Large model operation process security defense method suitable for end-to-end operation and maintenance all-in-one machine

Abstract

The invention provides a large model operation process security defense method suitable for an end-to-end operation and maintenance all-in-one machine, and relates to the technical field of large model security and operation and maintenance. According to the method, an'input layer-intermediate operation layer-output layer-cross-layer dispatching center 'framework is constructed, and the input layer screens credible data through optimization of a CRC algorithm and credibility pre-evaluation; the middle layer adopts hybrid encryption and an improved isolated forest algorithm to realize data encryption and anomaly detection; the output layer is combined with SHA-3-512 Hash verification and similarity matching to complete result verification; and the cross-layer dispatching center calculates a comprehensive risk value based on a weighted fusion algorithm, and dynamically adjusts a defense strategy. Wherein the risk value calculation is integrated with the interlayer weight and the risk change rate, and the encryption key period is dynamically updated along with the data volume. According to the method, the problems of incomplete protection and poor scene adaptability of the existing method are solved, and the real-time performance and resource limitation requirements of the operation and maintenance all-in-one machine are adapted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large model security and operation and maintenance, and particularly relates to a large model operation process security defense method suitable for an end-to-end operation and maintenance all-in-one machine. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, large models are widely used in end-to-end operation and maintenance all-in-one machines, which can realize intelligent analysis, fault prediction and automatic processing of operation and maintenance data, and significantly improve operation and maintenance efficiency. However, the large model operation process faces multiple security threats, including input data tampering, intermediate operation data leakage, operation node hijacking, and output result forgery. These threats not only lead to operation and maintenance decision errors, but also may leak sensitive operation and maintenance data, causing serious security risks.

[0003] Existing large model security defense methods are mostly aimed at cloud-based large models or single operation links, and lack a full-process, cross-layer collaborative defense mechanism for end-to-end operation and maintenance all-in-one machines. For example, some methods only focus on encrypted transmission of input data and do not systematically protect the intermediate operation process and output results. Some anomaly detection methods do not combine the hardware resource characteristics and operation scene characteristics of operation and maintenance all-in-one machines, resulting in low detection accuracy. At the same time, existing defense mechanisms are mostly statically configured and cannot dynamically adjust defense strategies according to the risks in the operation process, which is not adaptive enough. In addition, the mathematical models involved in existing methods are mostly general algorithms that are not optimized for the resource-constrained, real-time requirement, and other characteristics of operation and maintenance all-in-one machines, resulting in excessive system resource occupation during the defense process and affecting the efficiency of large model operation.

[0004] Therefore, there is an urgent need for a security defense method that can cover the entire process of large model operation, combine the characteristics of operation and maintenance all-in-one machines, and dynamically adjust to solve the problems of incomplete protection, poor adaptability, and excessive resource occupation in existing technologies. SUMMARY

[0005] The present application proposes a large model operation process security defense method suitable for an end-to-end operation and maintenance all-in-one machine, which realizes security protection for the entire process of large model operation input, intermediate process, and output, improves defense adaptability and effectiveness through cross-layer collaboration and dynamic adjustment mechanism, optimizes related algorithms and mathematical models to reduce resource occupation, and meets the real-time requirements of operation and maintenance all-in-one machines.

[0006] The present application proposes a large model operation process security defense method suitable for an end-to-end operation and maintenance all-in-one machine, the steps are as follows: S1, construct an operation and maintenance all-in-one machine large model operation security defense architecture, the architecture includes an operation input layer security module, an intermediate operation layer protection module, an output layer verification module, and a cross-layer security scheduling center; S2, the security module of the operation and maintenance layer preprocesses and verifies the operation and maintenance data input to the large model, and filters trustworthy input data through a data trust authentication algorithm; S3, the intermediate computing layer protection module encrypts intermediate data in real time during the large model computing process based on a dynamic encryption strategy, and monitors the status of computing nodes through an abnormal computing behavior detection algorithm. S4. The output layer verification module verifies the integrity and consistency of the output results of the large model operation and, in conjunction with the operation and maintenance scenario rule base, realizes the credibility rating of the output results. S5, the cross-layer security scheduling center receives security status data from modules at each layer, generates defense strategy adjustment instructions through the security risk assessment model, and realizes dynamic coordination of defense mechanisms at each layer.

[0007] In a further embodiment, the data trust authentication algorithm in step S2 includes data source authentication and data integrity verification. Data source authentication is achieved through digital signature verification based on public key infrastructure, and data integrity verification adopts an optimized cyclic redundancy check algorithm, the expression for which the verification value is calculated is:

[0008] In the formula, CRC is the Cyclic Redundancy Check value; The input is the i-th byte of the operation and maintenance data, where n is the total number of bytes; and k is the length of the checksum. To generate the modulus value corresponding to the polynomial.

[0009] In a further embodiment, the dynamic encryption strategy described in step S3 employs a hybrid encryption mechanism combining symmetric and asymmetric encryption. The symmetric encryption algorithm selected is AES-256-GCM mode, and its key update period T is dynamically adjusted based on the amount of intermediate computation data D. The adjustment formula is as follows:

[0010] In the formula, The basic key update cycle; The data volume impact coefficient; This represents the amount of intermediate computation data generated per unit of time.

[0011] In a further embodiment, the abnormal operation behavior detection algorithm in step S4 is based on the improved isolated forest algorithm. It identifies abnormal behavior by calculating the abnormal score S of the operation node. The formula for calculating the abnormal score is:

[0012] In the formula: For computing node data Average path length in an isolated tree; For a given sample size The average path length constant, when hour, ,in The harmonic number is approximately equal to 1. When abnormal scoring This is considered abnormal behavior. The threshold for anomaly scoring is 0.65.

[0013] In a further embodiment, the integrity verification in step S4 employs a hash verification mechanism, using the SHA-3-512 algorithm to calculate the hash value of the output result. Consistency verification is achieved by matching the output result with historical trusted results in the operation and maintenance scenario rule base. The similarity Sim calculation expression is:

[0014] In the formula, This represents the value of the i-th dimension of the feature vector of the current output result; This represents the value of the i-th dimension of the feature vector of historical reliable results; The dimension of the feature vector is [10, 50]; when When the output results are deemed consistent, they are considered acceptable. The operation and maintenance scenario rule base adopts an incremental update mechanism. The update trigger conditions include: the access of a new operation and maintenance scenario, the error rate of historical reliable results exceeding 5%, and the user's manual update command. During the rule base update process, a dual-database hot standby strategy is adopted to ensure that the output layer verification module runs normally during the update. After the update is completed, the new rule base is confirmed to be effective through consistency verification.

[0015] In a further embodiment, the cross-layer security scheduling center described in step S5 uses a distributed message queue to achieve communication between modules at each layer, reducing message transmission delay. satisfy Furthermore, it monitors the operating status of each module layer through a heartbeat detection mechanism, with a heartbeat detection interval. The timeout is 500ms. If no heartbeat response is received from the module for three consecutive times, the module failure emergency plan will be triggered.

[0016] In a further embodiment, the risk value calculation of the security risk assessment model in step S5 adopts an optimized weighted fusion algorithm, the expression of which is:

[0017] In the formula, The comprehensive security risk value for the large-scale model calculation process of the integrated operation and maintenance machine; This is the input layer security risk value, with a value range of [0,1], where 0 represents no risk and 1 represents extremely high risk; This is the security risk value for the intermediate computing layer, and its value range is [0,1]. This is the output layer security risk value, with a value range of [0,1]. , , These are the risk weight coefficients for the input layer, intermediate computation layer, and output layer, respectively, and they satisfy... ,in , , ; This is the risk change rate weighting coefficient, with a value of 0.15; The rate of change of the sum of the three risk values ​​over time reflects the speed of risk evolution.

[0018] In a further embodiment, during the data source authentication process, a credibility pre-assessment mechanism is employed for data from unknown sources. The formula for calculating the pre-assessment credibility P is as follows:

[0019] In the formula, This refers to the amount of verifiable valid data from data of unknown origin. Total amount of data from unknown sources; This is the data validity weight, with a value of 0.7. For the response time of the data source node; The standard response time is 500ms. As the response time weight, it takes a value of 0.3; when At that time, data from unknown sources may be allowed to enter the subsequent processing flow.

[0020] In a further embodiment, in the hybrid encryption mechanism, the RSA-4096 algorithm is selected as the asymmetric encryption algorithm to encrypt the symmetric encryption key, and a fragmented storage strategy is adopted for the encrypted key, dividing the key into... One fragment, of which At least The key can be reconstructed from a few fragments, among which The fragmented reconstruction is implemented using the Shamir secret sharing algorithm.

[0021] In a further embodiment, the improved isolated forest algorithm introduces computational scenario feature weight optimization, assigning weighted coefficients to features unique to the integrated operation and maintenance machine, such as CPU utilization, memory usage, and data processing throughput. The CPU utilization feature has a weight of 0.3, memory usage feature has a weight of 0.25, data processing throughput feature has a weight of 0.45, and the optimized average path length is... , For the total number of features, Let f be the path length of the f-th feature.

[0022] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention constructs a three-layer protection + cross-layer scheduling architecture covering input, intermediate operation and output, realizes all-round security defense in the large model operation process, and solves the problem of incomplete protection in existing methods; (2) Optimize the algorithm according to the characteristics of the integrated operation and maintenance machine, introduce hardware resource features and operation and maintenance scenario rules, and improve the accuracy of anomaly detection and data verification adaptability; (3) Through the cross-layer security dispatch center and risk assessment model, the defense strategy can be dynamically adjusted and the response can be rapid according to the risk changes. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall security defense architecture of the present invention.

[0024] Figure 2 This is a schematic diagram of the security verification process for the input layer. Detailed Implementation

[0025] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0026] Example 1 This embodiment discloses a security defense method for large-scale model computation processes applicable to end-to-end integrated operation and maintenance machines, including five core steps: constructing a security defense architecture, input layer security verification, intermediate computation layer protection, output layer verification, and cross-layer collaborative scheduling. The overall security defense architecture is shown below. Figure 1 As shown. Specifically: 1. Security defense architecture construction A four-layer defense architecture is constructed, comprising a security module for the input layer, a protection module for the intermediate computation layer, a verification module for the output layer, and a cross-layer security scheduling center. Specifically: The security module of the operation input layer is responsible for authenticating the source, verifying the integrity, and filtering the trustworthiness of input operation and maintenance data. Intermediate computation layer protection module: Implements encryption protection of intermediate data during the computation process and abnormal monitoring of computation nodes; Output layer verification module: performs integrity and consistency verification and reliability rating of the computation output results; Cross-layer security scheduling center: As the core of the defense architecture, it receives security status data uploaded by modules at each layer, generates policy adjustment instructions through risk assessment models, and realizes the coordinated operation of defense mechanisms at each layer.

[0027] 2. Security verification of the arithmetic input layer The input operation and maintenance data is processed using a data trust authentication algorithm, specifically including: Data source authentication: Based on public key infrastructure (PKI), the digital signature of the sender of the input data is verified, and only data that passes the signature verification proceeds to the next step of processing; Data integrity verification: The optimized cyclic redundancy check algorithm is used to calculate the CRC value of the data, and it is compared with the CRC value carried in the data. If they match, the data is determined to be intact. Data processing from unknown sources: For data from unknown sources without digital signatures, the credibility P is calculated through a credibility pre-evaluation mechanism. Only data with P ≥ 0.75 is allowed to enter the large model operation process.

[0028] The expression for calculating the CRC value for data integrity verification is as follows:

[0029] In the formula, This is the i-th byte of input operation and maintenance data. Total number of bytes of data The length of the check bit (value 32). To generate the modulus value corresponding to the polynomial, this optimized algorithm improves the sensitivity to detecting frequently tampered locations by adjusting the weight allocation of data bytes.

[0030] The formula for pre-evaluating the credibility of data from unknown sources is:

[0031] In the formula, This represents the amount of verifiable valid data from data of unknown origin. The total amount of data from unknown sources. The data validity weight is 0.7. For the response time of the data source node, The standard response time is 500ms. The response time weight is 0.3. By combining the validity of the data itself with the response performance of the source node, reasonable filtering of data from unknown sources can be achieved.

[0032] 3. Intermediate computing layer protection A dual protection system is implemented using a dynamic encryption strategy and an abnormal operation behavior detection algorithm. Hybrid encryption mechanism: The intermediate computation data is encrypted using the AES-256-GCM symmetric encryption algorithm, and the symmetric encryption key is protected using the RSA-4096 asymmetric encryption algorithm; at the same time, the symmetric key update cycle is dynamically adjusted based on the amount of intermediate computation data. The larger the amount of data, the shorter the update cycle, to ensure key security. Abnormal operation behavior detection: Based on the improved isolated forest algorithm, the system introduces features unique to the operation and maintenance machine, such as CPU utilization, memory usage, and data processing throughput, and assigns them corresponding weights. Abnormal behavior is identified by calculating the abnormal score S of the operation node data. An alarm is triggered when S>0.65.

[0033] The formula for adjusting the symmetric key update cycle is:

[0034] In the formula, The basic key update cycle is 3600s. The data volume influence coefficient ( byte⁻¹), This represents the amount of intermediate computation data generated per unit time (in bytes). This formula uses an exponential function to achieve a negative correlation between the key update cycle and the amount of data, ensuring security while avoiding resource waste caused by frequent updates.

[0035] The improved anomaly score calculation formula for the Isolation Forest algorithm is as follows:

[0036] in, The optimized average path length, The feature weights are (CPU utilization 0.3, memory usage 0.25, data processing throughput 0.45). Let f be the path length of the f-th feature; The average path length of the sample is a constant. This is a harmonic number. By introducing scene feature weights, the accuracy and scene adaptability of anomaly detection are improved.

[0037] 4. Output layer verification The output results are validated from two dimensions: completeness and consistency. Integrity verification: The hash value of the output result is calculated using the SHA-3-512 algorithm and compared with the intermediate hash value generated during the calculation process. If they match, the result is considered complete. Consistency verification: The output result is converted into a feature vector, and the similarity Sim is calculated with the historical reliable result feature vector in the operation and maintenance scenario rule base. When Sim≥0.8, the result is considered to be consistent. Trust rating: Combining the integrity and consistency verification results, the output results are divided into three levels: "trustworthy", "to be verified" and "untrustworthy". Only the results of the "trustworthy" level are used for subsequent operation and maintenance decisions.

[0038] The similarity calculation formula for consistency verification is:

[0039] In the formula, This represents the value of the i-th dimension of the current output feature vector. This represents the value of the i-th dimension of the feature vector of historical reliable results. The dimension of the feature vector is (10-50). This formula, by calculating the minimum-to-maximum ratio between vectors, can effectively reflect the degree of consistency between results and adapt to the multi-dimensional characteristics of operation and maintenance data.

[0040] 5. Cross-layer collaborative scheduling The cross-layer security dispatch center achieves inter-layer collaboration through the following mechanisms: Data communication: Secure communication between modules is achieved using a distributed message queue (such as Kafka), ensuring message transmission latency ≤100ms; Status monitoring: The operating status of each module is monitored through a heartbeat detection mechanism (every 500ms interval). If no heartbeat response is received for 3 consecutive times, the fault emergency plan will be triggered. Risk assessment: An optimized weighted fusion algorithm is used to calculate the comprehensive security risk value R. Based on the magnitude of R (R<0.3 is low risk, 0.3≤R<0.6 is medium risk, and R≥0.6 is high risk), defense strategy adjustment instructions are generated. Strategy execution: After receiving instructions, each module adjusts its own defense parameters (such as encryption strength, detection frequency, etc.) to achieve dynamic adaptation of the defense mechanism.

[0041] The formula for calculating the comprehensive safety risk value is:

[0042] In the formula, The input layer security risk value is (0-1). The security risk value for the intermediate computing layer is (0-1). Output layer security risk value (0-1); , , Risk weights for each layer, Weighted by the rate of change of risk; This represents the time-varying rate of change of the sum of the three layers of risk. By incorporating the risk change rate, this formula can promptly capture risk evolution trends and enhance the foresight of strategy adjustments.

[0043] Example 2 This embodiment provides a security defense method for large model calculation processes applicable to end-to-end integrated operation and maintenance machines. It is applied to industrial integrated operation and maintenance machines equipped with the GPT-4 lightweight model. The specific implementation steps are as follows: Step 1: Build a security defense architecture The embedded system of the integrated operation and maintenance machine deploys a security module for the input layer, a protection module for the intermediate operation layer, a verification module for the output layer, and a cross-layer security scheduling center. The cross-layer security scheduling center uses an ARM Cortex-A72 processor as its core processing unit and is equipped with a Kafka 2.8.1 distributed message queue for data communication.

[0044] Step 2: Security Verification of the Input Layer When the maintenance sensor node sends maintenance data such as device temperature and vibration to the integrated machine: The data source authentication unit calls the OpenSSL 3.0 library to verify the RSA digital signature carried in the data. If the signature verification is successful, the integrity verification will proceed. The integrity verification unit uses an optimized CRC algorithm to calculate the data check value. The total number of data bytes is n=1024 and k=32. If the calculated CRC value is consistent with the data carried value, the data is determined to be intact. Data sent by an unknown sensor node has no digital signature. The unknown data pre-evaluation unit detects N_valid=896, N_total=1024, T_response=420ms, calculates P=(896 / 1024)×0.7+(420 / 500)×0.3=0.784≥0.75, and allows the data to enter the calculation process.

[0045] Step 3: Intermediate Computation Layer Protection During the large model computation process: The hybrid encryption unit uses the AES-256-GCM algorithm to encrypt intermediate feature data, with a data generation rate of D = 2 × 10⁻⁶ per unit time. 7 byte, calculate key update period T=3600×e (-5×10⁻8×2×107) =3600×e (-1) =1321s, the symmetric key is updated every 1321s; at the same time, the symmetric key is encrypted using the RSA-4096 algorithm, the key is divided into 5 pieces, and stored in different hardware security modules. Only 3 pieces are needed to reconstruct the key; The abnormal behavior detection unit collects real-time data on CPU utilization (65%), memory usage (42%), and data processing throughput (1.2 × 10⁻⁶). 6Based on features such as byte / s, the optimized average path length E'(h(x)) is calculated as 0.3×1.8+0.25×1.5+0.45×2.1=1.83. With a sample size of n=1000, c(n)=2×(ln999+0.5772)-2×999 / 1000≈13.81. The anomaly score S=2^(-1.83 / 13.81)≈0.89>0.65, triggering an anomaly alarm.

[0046] Step 4: Output Layer Verification After the large model outputs the device fault prediction results: The integrity verification unit uses the SHA-3-512 algorithm to calculate the hash value of the result, which is “a3f…7d9”. This hash value is consistent with the intermediate hash value generated during the calculation process, thus determining that the integrity is complete. The consistency verification unit transforms the result into a 20-dimensional feature vector, calculates the similarity Sim with the historical reliable result vector of the same type of device in the rule base, and determines that the consistency is qualified if Sim=0.87≥0.8. The trust rating unit combines the two verification results, rates the output result as "trustworthy", and pushes it to the operation and maintenance decision system.

[0047] Step 5: Cross-layer collaborative scheduling During the operation of the cross-layer security dispatch center: The data communication module receives data from each layer of modules through Kafka, and the actual transmission delay is 85ms≤100ms; The status monitoring module sends a heartbeat detection signal every 500ms. If it fails to receive a response from the intermediate computing layer protection module three times in a row at a certain moment, the fault emergency plan is triggered and the backup protection module is activated. The risk assessment module collected data showing R_in=0.1, R_mid=0.7, and R_out=0.2, with a total risk time change rate of 0.05 / s. The comprehensive risk value R was calculated as R=0.25×0.1+0.45×0.7+0.3×0.2+0.15×0.05=0.445, which is classified as medium risk. The strategy execution module sends adjustment instructions to each layer of modules. The input layer increases the detection frequency to 1.5 times the original frequency, the intermediate layer shortens the key update cycle, and the output layer increases the similarity verification dimension to 30 dimensions, thereby realizing dynamic optimization of the defense strategy.

[0048] Example 3 The difference between this embodiment and Embodiment 2 is that it is applied to a data center operation and maintenance appliance equipped with the LLaMA2-7B model. The feature vector dimension m=35 for consistency verification, and the update trigger condition for the operation and maintenance scenario rule base is set to the access of a new operation and maintenance scenario and a historical reliable result error rate exceeding 5%. When a new air conditioning operation and maintenance scenario is accessed, the rule base initiates an incremental update, employing a dual-database hot standby strategy. During the update process, the output layer verification module operates normally. After the update is completed, consistency verification confirms the new rule base is effective, ensuring continuous and uninterrupted verification.

[0049] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A large model operation process security defense method suitable for end-to-end operation and maintenance all-in-one machine, characterized in that, Comprise the following steps: S1, construct operation and maintenance all-in-one machine big model operation security defense architecture, the architecture comprises operation input layer security module, middle operation layer protection module, output layer verification module and cross-layer security scheduling center; S2, the operation input layer security module pre-processes and security checks the operation and maintenance data input to the big model, and screens the trusted input data through a data trusted authentication algorithm; S3, the middle operation layer protection module encrypts the intermediate data in the big model operation process in real time based on a dynamic encryption strategy, and monitors the operation node state through an abnormal operation behavior detection algorithm; S4, the output layer verification module verifies the integrity and consistency of the big model operation output result, and realizes the trusted rating of the output result in combination with the operation and maintenance scene rule library; S5, the cross-layer security scheduling center receives the security state data of each layer module, generates defense strategy adjustment instructions through a security risk assessment model, and realizes the dynamic cooperation of each layer defense mechanism.

2. The large model operation process security defense method suitable for an end-to-end operation integrated machine according to claim 1, characterized in that, The data trusted authentication algorithm in step S2 comprises data source authentication and data integrity check, the data source authentication is realized through digital signature verification based on a public key infrastructure, and the data integrity check adopts an optimized cyclic redundancy check algorithm, and the check value calculation expression is: In the formula, CRC is a cyclic redundancy check value; is the i th byte of input operation and maintenance data, n is the total number of data bytes, and k is the length of the check bit. is the modulus corresponding to the generation polynomial.

3. The large model operation process security defense method suitable for an end-to-end operation integrated machine according to claim 1, characterized in that, The dynamic encryption strategy in step S3 adopts a hybrid encryption mechanism combining symmetric encryption and asymmetric encryption, the symmetric encryption algorithm selects the AES-256-GCM mode, the key update period T is dynamically adjusted based on the intermediate operation data volume D, and the adjustment formula is: In the formula, is a basic key update period; is a data quantity influence coefficient; is an intermediate operation data generation quantity per unit time.

4. The large model operation process security defense method suitable for an end-to-end operation integrated machine according to claim 1, characterized in that, The abnormal operation behavior detection algorithm in step S4 is based on an improved isolated forest algorithm, and the abnormal behavior is identified by calculating the abnormal score S of the operation node, and the abnormal score calculation formula is: where: is the operational node data is the average path length in the isolated tree; is the average path length constant for a given sample size when , where is the harmonic number, approximated as ; when the anomaly score is determined to be abnormal behavior, is the anomaly score threshold, with a value of 0.

65.

5. The large model operation process security defense method suitable for an end-to-end operation integrated machine according to claim 1, characterized in that, The integrity verification in step S4 adopts a hash check mechanism, selects the SHA-3-512 algorithm to calculate the hash value of the output result, and the consistency verification is realized by similarity matching between the output result and the historical trusted result in the operation and maintenance scene rule library, and the similarity Sim calculation expression is: In the formula, is the i-th dimension value of the current output result feature vector; is the i-th dimension value of the historical trusted result feature vector; is the dimension number of the feature vector, and the value range is [10, 50]; when the output result consistency is determined to be qualified. The operation and maintenance scene rule library adopts an incremental update mechanism, and the update trigger conditions include: new operation and maintenance scene access, historical trusted result error rate exceeding 5%, and user manual update instruction; During the rule library update process, a dual-library hot standby strategy is adopted to ensure the normal operation of the output layer verification module during the update period, and after the update is completed, the new rule library is confirmed to take effect through consistency check.

6. The large model operation process security defense method suitable for an end-to-end operation integrated machine according to claim 1, characterized in that, The cross-layer security scheduling center in step S5 uses a distributed message queue to realize the communication of each layer module, and the message transmission delay is satisfied , and the running state of each layer module is monitored through a heartbeat detection mechanism, and the heartbeat detection interval is 500 ms, and when no module heartbeat response is received for 3 times in succession, a module fault emergency plan is triggered.

7. The large model operation process security defense method suitable for an end-to-end operation integrated machine according to claim 1, characterized in that, The risk value calculation of the security risk assessment model in step S5 adopts an optimized weighted fusion algorithm, and the expression is: In the formula, is the comprehensive security risk value of the operation and maintenance all-in-one machine large model operation process; is the input layer security risk value, and the value range is [0, 1], 0 represents no risk, and 1 represents extremely high risk; is the intermediate operation layer security risk value, and the value range is [0, 1]; is the output layer security risk value, and the value range is [0, 1]; , , are respectively the risk weight coefficients of the input layer, the intermediate operation layer and the output layer, and satisfy , wherein , , ; is the risk change rate weight coefficient, and the value is 0.15; is the time change rate of the sum of the three-layer risk values, reflecting the risk evolution speed.

8. The large model operation process security defense method suitable for an end-to-end operation integrated machine according to claim 2, characterized in that, In the data source authentication process, the trustworthiness pre-evaluation mechanism is adopted for unknown source data, and the pre-evaluation trustworthiness P calculation formula is: In the formula, is the amount of verifiable valid data in unknown source data; is the total amount of unknown source data; is the data validity weight, with a value of 0.7; is the data source node response time; is the standard response time, with a value of 500 ms; is the response time weight, with a value of 0.3; when unknown source data is allowed to enter the subsequent processing flow.

9. The large model operation process security defense method suitable for an end-to-end operation integrated machine according to claim 3, characterized in that, In the hybrid encryption mechanism, the asymmetric encryption algorithm selects RSA-4096 algorithm for encrypting the symmetric encryption key, and a fragmentation storage strategy is adopted for the encrypted key, the key is divided into fragments, wherein at least fragments are required to reconstruct the key, wherein the fragmentation reconstruction is realized by using Shamir secret sharing algorithm.

10. The large model operation process security defense method suitable for an end-to-end operation integrated machine of claim 4, wherein, The improved isolation forest algorithm optimizes the CPU utilization rate, memory occupancy rate, data processing throughput and other characteristics of the operation and maintenance integrated machine by introducing operation scene characteristic weight , wherein the CPU utilization rate characteristic weight is 0.3, the memory occupancy rate characteristic weight is 0.25, the data processing throughput characteristic weight is 0.45, and the optimized average path length , is the total number of characteristics, is the path length of the fth characteristic.

Images