Large model generation content authenticity verification and traceability tracking system

Abstract

The invention relates to the technical field of artificial intelligence, and particularly discloses a large model generation content authenticity verification and traceability tracking system, which is characterized in that content generation data and content source data of a large model are acquired in real time, and contradictory fluctuation characteristic values are extracted in combination with fast Fourier transform; a source distribution characteristic value is calculated by using multi-dimensional statistical characteristics and an isolated forest algorithm; constructing the two types of features into a large model operation state feature vector, inputting the large model operation state feature vector into an artificial intelligence real evaluation model based on support vector machine training, carrying out fusion analysis, and outputting a quantitative real score; according to the method, the full-process automation from state sensing, intelligent evaluation to closed-loop control is realized, the intelligent level and the autonomous risk response capability of a large-scale model system are improved, and the real level of the currently generated content is judged according to the scoring result. Good application prospects and engineering practical values are realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a system for verifying the authenticity and tracing the source of content generated by a large model. Background Technology

[0002] Large-scale model-generated content refers to text, code, dialogue, or other forms of content automatically generated by large-scale pre-trained language models (such as GPT-4, Claude, and PaLM). These models, based on deep learning technology, acquire powerful language understanding and generation capabilities through training on massive amounts of data and are widely used in various fields. Content generation requires access to vast amounts of data, but the authenticity of this data directly impacts the quality of the generated content.

[0003] The existing technology has the following shortcomings:

[0004] Currently, research on monitoring the operational status of large-scale models mainly focuses on single-parameter detection or rule-based empirical judgment systems. In addition, some systems employ traditional statistical analysis or linear models for fault prediction. However, these systems often struggle to adapt to the nonlinear operational characteristics of large-scale models in complex dynamic environments, leading to problems such as delayed warnings, high false alarm rates, and insufficient risk identification capabilities. Furthermore, existing technologies lack effective data fusion and intelligent evaluation mechanisms, hindering the comprehensive perception and real-time response to the overall true state of large-scale models, thus restricting the intelligent and autonomous development of large-scale model systems. Summary of the Invention

[0005] The purpose of this invention is to provide a system for verifying the authenticity and tracing the source of content generated by large models, so as to solve the problems mentioned above.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A system for verifying the authenticity and tracing the source of content generated by a large model includes:

[0008] The data acquisition module collects content generation data and content source data of the large model in real time during the execution of content generation tasks in the large model;

[0009] An anomaly analysis module analyzes the content generation data and calculates the contradiction fluctuation characteristic value to assess the risk of anomalies in the current generation of the large model.

[0010] The source distribution analysis module analyzes the content source data and calculates the source distribution characteristic value based on the fluctuation of the source reliability, which is used to determine whether the large model is in an abnormal working state.

[0011] The feature fusion evaluation module constructs a large model running state feature vector from the contradictory fluctuation feature values ​​and the source distribution feature values, and inputs it into the trained artificial intelligence real evaluation model for fusion analysis.

[0012] The Real Decision Control Module determines the realism level of the content currently generated by the large model based on the analysis results, and triggers corresponding early warning mechanisms and emergency control strategies based on the realism level.

[0013] As a further aspect of the present invention: the assessment of the current generation realism anomaly risk of the large model specifically includes:

[0014] During the content generation task performed by the large model, the content generation data of the large model is collected in real time, analyzed, and the contradiction fluctuation characteristic value is calculated based on the degree of contradiction in the content generation data. It is then determined whether the contradiction fluctuation characteristic value is greater than or equal to the first preset threshold. If it is, the current generation of the large model is abnormal; otherwise, the current generation of the large model is normal.

[0015] As a further aspect of the present invention: the process for obtaining the contradictory fluctuation characteristic value is as follows:

[0016] During the monitoring period, the content generation data of the large model is collected in real time according to the time series. The real-time content generation data of the large model is subjected to fast Fourier transform to obtain the frequency domain representation, including: performing fast Fourier transform on the content generation data of the time series to obtain a set of complex spectral coefficients. Fast Fourier transform can convert the time domain signal into the frequency domain signal, which is convenient for analyzing the energy distribution of the signal.

[0017] Data is generated based on the content of the transformed large model, and the energy density of each frequency component is calculated.

[0018] Calculate the ratio of each frequency component to the total energy to obtain the spectral energy ratio;

[0019] Calculate the average value of the energy ratio of all frequency components, calculate the difference between each spectral energy ratio and the average value of all frequency component energy ratios, and take the absolute value to obtain the difference of the spectral energy ratio. Sum all the differences to obtain the contradictory fluctuation characteristic value.

[0020] As a further aspect of the present invention: the process of obtaining the source distribution feature value is as follows:

[0021] During the content generation task performed by the large model, the content source data of the large model is collected in real time, the content source data is analyzed, and the source distribution characteristic value is calculated based on the fluctuation of the source reliability. It is then determined whether the source distribution characteristic value is greater than or equal to the second preset threshold. If it is, the large model is in an abnormal working state; otherwise, the large model is in a normal working state.

[0022] As a further aspect of the present invention: the process of obtaining the source distribution feature value is as follows:

[0023] Obtain the content source dataset within the monitoring period;

[0024] Calculate the mean, standard deviation, and rate of change of the dataset from which the content is sourced;

[0025] Calculate the mean and standard deviation of the rate of change based on the rate of change of the content source dataset;

[0026] Calculate the local fluctuation values ​​of the data called by the large model;

[0027] A multidimensional feature vector is constructed by combining the mean, standard deviation, mean and standard deviation of the rate of change of the source dataset with the local fluctuation values ​​of the large model call data;

[0028] The isolated forest algorithm is used to train the feature vectors, generate multiple isolated trees, and then calculate the anomaly score for each data point.

[0029] Calculate the anomaly score for each data point based on the path length output by the isolated forest model.

[0030] The source distribution characteristic value is obtained by averaging the anomaly scores of all monitored data points.

[0031] As a further aspect of the present invention: the construction of a large model running state feature vector from the contradictory fluctuation feature values ​​and the source distribution feature values, and the input of this vector into a trained artificial intelligence real evaluation model for fusion analysis, specifically includes:

[0032] During the content generation task performed by the large model, the contradictory fluctuation feature values ​​and source distribution feature values ​​of the large model are obtained. These feature values ​​are then used to construct a feature vector representing the running state of the large model. This vector serves as the input to the AI ​​real evaluation model, minimizing the error between the predicted real score and the actual real score of the large model. This minimization serves as the training objective for the AI ​​real evaluation model. The AI ​​real evaluation model is then trained, and based on the trained AI real evaluation model, the real score of the large model is output. The AI ​​real evaluation model is a support vector machine model.

[0033] As a further aspect of the present invention: the training process of the artificial intelligence real evaluation model is as follows:

[0034] The contradictory fluctuation feature values ​​and source distribution feature values ​​in the historically collected large model running data are used to construct feature vectors, which are then used as input samples for the support vector machine model.

[0035] The corresponding actual scores are used as output labels to form a training sample set;

[0036] Select a support vector machine model suitable for regression tasks, set the kernel function type to RBF kernel, and initialize the model parameters including: penalty coefficient C and kernel function parameter γ;

[0037] The training sample set is input into the support vector machine model. The model searches for an optimal regression hyperplane by nonlinearly mapping the input feature vectors, which minimizes the deviation between the predicted true score and the actual true score.

[0038] As a further aspect of the present invention: the determination of the true level of the currently generated content of the large model specifically includes:

[0039] During the content generation task performed by the large model, it is determined whether the actual score of the large model is greater than or equal to the third preset threshold. If it is, the content generated by the large model is real; otherwise, the content generated by the large model is not real.

[0040] As a further aspect of the present invention: the step of determining the true level of the currently generated content of the large model based on the analysis results, and triggering corresponding early warning mechanisms and emergency control strategies based on the true level, specifically includes:

[0041] If the generated content is determined to be authentic, continue executing the current task; if the generated content is determined to have real potential risks, enter a risk state.

[0042] When a risk condition is determined, a corresponding multi-level early warning mechanism is activated, including: issuing audible and visual alarms, sending early warning information to remote control terminals, recording abnormal events, and generating fault logs.

[0043] The beneficial effects of this invention are:

[0044] (1) This invention constructs a high-precision, high-sensitivity operational status monitoring system by real-time acquisition of content generation data and content source data of a large model during the content generation process, combined with advanced signal processing and intelligent analysis algorithms. Specifically, in terms of content generation data processing, a fast Fourier transform technique is introduced to convert time-domain signals into frequency-domain signals, extract spectral energy distribution characteristics, and quantify the imbalance of power fluctuations, thereby achieving early identification of abnormal states in the generation of data. In the analysis of content source data, a multidimensional statistical feature extraction and isolated forest anomaly detection algorithm are adopted, comprehensively considering source reliability, local fluctuations, and overall change trends, to effectively determine whether the large model is in an abnormal working state. This system can not only capture minute changes in operational status but also maintain good robustness and adaptability in complex environments, significantly improving the system's ability to identify potential authenticity risks.

[0045] (2) This invention innovatively constructs an artificial intelligence real assessment model based on support vector machines, fully integrating multi-dimensional key feature information collected during the operation of the large model, including contradictory fluctuation feature values ​​and source distribution feature values, to form a representative feature vector of the large model's operating state. This model employs a support vector regression system suitable for regression analysis. During training, it aims to minimize the error between the predicted score and the actual score, achieving optimal regression hyperplane fitting in the nonlinear feature space through kernel function mapping, thereby obtaining a precise quantitative assessment capability of the realism of the large model's operation. In practical applications, this artificial intelligence real assessment model can receive input feature data from the sensor system in real time, output continuous real scores, and automatically classify the realism level of the currently generated content state according to preset scoring thresholds. Based on different risk levels, the system intelligently triggers corresponding early warning mechanisms and emergency control strategies, such as audible and visual alarms, remote communication notifications, and operation suspension, realizing a closed-loop control process from state perception, intelligent analysis, risk assessment to automatic response. This fully automated architecture not only improves the intelligence level of the large model system but also significantly enhances its autonomous decision-making and risk response capabilities in complex and dynamic environments. Attached Figure Description

[0046] The invention will now be further described with reference to the accompanying drawings.

[0047] Figure 1 This is a flowchart illustrating the authenticity verification and source tracing system for the large model-generated content of this invention.

[0048] Figure 2 This is a flowchart illustrating the process of obtaining contradictory fluctuation characteristic values ​​according to the present invention.

[0049] Figure 3 This is a flowchart illustrating the process of obtaining the source distribution characteristic values ​​according to the present invention.

[0050] Figure 4 This is a flowchart illustrating the training process of the artificial intelligence real evaluation model of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Please see Figure 1 As shown, this invention is a system for verifying the authenticity and tracing the source of content generated by a large model, comprising:

[0053] The data acquisition module collects content generation data and content source data of the large model in real time during the execution of content generation tasks in the large model;

[0054] An anomaly analysis module analyzes the content generation data and calculates the contradiction fluctuation characteristic value to assess the risk of anomalies in the current generation of the large model.

[0055] The source distribution analysis module analyzes the content source data and calculates the source distribution characteristic value based on the fluctuation of the source reliability, which is used to determine whether the large model is in an abnormal working state.

[0056] The feature fusion evaluation module constructs a large model running state feature vector from the contradictory fluctuation feature values ​​and the source distribution feature values, and inputs it into the trained artificial intelligence real evaluation model for fusion analysis.

[0057] The Real Decision Control Module determines the realism level of the content currently generated by the large model based on the analysis results, and triggers corresponding early warning mechanisms and emergency control strategies based on the realism level.

[0058] During the content generation task performed by the large model, content generation data and content source data of the large model are collected in real time, specifically including:

[0059] During the content generation task performed by the large model, multi-channel sensor modules deployed on the large model itself collect content generation data and content source data in real time. The content generation data reflects the authenticity of the large model's content generation in the current state; the content source data monitors the data retrieval status of the large model during continuous operation. The sensors for data acquisition and status monitoring are connected to the large model's main control system to ensure the real-time nature and accuracy of the collected data.

[0060] Furthermore, the data acquisition process includes timed sampling and analog-to-digital conversion of content generation data and content source data of the large model. The sampling frequency is set according to the dynamic characteristics of the large model operation, preferably 10 to 100 times per second. The acquired data is transmitted to the data processing module via a communication interface for subsequent feature extraction and real-world evaluation analysis, thereby achieving continuous, closed-loop real-world monitoring of the large model's operational status.

[0061] Analyzing the content generation data and calculating the contradiction fluctuation characteristic value are used to assess the risk of current generation anomalies in the large model, specifically including:

[0062] During the content generation task performed by the large model, the content generation data of the large model is collected in real time, analyzed, and the contradiction fluctuation characteristic value is calculated based on the degree of contradiction in the content generation data. It is then determined whether the contradiction fluctuation characteristic value is greater than or equal to the first preset threshold. If it is, the current generation of the large model is abnormal; otherwise, the current generation of the large model is normal.

[0063] The first preset threshold is set based on empirical data. It is easy to know that when all the content-generated data is real, there will be no contradictions within the content-generated data. Conversely, when some of the content-generated data is real and some is not, the unreal data will contradict the real data. The more contradictions there are, the lower the authenticity. However, when the contradictions are very small, they will not affect the content and can be ignored. Therefore, based on empirical data, the largest contradiction fluctuation characteristic value that will not affect the content can be obtained as the first preset threshold. The subsequent second and third preset thresholds are also set based on similar principles.

[0064] Please see Figure 2 As shown, the process of obtaining the contradictory fluctuation characteristic value is as follows:

[0065] During the monitoring period, the content generation data of the large model is collected in real time according to the time series. The real-time content generation data of the large model is subjected to fast Fourier transform to obtain the frequency domain representation, including: performing fast Fourier transform on the content generation data of the time series to obtain a set of complex spectral coefficients. Fast Fourier transform can convert the time domain signal into the frequency domain signal, which is convenient for analyzing the energy distribution of the signal.

[0066] Data is generated based on the content of the transformed large model, and the energy density of each frequency component is calculated using the following expression:

[0067] E a =|F a | 2 ;

[0068] Among them, E a Let F be the energy density of the a-th frequency component, representing a measure of the signal fluctuation intensity at the corresponding frequency. a Let represent the complex spectral coefficient of the a-th frequency component, where a represents the number of frequency components after the fast Fourier transform.

[0069] Calculate the ratio of each frequency component to the total energy to obtain the spectral energy ratio;

[0070] Calculate the average value of the energy ratio of all frequency components, calculate the difference between each spectral energy ratio and the average value of all frequency component energy ratios, and take the absolute value to obtain the difference of the spectral energy ratio. Sum all the differences to obtain the contradictory fluctuation characteristic value.

[0071] It should be noted that this invention innovatively introduces Fast Fourier Transform (FFT) technology to convert time-domain signals into frequency-domain signals through frequency domain analysis of the generated data from a large model. This allows for the extraction of the sum of the spectral energy ratio differences, reflecting power fluctuation characteristics, as a contradictory fluctuation feature value. When the generated data has authenticity issues, the data source is abnormal; that is, the source attributes of real and non-real data differ. This will be reflected in the frequency domain signal under FFT, leading to a significant difference between the frequency domain signals of real and non-real data. Therefore, the contradictory fluctuation feature value can be used to characterize this difference.

[0072] Analyzing the content source data and calculating source distribution characteristic values ​​based on the fluctuations in source reliability is used to determine whether the large model is in an abnormal operating state. Specifically, this includes:

[0073] During the content generation task performed by the large model, the content source data of the large model is collected in real time, the content source data is analyzed, and the source distribution characteristic value is calculated based on the fluctuation of the source reliability. It is then determined whether the source distribution characteristic value is greater than or equal to the second preset threshold. If it is, the large model is in an abnormal working state; otherwise, the large model is in a normal working state.

[0074] Please see Figure 3 As shown, the process of obtaining the source distribution feature values ​​is as follows:

[0075] Obtain the content source dataset within the monitoring period;

[0076] Calculate the mean, standard deviation, and rate of change of the dataset from which the content is sourced;

[0077] The expression for calculating the rate of change is:

[0078] ΔT i =T i+1 -T i ;

[0079] Where i represents each content source collection point, and T i ΔT represents the content collection value of the i-th content source collection point. i This represents the rate of change in the content collected at the (i+1)th collection point compared to the ith collection point.

[0080] Calculate the mean and standard deviation of the rate of change based on the rate of change of the content source dataset;

[0081] The expression for calculating the local fluctuation values ​​of the large model's call data is as follows:

[0082] R i =max(T) i ,T i+1 ,…,T i+k )-min(T i ,T i+1 ,…,T i+k );

[0083] Where k represents the sliding window size, R i This represents the local fluctuation value of the i-th content source collection point;

[0084] A multidimensional feature vector is constructed by combining the mean, standard deviation, mean and standard deviation of the rate of change of the source dataset with the local fluctuation values ​​of the large model call data;

[0085] The isolated forest algorithm is used to train the feature vectors, generate multiple isolated trees, and then calculate the anomaly score for each data point.

[0086] Calculate the anomaly score for each data point based on the path length output by the Isolation Forest model.

[0087] The calculation expression for the anomaly score is as follows:

[0088]

[0089] In the formula, S(T) i E(h(T)) represents the anomaly score of the i-th content source collection point. i )) indicates that T is in an isolated tree i The average path length when isolated, c(n) represents the normalization factor;

[0090] The source distribution characteristic value is obtained by averaging the anomaly scores of all monitored data points. The calculation expression is as follows:

[0091]

[0092] In the formula, n represents the total number of data collection points for the content source, and A T This represents the characteristic value of the source distribution.

[0093] It should be noted that this invention constructs a multi-dimensional feature vector, including mean, standard deviation, rate of change, and local fluctuation values, by statistically analyzing the dataset from which the large model content originates. The isolated forest algorithm is then used to model this feature vector, calculating anomaly scores for each data point. Finally, the source distribution feature value is obtained through the mean of these scores. This system can effectively identify genuine fluctuations caused by various factors during the operation of a large model, thereby accurately determining whether the large model is in an abnormal operating state. It exhibits strong robustness and adaptability, significantly improving the intelligence and precision of large model operation status monitoring.

[0094] The contradictory fluctuation characteristic values ​​and source distribution characteristic values ​​are used to construct a large model's operational state feature vector, which is then input into a trained artificial intelligence real-world evaluation model for fusion analysis. Specifically, this includes:

[0095] During the content generation task performed by the large model, the contradictory fluctuation feature values ​​and source distribution feature values ​​of the large model are obtained. These feature values ​​are then used to construct a feature vector representing the running state of the large model. This vector serves as the input to the AI ​​real evaluation model, minimizing the error between the predicted real score and the actual real score of the large model. This minimization serves as the training objective for the AI ​​real evaluation model. The AI ​​real evaluation model is then trained, and based on the trained AI real evaluation model, the real score of the large model is output. The AI ​​real evaluation model is a support vector machine model.

[0096] Please see Figure 4As shown, the training process of the artificial intelligence real evaluation model is as follows:

[0097] The contradictory fluctuation feature values ​​and source distribution feature values ​​in the historically collected large model running data are used to construct feature vectors, which are then used as input samples for the support vector machine model.

[0098] The corresponding actual scores are used as output labels to form a training sample set;

[0099] Select a support vector machine model suitable for regression tasks, set the kernel function type to RBF kernel, and initialize the model parameters including: penalty coefficient C and kernel function parameter γ;

[0100] Function: To control the model's tolerance for classification errors, and to balance model complexity and training error. A large C value means the model is more inclined to fit the training data (reducing misclassification), but may lead to overfitting (poor generalization ability). A small C value means more misclassification is allowed, the model is simpler (may be underfitting), but may have stronger generalization ability.

[0101] The kernel function parameter used here is the Gaussian kernel, which is the existing technology and has a general expression. Its function is to control the influence range of a single sample on the model. A large γ means a small influence range of the sample and a complex model (which may overfit). A small γ means a large influence range of the sample and a smooth model (which may underfit).

[0102] The training sample set is input into the support vector machine model. The model searches for an optimal regression hyperplane by nonlinearly mapping the input feature vectors, which minimizes the deviation between the predicted true score and the actual true score.

[0103] During the content generation task performed by the large model, it is determined whether the actual score of the large model is greater than or equal to the third preset threshold. If it is, the content generated by the large model is real; otherwise, the content generated by the large model is not real.

[0104] Based on the analysis results, the true level of the content currently generated by the large model is determined, and corresponding early warning mechanisms and emergency control strategies are triggered according to the true level, including:

[0105] The big model real score output by the AI ​​real assessment model is compared with the preset real score threshold. If the real score is greater than or equal to the third preset threshold, the generated content is determined to be in a real state and the current task continues. If the real score is less than the third preset threshold, the generated content is determined to have real hidden dangers and enters a risk state.

[0106] When a risky state is determined, the corresponding multi-level early warning mechanism is activated, including: issuing audible and visual alarms, sending early warning information to remote control terminals, recording abnormal events and generating fault logs;

[0107] Based on the risk level, corresponding emergency control strategies are automatically triggered, including: pausing the currently generated content actions, adjusting the large model's posture or path, and initiating an emergency return procedure to the real area, in order to ensure the authenticity of the large model itself and the operating environment.

[0108] Furthermore, this solution proposes a storage medium on which a computer-readable program is stored. When the computer-readable program is invoked, it executes the aforementioned system for verifying the authenticity and tracing the source of the large model-generated content.

[0109] It is understandable that the storage medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid-state drive (SSD).

[0110] In summary, the advantages of this invention are as follows: This invention aims to achieve real-time monitoring and intelligent evaluation of the operational status of large-scale models in complex working environments. The system deploys multi-channel sensor modules to collect real-time content generation and source data of the large-scale model, and combines this with Fast Fourier Transform (FFT) technology to extract contradictory fluctuation feature values ​​to identify abnormal energy consumption risks. Simultaneously, it utilizes statistical analysis and the Isolation Forest algorithm to construct a multi-dimensional feature vector for the large-scale model, calculating source distribution feature values ​​to determine whether the large-scale model is in an abnormal operating state. Further, the above two types of feature values ​​are used to construct a feature vector of the large-scale model's operational status, which is input into an AI-based real-world evaluation model trained on a Support Vector Machine (SVM). After fusion analysis, the system outputs the true score of the large-scale model. Based on the true score, the system can automatically determine the true level of the currently generated content and trigger corresponding early warning mechanisms and emergency control strategies, such as audible and visual alarms, remote notifications, and path adjustments.

[0111] This invention enhances the ability to perceive changes in key operating parameters of large-scale models by introducing frequency domain signal processing and anomaly detection algorithms, realizing a fully automated monitoring system from data acquisition, feature extraction, intelligent evaluation to closed-loop control. Compared to traditional systems, this invention has higher sensitivity and robustness, significantly improving the realism and reliability of large-scale model systems, and possesses promising engineering application prospects.

[0112] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A system for verifying the authenticity and tracing the source of content generated by a large model, characterized in that, include: The data acquisition module collects content generation data and content source data of the large model in real time during the execution of content generation tasks in the large model; An anomaly analysis module analyzes the content generation data and calculates the contradiction fluctuation characteristic value to assess the risk of anomalies in the current generation of the large model. The source distribution analysis module analyzes the content source data and calculates the source distribution characteristic value based on the fluctuation of the source reliability, which is used to determine whether the large model is in an abnormal working state. The feature fusion evaluation module constructs a large model running state feature vector from the contradictory fluctuation feature values ​​and the source distribution feature values, and inputs it into the trained artificial intelligence real evaluation model for fusion analysis. The Real Decision Control Module determines the realism level of the content currently generated by the large model based on the analysis results, and triggers corresponding early warning mechanisms and emergency control strategies based on the realism level.

2. The system for verifying the authenticity and tracing the source of large model-generated content according to claim 1, characterized in that, The assessment of the risk of real-world anomalies generated by the current large model specifically includes: During the content generation task performed by the large model, the content generation data of the large model is collected in real time, analyzed, and the contradiction fluctuation characteristic value is calculated based on the degree of contradiction in the content generation data. It is then determined whether the contradiction fluctuation characteristic value is greater than or equal to the first preset threshold. If it is, the current generation of the large model is abnormal; otherwise, the current generation of the large model is normal.

3. The system for verifying the authenticity and tracing the source of large model-generated content according to claim 2, characterized in that, The process for obtaining the contradictory fluctuation characteristic value is as follows: During the monitoring period, the content generation data of the large model is collected in real time according to the time series. The real-time content generation data of the large model is subjected to fast Fourier transform to obtain the frequency domain representation, including: performing fast Fourier transform on the content generation data of the time series to obtain a set of complex spectral coefficients. Fast Fourier transform can convert the time domain signal into the frequency domain signal, which is convenient for analyzing the energy distribution of the signal. Data is generated based on the content of the transformed large model, and the energy density of each frequency component is calculated. Calculate the ratio of each frequency component to the total energy to obtain the spectral energy ratio; Calculate the average value of the energy ratio of all frequency components, calculate the difference between each spectral energy ratio and the average value of all frequency component energy ratios, and take the absolute value to obtain the difference of the spectral energy ratio. Sum all the differences to obtain the contradictory fluctuation characteristic value.

4. The system for verifying the authenticity and tracing the source of large model-generated content according to claim 3, characterized in that, The process for obtaining the source distribution feature values ​​is as follows: During the content generation task performed by the large model, the content source data of the large model is collected in real time, the content source data is analyzed, and the source distribution characteristic value is calculated based on the fluctuation of the source reliability. It is then determined whether the source distribution characteristic value is greater than or equal to the second preset threshold. If it is, the large model is in an abnormal working state; otherwise, the large model is in a normal working state.

5. The system for verifying the authenticity and tracing the source of large model-generated content according to claim 4, characterized in that, The process for obtaining the source distribution feature values ​​is as follows: Obtain the content source dataset within the monitoring period; Calculate the mean, standard deviation, and rate of change of the dataset from which the content is sourced; Calculate the mean and standard deviation of the rate of change based on the rate of change of the content source dataset; Calculate the local fluctuation values ​​of the data called by the large model; A multidimensional feature vector is constructed by combining the mean, standard deviation, mean and standard deviation of the rate of change of the source dataset with the local fluctuation values ​​of the large model call data; The isolated forest algorithm is used to train the feature vectors, generate multiple isolated trees, and then calculate the anomaly score for each data point. Calculate the anomaly score for each data point based on the path length output by the isolated forest model; The source distribution characteristic value is obtained by averaging the anomaly scores of all monitored data points.

6. The system for verifying the authenticity and tracing the source of large model-generated content according to claim 5, characterized in that, The process of constructing a large model's operational state feature vector from the contradictory fluctuation characteristic values ​​and source distribution characteristic values, and then inputting it into a trained artificial intelligence real-world evaluation model for fusion analysis, specifically includes: During the content generation task performed by the large model, the contradictory fluctuation feature values ​​and source distribution feature values ​​of the large model are obtained. These feature values ​​are then used to construct a feature vector representing the running state of the large model. This vector serves as the input to the AI ​​real evaluation model, minimizing the error between the predicted real score and the actual real score of the large model. This minimization serves as the training objective for the AI ​​real evaluation model. The AI ​​real evaluation model is then trained, and based on the trained AI real evaluation model, the real score of the large model is output. The AI ​​real evaluation model is a support vector machine model.

7. The system for verifying the authenticity and tracing the source of large model-generated content according to claim 6, characterized in that, The training process of the AI-based real-world evaluation model is as follows: The contradictory fluctuation feature values ​​and source distribution feature values ​​in the historically collected large model running data are used to construct feature vectors, which are then used as input samples for the support vector machine model. The corresponding actual scores are used as output labels to form a training sample set; Select a support vector machine model suitable for regression tasks, set the kernel function type to RBF kernel, and initialize the model parameters including: penalty coefficient C and kernel function parameter γ; The training sample set is input into the support vector machine model. The model searches for an optimal regression hyperplane by nonlinearly mapping the input feature vectors, which minimizes the deviation between the predicted true score and the actual true score.

8. The system for verifying the authenticity and tracing the source of large model-generated content according to claim 7, characterized in that, The determination of the true level of the content currently generated by the large model specifically includes: During the content generation task performed by the large model, it is determined whether the actual score of the large model is greater than or equal to the third preset threshold. If it is, the content generated by the large model is real; otherwise, the content generated by the large model is not real.

9. The system for verifying the authenticity and tracing the source of large model-generated content according to claim 8, characterized in that, The process of determining the true level of the content currently generated by the large model based on the analysis results, and triggering corresponding early warning mechanisms and emergency control strategies based on the true level, specifically includes: If the generated content is determined to be authentic, continue executing the current task; if the generated content is determined to have real potential risks, enter a risk state. When a risk condition is determined, a corresponding multi-level early warning mechanism is activated, including: issuing audible and visual alarms, sending early warning information to remote control terminals, recording abnormal events, and generating fault logs.

Images