrence times of high-frequency cross-level access events and illegal access success, and the co-occurrence times of deep fake feature anomaly events and permission token expired status.

[0086] Divide the number of co-occurrences of permission token expiration status events and successful illegal access by the total number of permission token expiration to obtain the probability of successful illegal access when the permission token is expired; divide the number of co-occurrences of high-frequency cross-level access events and successful illegal access by the total number of high-frequency cross-level accesses to obtain the probability that high-frequency cross-level access leads to successful illegal access; divide the number of co-occurrences of deep fake feature anomaly events and permission token expiration status by the total number of deep fake feature anomalies to obtain the probability that deep fake feature anomalies lead to permission token expiration;

[0087] Connect the permission token status node and the abnormal access result node, and establish a directed edge from the permission token status node to the abnormal access result node. The expiration status of the permission token directly affects the probability of illegal access success / failure;

[0088] Connect the data access pattern node to the abnormal access result node and establish a directed edge from the data access pattern node to the abnormal access result node. High-frequency cross-level access directly affects the probability of successful illegal access.

[0089] Connect the deep fake feature node to the permission token status node, establish a directed edge from the deep fake feature node to the permission token status node, and trigger the permission token expiration state when the voiceprint similarity is lower than the voiceprint similarity threshold in the permission baseline parameters or the lip movement synchronization exceeds the lip movement synchronization threshold.

[0090] When the current time does not exceed the permission token start timestamp and the permission token expiration timestamp in the permission token time stamp, it is called the time stamp validity state. When the current time exceeds the permission token start timestamp and the permission token expiration timestamp in the permission token time stamp, it is called the time stamp invalid state. When the current time does not exceed the permission token start timestamp and the permission token expiration timestamp in the permission token time stamp, it is called the time stamp valid state. Convert the time stamp validity state and permission level into binary code, convert the data block access frequency and cross-level access count into unit time frequency, and convert the voiceprint similarity and lip movement synchronization into normalized deviation values.

[0091] The binary code, unit time frequency and normalized deviation value are input into the causal graph model; based on the frequency of occurrence of each abnormal access result node in the historical attack data, the conditional probabilities under different combinations of permission token status nodes and data access mode nodes are counted to obtain a conditional probability table; according to the conditional probability table, the conditional probability is transferred along the directed edges of the causal graph model, and the Bayesian formula is used to calculate the joint probability of each node. The joint probabilities of all nodes that directly affect the joint probability of the abnormal access result nodes are weighted and summed to obtain the abnormal probability value of illegal data access.

[0092] Use a sliding window to perform cosine similarity matching between the voiceprint embedding vector and the voiceprint embedding vector generated by user registration, count the proportion of windows below the voiceprint similarity threshold, and generate a voiceprint forgery probability value;

[0093] A dynamic time warping algorithm is used to align the time axis of the speech audio stream with the timing of the lip shape three-dimensional trajectory point set. The speech audio stream is framed and windowed, and plosive sound detection and vowel onset detection are performed to obtain speech key frames. The absolute time difference between the speech key frames and the corresponding lip key points is calculated, and the proportion of frames whose absolute time difference exceeds the lip movement synchronization threshold is counted to generate a lip movement abnormality probability value.

[0094] If the voiceprint forgery probability value is greater than the lip movement abnormality probability value, and the permission token is valid, the risk type is marked as "voiceprint forgery attack"; if the lip movement abnormality probability value is greater than the voiceprint forgery probability value, it is marked as "lip movement forgery attack"; if the voiceprint forgery probability value is equal to the lip movement abnormality probability value, multimodal feature cross-validation and temporal consistency analysis are used to further identify the attack type. If only a single attack type is found, it is marked as the corresponding risk type. If two attack types exist, the risk type is marked as "composite forgery attack";

[0095] If the number of cross-level accesses exceeds the cross-level access threshold in the permission baseline parameters and the permission token is expired, it is marked as a "permission token abuse attack"; if the hash value of the quantum watermark seed is inconsistent with the quantum watermark seed in the watermark response text, and the current voiceprint similarity is not lower than the voiceprint similarity threshold, and the lip movement synchronization is not higher than the lip movement synchronization threshold, it is marked as a "watermark tampering attack".

[0096] The abnormal probability value of illegal data access, voiceprint forgery probability value, lip movement abnormal probability value, risk type, attack timestamp and quantum watermark seed of permission token are integrated into an abnormality detection report.

[0097] It should also be noted that the integration of causal graph models and multimodal anomaly detection constructs a hierarchical risk identification system. By defining three types of nodes and directed edge relationships: permission token status, data access patterns, and deepfake features, discrete security events are converted into a computable joint probability distribution. Conditional probability modeling, linking permission token status nodes to abnormal access result nodes, quantifies the direct impact of token leakage on data security. The reverse causal relationship design, linking deepfake feature nodes to permission token status nodes, captures token status anomalies that may be induced by biometric anomalies. When calculating voiceprint forgery probability using a sliding window, setting a window size adaptive to the length of the speech content avoids feature truncation caused by a fixed window while accurately detecting synthetic tampering traces in localized audio segments. The application of a dynamic time warping algorithm to lip movement anomaly detection eliminates misjudgments caused by individual speaking rate differences by nonlinearly aligning voiceprint and lip movement time series data, making the calculation of lip movement synchronization deviation more consistent with biokinematic principles.

[0098] S4: Update the permission baseline parameters based on the anomaly detection report.

[0099] The specific steps are as follows:

[0100] Using a fixed time window as the statistical window, traverse the anomaly detection report in chronological order, count the number of triggers for each risk type, and divide the number of triggers for each risk type by the total duration of the statistical window to obtain the attack frequency of the risk type.

[0101] In the anomaly detection report, count the number of times the permission token expires and the number of times the illegal access succeeds. Divide the number of times the permission token expires and the number of times the illegal access succeeds by the total number of times the permission token expires to obtain the conditional probability value. Replace the probability of successful illegal access when the permission token expires in the causal graph model with the conditional probability value.

[0102] In the anomaly detection report, the number of co-occurrences of voiceprint similarity lower than the voiceprint similarity threshold and lip movement synchronization higher than the lip movement synchronization threshold in events with risk type "compound forgery attack" is counted; the number of co-occurrences of voiceprint similarity lower than the voiceprint similarity threshold and lip movement synchronization higher than the lip movement synchronization threshold is divided by the minimum value of the number of triggers with risk type "voiceprint forgery attack" and risk type "lip movement forgery attack" to obtain the joint trigger probability; the probability of the deep forgery feature anomaly in the causal graph model causing the expiration of the permission token is replaced by the joint trigger probability.

[0103] Divide the number of times events with the risk type "Permission Token Abuse Attack" are triggered by the total number of times cross-level accesses exceed the cross-level access threshold to obtain the conditional probability value of the permission token abuse attack. If the conditional probability value of the permission token abuse attack is higher than the historical benchmark value in the permission baseline parameters, the cross-level access threshold is reduced by a fixed ratio. If the conditional probability value of the permission token abuse attack is lower than the historical benchmark value in the permission baseline parameters, the cross-level access threshold is increased by a fixed ratio.

[0104] If the attack frequency of the risk type "compound forgery attack" exceeds the upper limit of the compound forgery attack frequency in the permission baseline parameters, the voiceprint similarity threshold and lip movement synchronization threshold will be reduced at a fixed ratio;

[0105] The number of consecutive triggering of events with the risk type of "watermark tampering attack" in adjacent time windows is counted to obtain the number of consecutive watermark tampering attack triggering times; if the number of consecutive watermark tampering attack triggering times exceeds the upper limit of the number of consecutive watermark tampering attack triggering times in the authority baseline parameters, the quantum watermark generation key is replaced.

[0106] It should also be noted that the adaptive parameter update mechanism based on attack frequency enables the continuous evolution of security defense capabilities. By statistically analyzing the ratio of the number of risk type triggers to the total duration, the threat level of various attacks can be dynamically quantified. When the conditional probability of a permission token abuse attack exceeds the historical baseline, the cross-level access threshold is proportionally reduced. This negative feedback adjustment mechanism automatically tightens the criteria for determining high-risk behavior without affecting normal business operations. The combined trigger probability calculation for composite forgery attacks analyzes the co-occurrence frequency of voiceprint and lip movement anomalies and dynamically adjusts the judgment threshold combination of multimodal features, maintaining a high detection accuracy even in the face of new deepfake technologies. The regular replacement strategy for quantum watermark generation keys, combined with statistical monitoring of the number of continuous tampering attacks, enhances the watermark system's anti-cracking capabilities, ensuring that even if part of the session watermark is cracked, the overall system security will not be compromised.

[0107] This embodiment also provides a computer device, which is suitable for the case of a hierarchical and decentralized management method for enterprise data based on a large language model RAG, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the hierarchical and decentralized management method for enterprise data based on a large language model RAG as proposed in the above embodiment.

[0108] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0109] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for implementing hierarchical and decentralized management of enterprise data based on a large language model RAG as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, disk or optical disk.

[0110] In summary, the present invention can capture the uncertainty of user behavior characteristics in real time through: multimodal biometric fusion and Shannon entropy calculation, solve the problem of lack of dynamic risk perception in permission decision-making, improve the ability to identify disguised attacks, ensure the uniqueness and unpredictability of permission tokens, and improve anti-counterfeiting capabilities; incorporate heterogeneous data such as permission status, access behavior, and biometric anomalies into a unified causal graph model, solve the problem of isolated analysis of risk factors, realize accurate modeling of complex attacks, and distinguish deep fake attacks.

[0111] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for hierarchical and decentralized management of enterprise data based on a large language model (RAG), characterized by: include, Acquire the speech audio stream and lip shape three-dimensional trajectory point set, extract the voiceprint embedding vector and motion vector, and fuse them into a fused feature vector. Calculate the Shannon entropy value of the fused feature vector. Calculate the voiceprint credibility weight and semantic risk weight through the attention mechanism. Combine the Shannon entropy value of the fused feature vector to generate a dynamic permission value, load the permission baseline parameters, and generate a permission token. The permission token includes a quantum watermark seed, a role identifier, a permission token expiration stamp, and a permission level. The authority baseline parameters include authority level mapping, weight baseline value, voiceprint similarity threshold, lip movement synchronization threshold, cross-level access count threshold and quantum watermark generation key; Based on the permission level and role identification, the permission data block is obtained through the retrieval enhancement generation framework. The hierarchy is divided according to the permission baseline parameters. The retrieval keyword is generated based on the permission token, and a hierarchical index is established to match the permission data block. The permission data block is reordered after role identification matching, and a response text is generated. The quantum watermark seed is converted into a binary stream and embedded into the response text to obtain the watermark response text. Based on the permission token, voiceprint embedding vector, motion vector, permission data block, and watermark response text, the permission token status node, abnormal access result node, data access pattern node, and deep fake feature node are defined. Directed edges are established, and binary codes, unit time frequencies, and normalized deviation values ​​are generated to form a conditional probability table. The abnormal probability of illegal data access is calculated based on the conditional probability table. Construct a causal graph model, calculate the abnormal probability of illegal data access, trigger deep fake identification based on the abnormal probability of illegal data access, generate the fake probability, mark the risk type, and record the abnormal detection report; Update permission baseline parameters based on anomaly detection reports; Based on the dynamic permission value and permission baseline parameters, a chaotic mapping algorithm is used to generate quantum watermark seeds; According to the permission baseline parameters, the dynamic permission value is converted into the permission level and permission token expiration stamp; Encode the quantum watermark seed, permission level and permission token expiration stamp into the permission token.

2. The method for hierarchical and decentralized management of enterprise data based on a large language model (RAG) as claimed in claim 1, characterized in that: The steps of obtaining the voice audio stream and lip shape three-dimensional trajectory point set, loading the permission baseline parameters, and generating the permission token are as follows: Extract the voiceprint embedding vector and motion vector from the speech audio stream and lip shape three-dimensional trajectory point set, fuse them into a fused feature vector, and calculate the Shannon entropy value of the fused feature vector; According to the Shannon entropy value of the fused feature vector and the voiceprint embedding vector, a linear transformation layer of the attention mechanism is used to generate the voiceprint query vector and the voiceprint key vector; Calculate the dot product of the voiceprint query vector and the voiceprint key vector to obtain the similarity score between the voiceprint query vector and the voiceprint key vector. Use the Softmax function to normalize the similarity score between the voiceprint query vector and the voiceprint key vector to generate the voiceprint credibility weight. According to the Shannon entropy value of the fused feature vector and the motion vector, a linear transformation layer of the attention mechanism is used to generate a motion query vector and a motion key vector. The similarity score between the motion query vector and the motion key vector is normalized using the Sigmoid function to generate a semantic risk weight. The dynamic permission value is calculated using the Sigmoid function based on the Shannon entropy value of the fused feature vector, the voiceprint credibility weight, and the semantic risk weight. Load permission baseline parameters.

3. The method for hierarchical and decentralized management of enterprise data based on a large language model (RAG) as claimed in claim 1, characterized in that: The specific steps of constructing the causal graph model and calculating the abnormal probability of illegal data access are as follows: Define permission token status nodes, abnormal access result nodes, data access mode nodes, and deep fake feature nodes; Establish directed edges from the permission token status node to the abnormal access result node, the data access pattern node to the abnormal access result node, and the deep fake feature node to the permission token status node; Generate binary codes, unit time frequencies, and normalized deviation values ​​based on the data block access frequency and cross-level access counts in the data access pattern node, deep fake feature nodes, and permission tokens; Generate a conditional probability table based on binary coding, unit time frequency and normalized deviation value; According to the conditional probability table, the abnormal probability value of illegal data access is calculated using the Bayesian formula.

4. The method for hierarchical and decentralized management of enterprise data based on a large language model (RAG) as claimed in claim 1, characterized in that: The specific steps of generating the forgery probability are as follows: Based on the voiceprint embedding vector, a sliding window is used to generate the voiceprint forgery probability value; According to the speech audio stream and the lip shape three-dimensional trajectory point set, the dynamic time warping algorithm is used to generate the lip movement abnormality probability value.

5. The method for hierarchical and decentralized management of enterprise data based on a large language model (RAG) as claimed in claim 4, characterized in that: The risk types marked include: marking voiceprint forgery attacks, lip movement forgery attacks and composite forgery attacks based on the voiceprint forgery probability value and the lip movement abnormality probability value; marking permission token abuse attacks based on the number of cross-level accesses in the data access pattern node; and marking watermark tampering attacks based on the quantum watermark seed; The recording of the anomaly detection report refers to recording the abnormal probability value of illegal data access, the voiceprint forgery probability value, the lip movement abnormal probability value, the risk type, the attack timestamp and the quantum watermark seed of the authority token as an anomaly detection report.

6. The method for hierarchical and decentralized management of enterprise data based on a large language model (RAG) as claimed in claim 1, characterized in that: The specific steps for updating the permission baseline parameters based on the anomaly detection report are as follows: Anomaly detection reports statistically analyze the frequency of attacks based on risk types; Adjust the cross-level access count threshold and voiceprint similarity threshold in the permission baseline parameters based on the attack frequency of the risk type; Based on the anomaly detection report, the number of consecutive triggering of watermark tampering attacks is counted, and the quantum watermark generation key in the permission baseline parameters is replaced.

7. The method for hierarchical and decentralized management of enterprise data based on a large language model (RAG) as claimed in claim 1, characterized in that: The reordering of the permission data blocks after matching the role identifiers refers to defining a priority sorting rule, and reordering the permission data blocks after matching the role identifiers using the priority sorting rule.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for hierarchical and decentralized management of enterprise data based on the large language model RAG are implemented as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for hierarchical and decentralized management of enterprise data based on a large language model RAG are implemented as described in any one of claims 1 to 7.

Images