AI creation monitoring method and system based on big data processing

Abstract

The invention discloses an AI creation monitoring method and system based on big data processing, and the method comprises the following steps: analyzing a to-be-monitored content through a multi-modal feature extraction model, extracting the multi-dimensional features of the to-be-monitored content, and generating a corresponding standardized feature vector; inputting the standardized feature vector into a pre-trained multivariate classification model, and outputting a comprehensive judgment result; based on the comprehensive judgment result, a preset dynamic monitoring strategy library is inquired, and a corresponding monitoring action is determined and executed; the multi-modal feature extraction model and the multivariate classification model are constructed, execution results of monitoring actions are collected, a feedback data set is formed, incremental updating and optimization are conducted on the multi-modal feature extraction model and the multivariate classification model, in the implementation process of the method, multi-dimensional feature extraction is conducted on various AI creation contents through construction of the multi-modal feature extraction model, comprehensive judgment is conducted through the multivariate classification model, and the multi-modal feature extraction model and the multivariate classification model are combined. And executing a corresponding monitoring action based on a comprehensive judgment result, thereby realizing a multi-modal monitoring process of the AI creation content.

Description

Technical Field

[0001] This invention relates to the field of big data processing, and specifically to an AI-based creative monitoring method and system based on big data processing. Background Technology

[0002] With the development of artificial intelligence technologies such as deep learning, generative adversarial networks (GAN), large language models (LLM), and diffusion models, AI-generated content (AIGC) has entered a stage of explosive growth. AI has evolved from a simple auxiliary tool into a subject capable of creating content independently or collaboratively. Its application areas include all media forms such as text, images, audio, and video, giving rise to emerging business models such as news writing, algorithmic music composition, digital human live streaming, and AI painting.

[0003] Compared to traditional human creation, AI creation exhibits significantly different technical characteristics, such as large scale and high efficiency. AI can generate massive amounts of content in a short time. It also possesses high fidelity and concealment. High-level AIGC works (such as deepfake videos and AI dialogues) are perceptually deceptive and difficult to distinguish from their authenticity by ordinary means. Furthermore, the randomness and uncertainty of content generation, based on probability models, make AI output unpredictable. This may lead to discriminatory content, non-compliant content, or factual errors due to training data bias. Therefore, monitoring AI creation has become a crucial link in ensuring the authenticity of information and maintaining the order of dissemination.

[0004] In existing technologies, monitoring of AI-generated content mainly includes technical detection methods and content tracing and authentication methods. Technical detection methods are further subdivided into digital watermarking, statistical feature analysis, and model-based fingerprinting. Digital watermarking achieves tracking by embedding invisible identifiers in the generated content, but suffers from poor compatibility and is easily removed. Statistical feature analysis relies on subtle anomalies in the frequency domain, pixel distribution, or language structure of AI-generated content for identification, but its generalization ability to new models is weak. Model-based fingerprinting, while effectively associating with the generation model, is only applicable to closed systems with known architectures. Content tracing and authentication methods rely on blockchain or trusted third parties to record the content generation path, but this is difficult to enforce in decentralized environments and poses a risk of data privacy leaks. Furthermore, existing monitoring systems mostly focus on single-modal content detection, lacking cross-modal correlation analysis capabilities and failing to cope with the complex scenarios of multi-channel collaborative AI creation, resulting in certain monitoring blind spots.

[0005] Therefore, there is an urgent need for an AI-based creative monitoring method and system based on big data processing to solve the above problems.

[0006] It should be noted that the information disclosed in this background section is only for understanding the background technology of the present invention, and therefore may include information that does not constitute prior art. Summary of the Invention

[0007] The purpose of this invention is to solve the problem that there are many types of AI-generated content, making it difficult to monitor them uniformly.

[0008] To address the aforementioned technical issues, this application proposes an AI-driven content creation monitoring method based on big data processing. The method is executed by a monitoring server and includes the following steps:

[0009] S100: The multimodal feature extraction model deployed on the monitoring server is used to parse the content to be monitored, extract its multi-dimensional features, and generate corresponding standardized feature vectors; wherein, the content to be monitored is digital content in text, image, audio or video format; the multi-dimensional features include at least style features, semantic features, statistical features and code features;

[0010] S200: Input the standardized feature vector into a pre-trained multivariate classification model. The multivariate classification model outputs a comprehensive judgment result for the content to be monitored. The comprehensive judgment result includes an AI generation probability value P_ai, and if it is generated by AI, it further includes the creation type identifier T_type and the generation source credibility identifier S_source.

[0011] S300: Based on the comprehensive judgment result, query the preset dynamic monitoring strategy library, determine and execute the corresponding monitoring action; the level of the monitoring action is dynamically adjusted according to the combination of AI generation probability value, creation type identifier and generation source credibility identifier;

[0012] S400: Collect the execution results of the monitoring actions, form a feedback dataset, and use the feedback dataset to incrementally update and optimize the multimodal feature extraction model and the multivariate classification model.

[0013] Furthermore, in step S100, the extraction of style features specifically includes:

[0014] For the text content, extract linguistic features and calculate the deviation vector Style_text between it and multiple preset human author style templates;

[0015] For image content, extract its color distribution histogram, texture features, and frequency domain features, and calculate its similarity vector Style_image with the image style in the preset AI image style library;

[0016] For audio content, extract its Mel frequency cepstral coefficients, fundamental frequency profile, and spectral centroid, and calculate its similarity vector Style_audio with the output audio of various preset speech synthesis models;

[0017] The above vectors are normalized and concatenated to form a unified style feature vector V_style.

[0018] Furthermore, in step S100, the extraction of semantic features specifically includes:

[0019] The content to be monitored is subjected to deep semantic encoding using a pre-trained large language model or vision model to obtain a high-dimensional semantic embedding vector V_semantic.

[0020] Calculate the cosine similarity between V_semantic and multiple typical vectors in a pre-defined AI-generated semantic common pattern library to form a semantic consistency vector V_consistency; the AI-generated semantic common pattern library contains a set of vectors encoded from known AI-generated content.

[0021] The semantic features are composed of V_semantic and V_consistency.

[0022] Furthermore, in step S100, the extraction of statistical features specifically includes:

[0023] For text content, calculate its information entropy, character repetition rate, and outliers in the frequency of use of specific punctuation marks;

[0024] For image content, calculate the local correlation of pixel values ​​and the statistics of compression artifacts;

[0025] For audio / video content, calculate its signal-to-noise ratio and outlier values ​​of coding features;

[0026] After normalizing the above statistics, a statistical feature vector V_statistical is formed.

[0027] Furthermore, in step S100, the extraction of code features targets executable files or script-like content, specifically including:

[0028] Analyze its code structure and extract features from the function call graph, control flow graph, and referenced external libraries;

[0029] Calculate the structural similarity between the stated code structure and the output code of known AI code generation tools;

[0030] This forms the code feature vector V_code.

[0031] Furthermore, in step S200, the multi-class classification model is trained in the following manner:

[0032] Collect labeled human-created content and AI-created content as training samples. The labeling information includes at least: whether it is AI-generated, the specific type of AI creation, and the AI ​​tool or model used.

[0033] The initial deep learning model is trained using the training sample set. The initial deep learning model adopts a multi-task learning architecture, and its output layer includes at least: a binary classification output head for judging the probability of AI generation, a multi-classification output head for distinguishing the type of creation, and a multi-classification output head for identifying the source of generation.

[0034] In the comprehensive judgment result, P_ai comes from the binary classification output head, T_type comes from the multi-classification output head, and S_source comes from the generation source identification output head.

[0035] Furthermore, in step S200, the calculation of the source credibility identifier S_source also introduces a credibility evaluation function:

[0036] C = f(P_ai, P_type, P_source) = α× P_ai +β× P_type +γ× P_source, where P_ai is the AI ​​generation probability, P_type is the model's confidence in the decision type T_type on the multi-class output head, P_source is the model's confidence in the decision source S_source on the generation source recognition output head, and α, β, γ are adjustable weight coefficients, and α+β+γ=1;

[0037] When C is below the first threshold θ_low, S_source is marked as "uncertain"; when C is above the second threshold θ_high, S_source is output according to the model's judgment result; when C is between θ_low and θ_high, the low confidence processing procedure is triggered.

[0038] Furthermore, in step S300, the dynamic monitoring strategy library contains multiple strategy rules, each consisting of a trigger condition and a monitoring action; the trigger condition is a logical expression based on P_ai, T_type, and S_source; the monitoring action includes multiple levels:

[0039] Level L1: When P_ai < δ1, execute action A1: record the content ID and judgment result in the log, and perform no further operations;

[0040] Level L2: When δ1 ≤ P_ai < δ2 and S_source is not "uncertain", execute action A2: record detailed information and generate a mild alert notification;

[0041] Level L3: When P_ai ≥ δ2, or T_type belongs to the high-risk type list, execute action A3: intercept or restrict the spread of the content, and initiate the deep source tracing analysis subprocess;

[0042] Level L4: When P_ai ≥ δ2 and S_source points to an unauthorized or malicious AI model, perform action A4: immediately isolate the content, generate a high-risk alert, and attempt to trace the uploader's identity.

[0043] Furthermore, step S400 further includes the following steps:

[0044] The collected manual review records, user reports and feedback, and in-depth source tracing analysis results generated during the monitoring process constitute a labeled feedback dataset D_feedback;

[0045] The multivariate classification model is incrementally trained periodically using D_feedback to adjust the model parameters;

[0046] Analyze the cases in D_feedback that are judged incorrectly, and reversely adjust the feature extraction focus or structure of the multimodal feature extraction model.

[0047] Furthermore, the method is applied to an AI-driven content creation monitoring system based on big data processing, the system comprising:

[0048] The multi-dimensional feature extraction module is used to parse the content to be monitored, extract its multi-dimensional features, and generate corresponding standardized feature vectors through a multi-modal feature extraction model deployed on the monitoring server; wherein, the content to be monitored is digital content in text, image, audio, or video format; the multi-dimensional features include at least style features, semantic features, statistical features, and code features;

[0049] The multivariate classification decision module inputs the standardized feature vector into a pre-trained multivariate classification model. The multivariate classification model outputs a comprehensive judgment result for the content to be monitored. The comprehensive judgment result includes an AI generation probability value, and if it is generated by AI, it further includes the creation type identifier and the credibility identifier of the generation source.

[0050] The execution module is used to query a preset dynamic monitoring strategy library based on the comprehensive judgment result, determine and execute the corresponding monitoring action; the level of the monitoring action is dynamically adjusted according to the combination of AI generation probability value, creation type identifier and generation source credibility identifier;

[0051] The closed-loop feedback module is used to collect the execution results of the monitoring actions, form a feedback dataset, and use the feedback dataset to incrementally update and optimize the multimodal feature extraction model and the multivariate classification model.

[0052] The beneficial effects of this invention are: by constructing a multimodal feature extraction model, multi-dimensional feature extraction is performed on various AI-created content, and a comprehensive judgment is made through a multi-classification model. Based on the comprehensive judgment result, corresponding monitoring actions are executed, thereby realizing the multimodal monitoring process of AI-created content.

[0053] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating an AI-based content creation monitoring method based on big data processing according to the present invention.

[0055] Figure 2 This is a schematic diagram of the module structure of an AI creation monitoring system based on big data processing according to the present invention. Detailed Implementation

[0056] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0057] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. 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 should fall within the scope of protection of the present invention.

[0058] Since current generative AI creation encompasses multiple formats such as text, images, and videos, traditional monitoring methods struggle to cover multimodal content generation scenarios, resulting in low monitoring efficiency and a high risk of missed detections. Therefore, this invention addresses these issues.

[0059] Reference Figure 1As shown, an embodiment of the present invention proposes an AI creation monitoring method based on big data processing. This method is executed through a monitoring server, which refers to a computing device with data analysis, processing, and storage capabilities. The method includes the following steps:

[0060] S100: The multimodal feature extraction model deployed on the monitoring server is used to parse the content to be monitored, extract its multi-dimensional features, and generate corresponding standardized feature vectors; wherein the content to be monitored is digital content in text, image, audio or video format, and the multi-dimensional features include at least style features, semantic features, statistical features and code features.

[0061] During the monitoring of AI creation, it is necessary to analyze the monitored content to extract multi-dimensional features. This is because AI creation not only includes text content but also multiple modal information such as images, audio, and video. These multi-modal information contain various feature types. For example, style features reflect the expression of the creation, semantic features reflect the theme and meaning of the content, statistical features describe the data distribution pattern, and code features are associated with the technical traces of the generation process. Therefore, it is necessary to extract these multi-dimensional features and generate corresponding standardized feature vectors. These standardized feature vectors can be represented using a unified numerical format, such as floating-point arrays.

[0062] For example, parsing the monitored content and extracting its multi-dimensional features can be achieved using a multimodal feature extraction model. This model can be deployed on the monitoring server, directly utilizing the server's computing resources to complete the multi-dimensional feature extraction. The model can be built on a deep neural network, integrating text encoders, image encoders, audio encoders, and video encoders to process data inputs from different modalities. Through a cross-modal alignment mechanism, the features of each modality are mapped to a unified semantic space, ensuring that the generated standardized feature vectors are consistent and comparable, thereby improving the accuracy and efficiency of subsequent monitoring and analysis.

[0063] Specifically, the extraction of style features includes:

[0064] For text content, linguistic features such as word frequency distribution, syntactic complexity, n-gram statistical features, and frequency of rhetorical device usage are extracted to reflect its writing style and expression habits. The deviation vector Style_text between the text content and various preset human author style templates is calculated as a style similarity index. Among them, n-gram statistical features refer to the frequency distribution of n consecutive words or characters in the text, which is used to characterize the local structural rules of language. It usually consists of a multi-order statistical sequence from 2-gram to 5-gram. The specific application method can refer to the existing technology, which will not be elaborated in this embodiment. The deviation vector Style_text is obtained by calculating the Euclidean distance or cosine similarity between the text to be monitored and various human author style templates in terms of word frequency distribution, syntactic structure, and rhetorical device usage. The smaller the value, the closer it is to the human writing style.

[0065] For image content, its color distribution histogram, texture features, and frequency domain features are extracted, and a similarity vector Style_image is calculated between it and the image style in the preset AI image style library. The color distribution histogram reflects the overall distribution of hue, saturation, and brightness of the image; texture features extract the image's detail and structural regularity through local binary mode or Gabor filtering; and frequency domain features utilize Fourier transform to analyze high-frequency noise and periodic artifacts, effectively identifying the synthetic traces unique to the generative model. Combined with the preset AI image style database, the similarity vector Style_image of the image in the above feature space is calculated to quantify its style correlation with the output of typical generative models, thereby assisting in determining the source of the content's creation. The preset AI image style database contains style feature clusters formed by image samples trained on public datasets by mainstream generative models such as Stable Diffusion, DALL-E, and MidJourney. Cosine similarity or Mahalanobis distance is used to measure the degree of matching between the test image and the styles generated by various models, thus achieving source tracing analysis.

[0066] For audio content, its Mel-frequency cepstral coefficients, fundamental frequency profile, and spectral centroid are extracted, and a similarity vector Style_audio is calculated between it and the output audio of various preset speech synthesis models. Among them, the Mel-frequency cepstral coefficients are used to characterize the short-time spectral envelope features of the audio, reflecting the resonance structure of the sound; the fundamental frequency profile describes the dynamic change trend of pitch, which can effectively distinguish the prosodic differences between natural human voice and synthesized speech; and the spectral centroid indicates the concentrated position of audio energy in the frequency domain, helping to identify the smoothness and regularity unique to electronically generated signals. By comparing the extracted features with the style clusters of audio generated by preset speech synthesis models (such as Tacotron, WaveNet, FastSpeech), the cosine similarity is calculated to obtain Style_audio, quantifying the matching degree between the test audio and various AI speech generation systems, and thus determining whether it is synthesized by the model and its possible source path.

[0067] It should be noted that for video content, the above image and audio feature extraction methods can be combined to calculate the Style_image of the video frame and the Style_audio of the audio stream, and then spatiotemporal consistency analysis can be fused to detect the alignment relationship between the image and sound on the time axis and identify whether there are deep synthesis traces such as face swapping, voice cloning or lip-syncing. Therefore, it will not be described in detail in this embodiment.

[0068] In addition, time series analysis models can be introduced to dynamically model the style vectors of consecutive frames in the video, capturing subtle temporal inconsistencies in the generated content and further improving detection accuracy.

[0069] After generating the deviation vector Style_text, similarity vector Style_image, and Style_audio, normalization and concatenation are required to generate a unified style feature vector V_style. Specifically, this unified style feature vector V_style is horizontally concatenated after L2 normalization of each sub-vector to ensure consistent numerical scales across different modalities. This integrates multi-dimensional style information from text, image, and audio to form a comprehensive representation of the content creation mode. For example, the text's Style_text, the image's Style_image, and the audio's Style_audio are concatenated in modal order to form a joint feature vector, and L2 normalization is used to eliminate dimensional differences, ultimately constituting a unified multi-modal style representation V_style. This style feature vector can be expressed as follows: V_style = [Norm(Style_text), Norm(Style_image), ... Norm(Style_audio)], where Norm(·) represents the L2 normalization operation, ensuring that the contributions of each modal sub-vector are balanced under a unified dimension; this vector not only retains the fine-grained features of multimodal generation traces, but also strengthens the cross-modal collaborative discrimination capability through linear combination in high-dimensional space, making the minor contradictions between semantics and form of forged content explicit.

[0070] Based on the above embodiments, the semantic feature extraction step specifically includes:

[0071] The content to be monitored is subjected to deep semantic encoding using a pre-trained large language model or vision model to obtain a high-dimensional semantic embedding vector V_semantic.

[0072] Calculate the cosine similarity between V_semantic and multiple typical vectors in a pre-defined "AI-generated semantic common pattern library" to form a semantic consistency vector V_consistency; the "AI-generated semantic common pattern library" contains a set of vectors representing typical patterns such as logical fallacies, factual errors, and content gaps, obtained by encoding known AI-generated content, and the semantic features are jointly composed of V_semantic and V_consistency.

[0073] Semantic features refer to the feature representations that reflect whether the content conforms to the human knowledge system and logical structure at a deep semantic level. For example, semantic features can identify whether there is conceptual confusion, causal inversion, or common sense error in the text, whether there is structural distortion in the image that does not conform to the laws of physics, and whether there is a lack of coherence in the speech logic in the audio content. In the process of extracting semantic features, it is necessary to use a pre-trained large-scale language model or vision model to perform deep semantic encoding on the content to be monitored, so as to obtain a high-dimensional semantic embedding vector V_semantic. This high-dimensional semantic embedding vector is a mathematical representation that can capture the deep semantic structure of the input content. Each dimension implicitly corresponds to a certain abstract semantic attribute, such as object relationship, scene logic, or emotional tendency. Its expression can be V_semantic∈R^d, where d is the hidden layer dimension of the model, usually between 768 and 4096, and R^d represents the d-dimensional real space. This vector maps the original input to the latent space through the deep nonlinear transformation of the Transformer architecture, thereby capturing the context-sensitive semantics of the input content.

[0074] For example, semantic features are jointly composed of V_semantic and V_consistency. V_semantic represents the semantic structure of the content itself, while V_consistency quantifies its closeness to typical AI-generated patterns. V_consistency, which is composed of the cosine similarity value of the high-dimensional semantic embedding vector V_semantic and each typical vector in the preset pattern library, reflects the comprehensive features of the content in two dimensions: semantic rationality and AI generation tendency. After concatenation, the two are linearly transformed to output the final semantic feature representation. The linear transformation can achieve dimensionality reduction and feature fusion through a learnable weight matrix. This weight matrix can be expressed as W ∈ R^{k×(d+m)}, where k is the output feature dimension, d is the original dimension of V_semantic, and m is the number of typical vectors in the pattern library. The concatenated high-dimensional vector is mapped to the low-dimensional semantic feature space through matrix multiplication, realizing the joint representation of the deep semantics of the input content and the consistency of the AI-generated patterns.

[0075] Based on the above embodiments, the extraction of statistical features specifically includes:

[0076] For text content, calculate its information entropy, character repetition rate, and outliers in the frequency of use of specific punctuation marks;

[0077] For image content, calculate the local correlation of pixel values ​​and the statistics of compression artifacts;

[0078] For audio / video content, calculate its signal-to-noise ratio and outlier values ​​of coding features;

[0079] After normalizing the above statistics, a statistical feature vector V_statistical is formed.

[0080] In text content, information entropy is calculated based on the probability distribution of characters or words, reflecting the randomness and predictability of the text. Low entropy values ​​may indicate templated generation, while high entropy values ​​may indicate chaotic content or noise interference. Character repetition rate is used to detect abnormal use of consecutive repeated symbols, such as excessive exclamation marks and ellipses, which may expose non-human writing habits. The frequency of specific punctuation marks is identified by statistically analyzing the occurrence density of commas, periods, quotation marks, etc., to identify patterns that do not conform to conventional language norms.

[0081] In image content, local correlation of pixel values ​​reflects the linear or nonlinear dependence between adjacent pixels. Natural images usually have strong local continuity, while AI-generated images may have abnormal correlation distribution due to periodic artifacts introduced during the upsampling process. Compression artifact statistics, by analyzing indicators such as quantization matrix residues and block effect intensity, identify whether the image has undergone multiple lossy compressions, and indirectly determine its propagation path and generation source.

[0082] In audio / video content, the signal-to-noise ratio (SNR) reflects the power ratio of the original signal to the noise introduced by the encoding. A low SNR may indicate traces of multiple transcoding or synthesis, while outliers in the encoding features can be used to identify non-authentic content by analyzing the degree to which technical parameters such as sampling rate and bitstream structure deviate from the normal distribution.

[0083] Extracting statistical features can effectively capture abnormal patterns in content at the low-level signal level, making up for the shortcomings of semantic features in modeling surface statistical regularities.

[0084] Based on the above embodiments, code feature extraction typically targets only executable files or script-like content, such as scripting languages ​​like Python and JavaScript. Specifically, the code feature extraction process includes:

[0085] Analyze its code structure and extract features from the function call graph, control flow graph, and referenced external libraries;

[0086] Calculate the structural similarity between the stated code structure and the output code of known AI code generation tools;

[0087] This forms the code feature vector V_code.

[0088] Code structure parsing is primarily achieved through Abstract Syntax Tree (AST) traversal. The function call graph mainly reflects the depth and frequency distribution of calls between modules. Control flow graph features characterize the complexity and branch density of the program execution path, while external library references reflect the code's preference for external dependencies. AI-generated code often exhibits abnormal patterns such as excessively short call chains, overly regular control flow, or frequent introduction of uncommon libraries. Therefore, effective discrimination criteria can be constructed based on these features.

[0089] The calculation of the structural similarity between the code structure and the output code of known AI generation tools can be quantified using graph matching algorithms or vector space models. This transforms the function call graph and control flow graph into comparable topological features. Combined with the statistical deviation of external library reference frequency, the degree of deviation from typical human programming behavior is comprehensively evaluated. The final code feature vector V_code can be represented as V_code = [f_1, f_2, ..., f_n], where f_i represents the i-th normalized feature component, covering graph structure similarity score, control flow complexity ratio, external library reference entropy value, and call chain length statistics, etc.

[0090] After obtaining the style feature vector, semantic feature vector, statistical feature vector, and code feature vector, these feature vectors need to be concatenated to output a standardized feature vector. This standardized feature vector should have a fixed dimension and uniform numerical ranges for each component. Normalization is usually used to achieve this. In this embodiment, a fully connected layer can be used for dimensionality reduction and standardization. A fully connected layer is a neural network structure that can perform linear transformations and nonlinear activations on the concatenated high-dimensional feature vectors to achieve feature fusion and dimensionality reduction. Therefore, it is suitable for integrating various feature vectors and outputting a unified standardized feature vector V_final. Using a fully connected layer can effectively retain key discriminative information while suppressing redundant feature interference. Its weight matrix is ​​automatically learned through training data and can dynamically adjust the contribution of each sub-feature. For specific application methods, please refer to Chinese Invention Patent Publication No. CN117496388A, which will not be detailed in this embodiment.

[0091] S200: Input the standardized feature vector into a pre-trained multivariate classification model. The multivariate classification model outputs a comprehensive judgment result for the content to be monitored. The comprehensive judgment result includes an AI generation probability value P_ai, and if it is generated by AI, it further includes the creation type identifier T_type and the generation source credibility identifier S_source.

[0092] After obtaining the standardized feature vector of the content to be detected, it is input into a multivariate classification model based on an ensemble learning architecture. This model is formed by fusing multiple weak classifiers through an adaptive weighting strategy and generates a comprehensive judgment result that includes the probability value of AI generation, the creation type identifier if it is AI-generated, and the credibility identifier of the generation source.

[0093] For example, in one possible embodiment, the training process of a multi-class classification model may include the following steps:

[0094] Collect a massive amount of labeled human-created content and AI-created content as training samples. The labeling information includes at least: whether it is AI-generated, the specific type of AI creation, and the AI ​​tools or models used.

[0095] The initial deep learning model is trained using the training sample set. The initial deep learning model adopts a multi-task learning architecture, and its output layer includes at least: a binary classification output head for judging the probability of AI generation, a multi-classification output head for distinguishing the type of creation, and a multi-classification output head for identifying the source of generation.

[0096] In the comprehensive judgment result, P_ai comes from the binary classification output head, T_type comes from the multi-classification output head, and S_source comes from the generation source identification output head. Here, P_ai represents the probability value that the content is generated by AI, with a value range of [0,1]. T_type identifies the creation type to which the content belongs, including creation type identifiers such as text, image, audio, or video. S_source indicates the generation source credibility identifier of the specific AI model or tool used to generate the content. The three together constitute a multi-dimensional judgment output on the source and attributes of the content, effectively supporting subsequent content review.

[0097] The calculation of the source credibility identifier S_source also introduces a credibility evaluation function: C = f(P_ai, P_type, P_source) = α × P_ai + β × P_type + γ × P_source, where P_ai is the AI ​​generation probability, P_type is the model's confidence in the judgment type T_type on the multi-class output head, P_source is the model's confidence in the judgment source S_source on the source identification output head, and α, β, γ are adjustable weight coefficients, and α+β+γ=1;

[0098] Among them, P_type is calculated based on the maximum probability value of the corresponding category label output by the softmax layer, reflecting the reliability of the model's classification result of the creation type; P_source is similarly taken as the highest confidence level of the predicted source category in the generation source identification branch.

[0099] The confidence assessment function C can be based on a weighted linear combination function or extended to a nonlinear fusion function based on a neural network. This allows for adaptive adjustment of the contribution of each dimension's confidence level according to the actual scenario requirements, improving overall judgment accuracy while ensuring computational efficiency.

[0100] When C is below the first threshold θ_low, S_source is marked as "uncertain"; when C is above the second threshold θ_high, S_source is output according to the model's judgment result; when C is between θ_low and θ_high, the low confidence processing procedure is triggered.

[0101] For example, in a content moderation system, if a text fragment, after model analysis, outputs P_ai=0.93, T_type=“text”, P_type=0.88, S_source=“WPSAI”, P_source=0.76, and weighting coefficients α=0.5, β=0.3, γ=0.2, the calculated C=0.5×0.93+0.3×0.88+0.2×0.76=0.887; if θ_high=0.85 is set, then since C>θ_high, the judgment result is reliable, and the effective output of S_source is “WPSAI”, which is used for subsequent source tracing and labeling. If C=0.82 and θ_low=0.80, then it enters the low reliability processing flow, and the system will initiate manual review or additional metadata verification mechanism to confirm the reliability of the source;

[0102] S300: Based on the comprehensive judgment result, query the preset dynamic monitoring strategy library, determine and execute the corresponding monitoring action; the level of the monitoring action is dynamically adjusted according to the combination of AI generation probability value, creation type identifier and generation source credibility identifier.

[0103] After outputting a comprehensive judgment result through a multivariate classification model, the system automatically matches the corresponding monitoring strategy rules based on the preset dynamic monitoring strategy library, the range of AI generation probability values, the specific category of the creation type identifier, and the level of credibility of the generation source.

[0104] The dynamic monitoring strategy library contains multiple strategy rules, each consisting of a trigger condition and a monitoring action. The trigger condition is a logical expression based on P_ai, T_type, and S_source. The monitoring actions include multiple levels.

[0105] Level L1: When P_ai < δ1, execute action A1: record the content ID and judgment result in the log, and perform no further operations;

[0106] Level L2: When δ1 ≤ P_ai < δ2 and S_source is not "uncertain", execute action A2: record detailed information and generate a mild alert notification;

[0107] Level L3: When P_ai ≥ δ2, or T_type belongs to the high-risk type list, execute action A3: intercept or restrict the spread of the content, and initiate the deep source tracing analysis subprocess;

[0108] Level L4: When P_ai ≥ δ2 and S_source points to an unauthorized or malicious AI model, perform action A4: immediately isolate the content, generate a high-risk alert, and attempt to trace the uploader's identity.

[0109] For example, when δ1=0.6 and δ2=0.85, if a certain content P_ai=0.91, T_type=“image”, and S_source=“unknown model”, then the trigger condition meets the L4 level, executes the A4 action, and the system automatically isolates the content and reports it to the security center; if P_ai=0.75, T_type=“text”, and S_source=“WPSAI”, then it enters the L2 level, only generates a mild warning and records metadata.

[0110] At level L3, the deep source analysis subprocess includes the following steps:

[0111] First, the implicit identifier metadata of the content is retrieved, and the AIGC field, generation service provider information, and unique number are parsed. Then, the integrity of the file is verified by hash check value, and the digital signature is compared to confirm the authenticity of the source. If the metadata is missing or the verification fails, a reverse query request is initiated to the generation source to trace the original creation node. At the same time, correlation analysis is performed by combining IP logs and account behavior sequences to locate potential risk subjects. Finally, the source tracing results are written to the security audit log, and the corresponding handling process is initiated according to the platform rules.

[0112] S400: Collect the execution results of the monitoring actions, form a feedback dataset, and use the feedback dataset to incrementally update and optimize the multimodal feature extraction model and the multivariate classification model.

[0113] Since the entire monitoring process involves the collaborative operation of multiple models, it is also necessary to form a feedback dataset based on the execution results of the monitoring actions. This feedback dataset is then used to incrementally train the multimodal feature extraction model and the multi-class classification model, dynamically adjusting the threshold boundaries of δ1 and δ2 and the classification decision weights to improve the model's recognition accuracy for new AI-generated content.

[0114] Specifically, step S400 includes the following steps:

[0115] The collected manual review records, user reports and feedback, and in-depth source tracing analysis results generated during the monitoring process constitute a labeled feedback dataset D_feedback;

[0116] The multivariate classification model is incrementally trained periodically using D_feedback to adjust the model parameters;

[0117] Analyze the cases in D_feedback that are judged incorrectly, and reversely adjust the feature extraction focus or structure of the multimodal feature extraction model.

[0118] During the collection of feedback datasets, not only model-generated results can be used, but also manual review records, user reports, and other sources of supplementary labels can be cited to form a more complete feedback dataset. When incrementally training the multi-classification model, online learning or small-batch update strategies can be adopted to avoid the resource consumption caused by full retraining. Simultaneously, the characteristics of incorrectly judged samples in the feedback dataset can be analyzed to identify cognitive biases of the model in specific modalities or scenarios, thereby optimizing the feature extraction process of the multimodal feature extraction model. For example, to address the problem of over-reliance on texture redundancy features in image modalities, the semantic layer feature weights can be dynamically enhanced. When analyzing new features, the analysis dimensions can be increased by introducing temporal behavior patterns and cross-modal correlation features to improve the model's ability to identify hidden AIGC content.

[0119] Based on the above embodiments, another embodiment of the present invention proposes an AI creation monitoring system based on big data processing. This system runs an AI creation monitoring method based on big data processing as described in the foregoing embodiments. The system includes:

[0120] The multi-dimensional feature extraction module is used to parse the content to be monitored, extract its multi-dimensional features, and generate corresponding standardized feature vectors through a multi-modal feature extraction model deployed on the monitoring server; wherein, the content to be monitored is digital content in text, image, audio, or video format; the multi-dimensional features include at least style features, semantic features, statistical features, and code features;

[0121] The multivariate classification decision module inputs the standardized feature vector into a pre-trained multivariate classification model. The multivariate classification model outputs a comprehensive judgment result for the content to be monitored. The comprehensive judgment result includes an AI generation probability value, and if it is generated by AI, it further includes the creation type identifier and the credibility identifier of the generation source.

[0122] The execution module is used to query a preset dynamic monitoring strategy library based on the comprehensive judgment result, determine and execute the corresponding monitoring action; the level of the monitoring action is dynamically adjusted according to the combination of AI generation probability value, creation type identifier and generation source credibility identifier;

[0123] The closed-loop feedback module is used to collect the execution results of the monitoring actions, form a feedback dataset, and use the feedback dataset to incrementally update and optimize the multimodal feature extraction model and the multivariate classification model.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI creation monitoring method based on big data processing, the method being executed by a monitoring server, characterized in that: The method includes the following steps: S100: The multimodal feature extraction model deployed on the monitoring server is used to parse the content to be monitored, extract its multi-dimensional features, and generate corresponding standardized feature vectors; wherein, the content to be monitored is digital content in text, image, audio or video format; the multi-dimensional features include at least style features, semantic features, statistical features and code features; S200: Input the standardized feature vector into a pre-trained multi-class classification model. The multi-class classification model outputs a comprehensive judgment result for the content to be monitored. The comprehensive judgment result includes an AI generation probability value P_ai, and if it is generated by AI, it further includes the creation type identifier T_type and the generation source credibility identifier S_source. S300: Based on the comprehensive judgment result, query the preset dynamic monitoring strategy library, determine and execute the corresponding monitoring action; the level of the monitoring action is dynamically adjusted according to the combination of AI generation probability value, creation type identifier and generation source credibility identifier; S400: Collect the execution results of the monitoring actions, form a feedback dataset, and use the feedback dataset to incrementally update and optimize the multimodal feature extraction model and the multivariate classification model.

2. The AI creation monitoring method based on big data processing according to claim 1, characterized in that: In step S100, the extraction of style features specifically includes: For the text content, extract linguistic features and calculate the deviation vector Style_text between it and multiple preset human author style templates; For image content, extract its color distribution histogram, texture features, and frequency domain features, and calculate its similarity vector Style_image with the image style in the preset AI image style library; For audio content, extract its Mel frequency cepstral coefficients, fundamental frequency profile, and spectral centroid, and calculate its similarity vector Style_audio with the output audio of various preset speech synthesis models; The above vectors are normalized and concatenated to form a unified style feature vector V_style. 3.The AI creation monitoring method based on big data processing according to claim 2, characterized in that: In step S100, the extraction of semantic features specifically includes: The content to be monitored is subjected to deep semantic encoding using a pre-trained large language model or vision model to obtain a high-dimensional semantic embedding vector V_semantic. Calculate the cosine similarity between V_semantic and multiple typical vectors in a pre-defined AI-generated semantic common pattern library to form a semantic consistency vector V_consistency; the AI-generated semantic common pattern library contains a set of vectors encoded from known AI-generated content. The semantic features are composed of V_semantic and V_consistency.

4. The AI creation monitoring method based on big data processing according to claim 3, characterized in that: In step S100, the extraction of statistical features specifically includes: For text content, calculate its information entropy, character repetition rate, and outliers in the frequency of use of specific punctuation marks; For image content, calculate the local correlation of pixel values ​​and the statistics of compression artifacts; For audio / video content, calculate its signal-to-noise ratio and outlier values ​​of coding features; After normalizing the above statistics, a statistical feature vector V_statistical is formed.

5. The AI creation monitoring method based on big data processing according to claim 4, characterized in that: In step S100, the extraction of code features targets executable files or script-like content, specifically including: Analyze its code structure and extract features from the function call graph, control flow graph, and referenced external libraries; Calculate the structural similarity between the stated code structure and the output code of known AI code generation tools; This forms the code feature vector V_code.

6. The AI creation monitoring method based on big data processing according to claim 1, characterized in that: In step S200, the multivariate classification model is trained in the following manner: Collect labeled human-created content and AI-created content as training samples. The labeling information includes at least: whether it is AI-generated, the specific type of AI creation, and the AI ​​tool or model used. The initial deep learning model is trained using the training sample set. The initial deep learning model adopts a multi-task learning architecture, and its output layer includes at least: a binary classification output head for judging the probability of AI generation, a multi-classification output head for distinguishing the type of creation, and a multi-classification output head for identifying the source of generation. In the comprehensive judgment result, P_ai comes from the binary classification output head, T_type comes from the multi-classification output head, and S_source comes from the generation source identification output head.

7. The AI ​​creation monitoring method based on big data processing according to claim 1, characterized in that: In step S200, the calculation of the source credibility identifier S_source also introduces a credibility evaluation function: C = f(P_ai, P_type, P_source) = α×P_ai +β×P_type +γ×P_source, where P_ai is the AI ​​generation probability, P_type is the model's confidence in the decision type T_type at the multi-class output head, P_source is the model's confidence in the decision source S_source at the generation source identification output head, and α, β, γ are adjustable weight coefficients, and α+β+γ=1; When C is below the first threshold θ_low, S_source is marked as "uncertain"; when C is above the second threshold θ_high, S_source is output according to the model's judgment result; when C is between θ_low and θ_high, the low confidence processing procedure is triggered.

8. The AI creation monitoring method based on big data processing according to claim 7, characterized in that: In step S300, the dynamic monitoring strategy library contains multiple strategy rules, each consisting of a trigger condition and a monitoring action; the trigger condition is a logical expression based on P_ai, T_type, and S_source; the monitoring action includes multiple levels: Level L1: When P_ai < δ1, execute action A1: record the content ID and judgment result in the log, and perform no further operations; Level L2: When δ1 ≤ P_ai < δ2 and S_source is not "uncertain", execute action A2: record detailed information and generate a mild alert notification; Level L3: When P_ai ≥ δ2, or T_type belongs to the high-risk type list, execute action A3: intercept or restrict the spread of the content, and initiate the deep source tracing analysis subprocess; Level L4: When P_ai ≥ δ2 and S_source points to an unauthorized or malicious AI model, perform action A4: immediately isolate the content, generate a high-risk alert, and attempt to trace the uploader's identity. 9.The AI creation monitoring method based on big data processing of claim 1, wherein: Step S400 further includes the following steps: The collected manual review records, user reports and feedback, and in-depth source tracing analysis results generated during the monitoring process constitute a labeled feedback dataset D_feedback; The multivariate classification model is incrementally trained periodically using D_feedback to adjust the model parameters; Analyze the cases in D_feedback that are judged incorrectly, and reversely adjust the feature extraction focus or structure of the multimodal feature extraction model.

10. The AI creation monitoring method based on big data processing according to any one of claims 1 to 9, characterized in that: The method is applied to an AI-driven content creation monitoring system based on big data processing, the system comprising: The multi-dimensional feature extraction module is used to parse the content to be monitored, extract its multi-dimensional features, and generate corresponding standardized feature vectors through a multi-modal feature extraction model deployed on the monitoring server; wherein, the content to be monitored is digital content in text, image, audio, or video format; the multi-dimensional features include at least style features, semantic features, statistical features, and code features; The multivariate classification decision module inputs the standardized feature vector into a pre-trained multivariate classification model. The multivariate classification model outputs a comprehensive judgment result for the content to be monitored. The comprehensive judgment result includes an AI generation probability value, and if it is generated by AI, it further includes the creation type identifier and the credibility identifier of the generation source. The execution module is used to query a preset dynamic monitoring strategy library based on the comprehensive judgment result, determine and execute the corresponding monitoring action; the level of the monitoring action is dynamically adjusted according to the combination of AI generation probability value, creation type identifier and generation source credibility identifier; The closed-loop feedback module is used to collect the execution results of the monitoring actions, form a feedback dataset, and use the feedback dataset to incrementally update and optimize the multimodal feature extraction model and the multivariate classification model.

Images