Private network content copyright monitoring and evidence obtaining system based on AI and large model

Abstract

The invention discloses a private network content copyright monitoring and evidence obtaining system based on AI and a large model, which utilizes AI and large model technologies to carry out copyright monitoring and evidence obtaining on multimedia content in a private network environment, and carries out deep semantic understanding and cross-modal feature extraction through a large model processor to generate unified semantic representation. The method comprises the steps that a copyright content database is established and stored in a copyright content knowledge base, the copyright content knowledge base is used for storing metadata of original content protected by copyright, unified semantic representation and copyright declarations in a natural language form provided by a copyright party, and the copyright declarations are converted into query vectors through a semantic understanding technology; the method comprises the following steps: acquiring a semantic representation of a multimedia content, performing similarity calculation with the semantic representation of the multimedia content, identifying infringement content, when the infringement content is identified, recording an original source, publishing time and publisher information of the infringement content, performing differentiation analysis, generating an evidence chain, displaying a copyright monitoring result through a user interface, generating infringement alarm information, and presenting details of the evidence chain.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a private network content copyright monitoring and evidence collection system based on AI and large models. Background Technology

[0002] With the rapid development of internet technology and the explosive growth of digital content, private networks (PVAs) have become an important carrier for information flow and content creation. Content within PVAs includes, but is not limited to, corporate documents, internal training materials, industry reports, business presentations, and customer data, forming a large amount of highly concentrated sensitive information and intellectual property. However, as the use of PVAs expands and diversifies, especially when content is shared and collaborated among multiple departments and employees, copyright infringement issues are becoming increasingly serious.

[0003] Traditional content copyright protection mechanisms on private networks typically rely on technologies such as digital watermarking, content fingerprinting, access control, and permission management. While these technologies can limit the illegal copying and dissemination of content to some extent, their limitations are becoming increasingly apparent with the diversification and complexity of user-generated content and the mixed editing of multimedia content such as video, audio, and text. This is especially true in private network environments, where frequent changes, modifications, and reorganizations of content make precise matching-based monitoring methods ineffective. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a private network content copyright monitoring and evidence collection system based on AI and large models to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a private network content copyright monitoring and evidence collection system based on AI and large models, specifically including a content acquisition module, a multimodal content analysis module, a copyright content knowledge base, a zero-sample copyright identification module, an infringement behavior evidence collection module, and a visualization and alarm module;

[0006] The content acquisition module is used to acquire the multimedia content to be monitored from the private network environment, including text, images, audio, video and any mixture thereof;

[0007] The multimodal content analysis module uses a large model processor to perform deep semantic understanding and cross-modal feature extraction on the acquired multimedia content, so as to generate a unified semantic representation of the multimedia content and store it in the copyright content knowledge base.

[0008] The copyright content knowledge base is used to store metadata of the original content protected by copyright, a unified semantic representation generated by the multimodal content analysis module, and a copyright statement in natural language form provided by the copyright holder.

[0009] The zero-sample copyright identification module receives a natural language copyright statement from the copyright holder, uses the large model processor to perform semantic understanding on the natural language copyright statement, converts it into a semantic query vector, and performs semantic similarity calculation and matching between the semantic query vector and the unified semantic representation generated by the multimodal content analysis module to identify infringing content.

[0010] The infringement evidence collection module records the original source, publication time, publisher identity information, modification history, semantic similarity score and matching area of ​​the infringing content when the zero-sample copyright recognition module identifies infringing content. It then performs differential analysis on the infringing content and the original copyright content through a large model processor and generates a chain of evidence.

[0011] The visualization and alerting module provides a user interface to display real-time copyright monitoring results, generate infringement alert information, present complete evidence chain details, and allow users to manage copyright statements and query historical monitoring data.

[0012] Preferably, as a preferred embodiment of the AI-based and large-scale model-based private network content copyright monitoring and evidence collection system of the present invention, it includes a content acquisition module for acquiring multimedia content to be monitored from the private network environment, including text, images, audio, video, and any mixture thereof, specifically including the following:

[0013] For static resources of file servers, databases and internal websites deployed within the private network, a targeted crawler is used to crawl the content according to a preset strategy in order to obtain the multimedia files and data contained therein.

[0014] Traffic mirroring devices are deployed at key network nodes of the private network, and the mirrored traffic is imported into a sniffer. The sniffer is capable of performing deep packet inspection on data packets circulating within the private network, and reconstructing the multimedia content of files, email attachments, and instant messaging messages that were encapsulated during transmission.

[0015] Preferably, as a preferred embodiment of the AI- and large-scale model-based private network content copyright monitoring and evidence collection system of the present invention, the multimodal content analysis module utilizes a large-scale model processor to perform deep semantic understanding and cross-modal feature extraction on the acquired multimedia content to generate a unified semantic representation of the multimedia content and store it in a copyright content knowledge base, specifically including the following:

[0016] The acquired multimedia content is subjected to modal recognition and preliminary analysis, and is decomposed into text modal sub-content. Image modal sub-content Audio modal sub-content Video modal sub-content For each identified modal sub-content, a preliminary modal feature vector is generated through the corresponding encoder. , , , , representing the modal feature vectors of text modality, image modality, audio modality, and video modality, respectively, further including:

[0017] Text modality: This involves word segmentation and word embedding, using a pre-trained word embedding model to convert the text sequence into an embedding sequence. And a shallow Transformer encoder is used to generate a context-aware preliminary modal feature vector, represented as follows: ;

[0018] Image modality: Perform size normalization and pixel value standardization, and scale image I to a fixed size. After normalization, a lightweight convolutional neural network is used to extract its visual features, and the initial modality feature vector is represented as... Where H, W, and D represent height, width, and depth, respectively;

[0019] Audio modality: A spectrogram is generated by unifying the sampling rate and performing a short-time Fourier transform. This spectrogram is treated as an image, and its features are extracted using a lightweight CNN similar to that used for image modality. The initial modality feature vector is represented as follows: ;

[0020] Video modality: This refers to the video... Decomposed into frame sequences And extract the accompanying audio. For each frame Generate its initial feature vector For accompanying audio Its features are obtained by processing according to audio modality. The initial feature vector of a video modality is represented as a set of frame feature sequences. ,in, This represents the audio feature vector extracted from the audio track of the video file;

[0021] The large model processor is a large-scale pre-trained model based on the Transformer architecture. It utilizes self-attention mechanisms and feedforward neural networks to pre-train on massive amounts of text, image, audio, and video data through self-supervised learning. Based on preliminary modal feature vectors, it generates a unified semantic representation, further including:

[0022] Preliminary feature vectors of different modalities , , , It maps to the embedding space shared by the large model through a linear projection layer, and adds positional encoding P to form a token sequence for the model input. ,in, This represents the operation of mapping the initial feature vectors of each modality to a shared embedding space through a linear projection layer;

[0023] The projected token sequence The Transformer encoder fed into the model dynamically calculates the association weight between any two tokens in the sequence through its self-attention mechanism, using the following formula: Where Q, K, and V are derived from the input token sequence The query, key, and value matrices obtained through different linear transformation matrices The dimension of the key vector is used to scale the dot product result;

[0024] After processing by multiple Transformer encoders, the large model processor outputs a final hidden state sequence. Pooling is then performed on this sequence to generate a fixed-dimensional unified semantic representation. This unified semantic representation, along with the metadata of the original multimedia content, is stored in a copyright knowledge base. For copyrighted original content, the unified semantic representation is marked as... The copyright content knowledge base stores copyright statements in natural language form provided by the copyright holder.

[0025] Preferably, as a preferred embodiment of the AI- and large-scale model-based private network content copyright monitoring and evidence collection system of the present invention, the zero-sample copyright identification module receives a natural language copyright statement from the copyright holder, uses the large-scale model processor to perform semantic understanding on the natural language copyright statement, converts it into a semantic query vector, and performs semantic similarity calculation and matching between the semantic query vector and the unified semantic representation generated by the multimodal content analysis module to identify infringing content. Specifically, it includes the following:

[0026] Representing the copyright statement in natural language form provided by the copyright holder as a lexical sequence ,in, It consists of the tokens in the copyright statement, and performs text preprocessing on the copyright statement, including word segmentation, removal of stop words, and spell normalization, to form a standardized text sequence;

[0027] The preprocessed copyright notice is subjected to deep semantic encoding using a large model processor. The text sequence is taken as input and fed into the text input terminal of the large model processor. Through the self-attention mechanism in its Transformer architecture, it captures the contextual semantic information of the copyright statement and outputs a fixed-dimensional text. semantic query vector ;

[0028] The generated semantic query vector With the unified semantic representation generated by the multimodal content analysis module in the copyright content knowledge base Perform comparisons and calculate unified semantic representations. With semantic query vector In the shared The semantic similarity in the 3D semantic space is calculated using the following formula:

[0029] ,in, Represents the semantic similarity function. It is the i-th unified semantic representation The j-th dimension component, It is a semantic query vector The j-th dimension component is calculated, and the calculated similarity is compared with a preset similarity threshold. When comparing, The system determines that the content to be monitored generated by the multimodal content analysis module is highly semantically related to the original content described in the copyright statement, thus constituting infringement; by adjusting the similarity threshold... It can balance recall and precision;

[0030] Based on the matching results, all content to be monitored that is highly similar to the copyright statement in semantics is identified, and the list of identified infringing content is output. The list of infringing content includes its metadata, unified semantic representation, and similarity score with the copyright statement.

[0031] Preferably, as a preferred embodiment of the AI- and large-scale model-based private network content copyright monitoring and evidence collection system of the present invention, the infringement evidence collection module records the original source, publication time, publisher identity information, modification history, semantic similarity score and matching area of ​​the infringing content when the zero-sample copyright identification module identifies infringing content. The large-scale model processor then performs differential analysis on the infringing content and the original copyright content and generates an evidence chain, specifically including the following:

[0032] The system automatically captures and records full-dimensional metadata related to infringing content and integrates it into an electronic evidence archive associated with a unique identifier for the infringing content. This electronic evidence archive includes basic identification information, time and identity information, content comparison information, and core feature data. The basic identification information includes the unique identifier of the identified infringing content and the specific network link when the infringing content was discovered. The time and identity information includes the publication time, last modification time, and the publisher's / creator's private network identity information. The content comparison information includes a semantic similarity score between the infringing content and the copyright statement; for image and video content, it records the coordinates of local areas highly similar to the original content; for text content, it records the start and end positions of similar paragraphs. The core feature data includes a unified semantic representation vector of the infringing content and preliminary feature vectors for each modality.

[0033] The differential analysis compares the subtle differences between infringing content and original copyrighted content within their respective modalities, identifying semantic substitution, sentence restructuring, and paragraph deletion in text; detecting watermark removal, partial occlusion, element addition / deletion, and color adjustment in images; analyzing audio cutting, splicing, and track separation; and combining network traffic logs and access control lists within the private network to analyze the potential propagation path of infringing content within the private network and identify secondary propagation nodes. When different versions of infringing content related to the same copyrighted content exist, a content derivation relationship graph is constructed by comparing their unified semantic representations and preliminary feature vectors, visually demonstrating the evolution process of infringing content and the kinship between different variants.

[0034] All the above information is automatically combined into a standardized electronic evidence chain report, and blockchain technology is used to embed an immutable digital timestamp for each key event and data record in the evidence chain.

[0035] Preferably, as a preferred embodiment of the AI-based and large-scale model-based private network content copyright monitoring and evidence collection system of the present invention, the visualization and alarm module provides a user interface to display real-time copyright monitoring results, generate infringement alarm information, present complete evidence chain details, and allow users to manage copyright statements and query historical monitoring data, specifically including the following:

[0036] The newly identified infringing content is displayed in a list format. Each record includes a thumbnail of the infringing content, the type of infringement, the similarity score, the discovery time, and the initial source. It also supports quick jump to the detailed evidence chain page and uses charts to show the trend of the number of infringements over time, the distribution of infringements by different content types and platforms, to help users understand the dynamics and patterns of infringement.

[0037] Based on user-defined alarm rules, when the similarity exceeds a preset threshold, an infringement alarm is automatically generated. The alarm is promptly notified to authorized users via in-site messages, emails, SMS, and API callbacks. When a user clicks on an infringement alarm record, they are navigated to a detailed page that fully presents the electronic evidence archive generated by the infringement evidence collection module. All metadata in the electronic evidence archive is displayed in a clear and easy-to-read format, including: basic identification information, time and identity information, content comparison information, and core feature data. The page also embeds a propagation path diagram and a content derivative relationship diagram, allowing users to interact with the graph, click on nodes to view details, and intuitively understand the spread path and variant relationships of the infringing content.

[0038] On the other hand, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements the functional modules of a private network content copyright monitoring and evidence collection system based on AI and large models as described above.

[0039] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements an AI-based and large-scale model-based private network content copyright monitoring and evidence collection system as described above.

[0040] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0041] This invention eliminates the need for a pre-stored infringement sample library. It can automatically generate semantic query vectors simply through natural language declarations, enabling flexible and dynamic infringement monitoring. When infringing content is identified, it integrates metadata, propagation path analysis, and content derivation graphs to generate a structured evidence chain with timestamps and blockchain storage. This achieves automated solidification of complete evidence in infringement cases, ensuring the immutability of the evidence. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0043] Figure 1 This is a flowchart of a method for monitoring and obtaining evidence of content copyright on a private network based on AI and large models, according to the present invention.

[0044] Table 1 is a data record table of the simulation experiment of a private network content copyright monitoring and evidence collection system based on AI and large model according to the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0047] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0048] Example 1

[0049] This embodiment provides, for example Figure 1 The system described is a private network content copyright monitoring and evidence collection system based on AI and large models. Specifically, it includes a content acquisition module, a multimodal content analysis module, a copyright content knowledge base, a zero-sample copyright identification module, an infringement behavior evidence collection module, and a visualization and alarm module.

[0050] The content acquisition module is used to acquire the multimedia content to be monitored from the private network environment, including text, images, audio, video and any mixture thereof;

[0051] The multimodal content analysis module uses a large model processor to perform deep semantic understanding and cross-modal feature extraction on the acquired multimedia content, so as to generate a unified semantic representation of the multimedia content and store it in the copyright content knowledge base.

[0052] The copyright content knowledge base is used to store metadata of the original content protected by copyright, a unified semantic representation generated by the multimodal content analysis module, and a copyright statement in natural language form provided by the copyright holder.

[0053] The zero-sample copyright identification module receives a natural language copyright statement from the copyright holder, uses the large model processor to perform semantic understanding on the natural language copyright statement, converts it into a semantic query vector, and performs semantic similarity calculation and matching between the semantic query vector and the unified semantic representation generated by the multimodal content analysis module to identify infringing content.

[0054] The infringement evidence collection module records the original source, publication time, publisher identity information, modification history, semantic similarity score and matching area of ​​the infringing content when the zero-sample copyright recognition module identifies infringing content. It then performs differential analysis on the infringing content and the original copyright content through a large model processor and generates a structured, traceable, and timestamped evidence chain.

[0055] The visualization and alerting module provides a user interface to display real-time copyright monitoring results, generate infringement alert information, present complete evidence chain details, and allow users to manage copyright statements and query historical monitoring data.

[0056] In this embodiment, the content acquisition module is specifically described. This module is used to acquire the multimedia content to be monitored from the private network environment, including text, images, audio, video, and any mixture thereof. Specifically, it includes the following:

[0057] For static resources of file servers, databases and internal websites deployed within the private network, a targeted crawler is used to crawl the content according to a preset strategy in order to obtain the multimedia files and data contained therein.

[0058] Traffic mirroring devices are deployed at key network nodes of the private network, and the mirrored traffic is imported into a sniffer. The sniffer is capable of performing deep packet inspection on data packets circulating within the private network, and reconstructing the multimedia content of files, email attachments, and instant messaging messages that were encapsulated during transmission.

[0059] In this embodiment, the multimodal content analysis module is specifically described. This module utilizes a large model processor to perform deep semantic understanding and cross-modal feature extraction on the acquired multimedia content to generate a unified semantic representation of the multimedia content, which is then stored in a copyright content knowledge base. Specifically, it includes the following:

[0060] The acquired multimedia content is subjected to modal recognition and preliminary analysis, and is decomposed into text modal sub-content. Image modal sub-content Audio modal sub-content Video modal sub-content For each identified modal sub-content, a preliminary modal feature vector is generated through the corresponding encoder. , , , , representing the modal feature vectors of text modality, image modality, audio modality, and video modality, respectively, further including:

[0061] Text modality: This involves word segmentation and word embedding, using a pre-trained word embedding model to convert the text sequence into an embedding sequence. And a shallow Transformer encoder is used to generate a context-aware preliminary modal feature vector, represented as follows: ;

[0062] Image modality: Perform size normalization and pixel value standardization, and scale image I to a fixed size. After normalization, a lightweight convolutional neural network is used to extract its visual features, and the initial modality feature vector is represented as... Where H, W, and D represent height, width, and depth, respectively;

[0063] Audio modality: A spectrogram is generated by unifying the sampling rate and performing a short-time Fourier transform. This spectrogram is treated as an image, and its features are extracted using a lightweight CNN similar to that used for image modality. The initial modality feature vector is represented as follows: ;

[0064] Video modality: This refers to the video... Decomposed into frame sequences And extract the accompanying audio. For each frame Generate its initial feature vector For accompanying audio Its features are obtained by processing according to audio modality. The initial feature vector of a video modality is represented as a set of frame feature sequences. ,in, This represents the audio feature vector extracted from the audio track of the video file;

[0065] The large model processor is a large-scale pre-trained model based on the Transformer architecture. It utilizes self-attention mechanisms and feedforward neural networks to pre-train on massive amounts of text, image, audio, and video data through self-supervised learning. Based on preliminary modal feature vectors, it generates a unified semantic representation, further including:

[0066] Preliminary feature vectors of different modalities , , , It maps to the embedding space shared by the large model through a linear projection layer, and adds positional encoding P to form a token sequence for the model input. ,in, This represents the operation of mapping the initial feature vectors of each modality to a shared embedding space through a linear projection layer;

[0067] The projected token sequence The Transformer encoder fed into the model dynamically calculates the association weight between any two tokens in the sequence through its self-attention mechanism, using the following formula: Where Q, K, and V are derived from the input token sequence The query, key, and value matrices obtained through different linear transformation matrices The dimension of the key vector is used to scale the dot product result;

[0068] After processing by multiple Transformer encoders, the large model processor outputs a final hidden state sequence. Pooling is then performed on this sequence to generate a fixed-dimensional unified semantic representation. This unified semantic representation, along with the metadata of the original multimedia content, is stored in a copyright knowledge base. For copyrighted original content, the unified semantic representation is marked as... The copyright content knowledge base stores copyright statements in natural language form provided by the copyright holder.

[0069] In this embodiment, the zero-sample copyright identification module is specifically described. This module receives a natural language copyright statement from the copyright holder, uses the large model processor to perform semantic understanding on the natural language copyright statement, converts it into a semantic query vector, and then calculates and matches the semantic query vector with the unified semantic representation generated by the multimodal content analysis module to identify infringing content. Specifically, this includes the following:

[0070] Representing the copyright statement in natural language form provided by the copyright holder as a lexical sequence ,in, It consists of the tokens in the copyright statement, and performs text preprocessing on the copyright statement, including word segmentation, removal of stop words, and spell normalization, to form a standardized text sequence;

[0071] The preprocessed copyright notice is subjected to deep semantic encoding using a large model processor. The text sequence is taken as input and fed into the text input terminal of the large model processor. Through the self-attention mechanism in its Transformer architecture, it captures the contextual semantic information of the copyright statement and outputs a fixed-dimensional text. semantic query vector ;

[0072] The generated semantic query vector With the unified semantic representation generated by the multimodal content analysis module in the copyright content knowledge base Perform comparisons and calculate unified semantic representations. With semantic query vector In the shared The semantic similarity in the 3D semantic space is calculated using the following formula:

[0073] ,in, Represents the semantic similarity function. It is the i-th unified semantic representation The j-th dimension component, It is a semantic query vector The j-th dimension component is calculated, and the calculated similarity is compared with a preset similarity threshold. When comparing, The system determines that the content to be monitored generated by the multimodal content analysis module is highly semantically related to the original content described in the copyright statement, thus constituting infringement; by adjusting the similarity threshold... It can balance recall and precision;

[0074] Based on the matching results, all content to be monitored that is highly similar to the copyright statement in semantics is identified, and the list of identified infringing content is output. The list of infringing content includes its metadata, unified semantic representation, and similarity score with the copyright statement.

[0075] In this embodiment, the infringement evidence collection module is specifically described. When the zero-sample copyright identification module identifies infringing content, this module records the original source, publication time, publisher's identity information, modification history, semantic similarity score, and matching region of the infringing content. A large model processor then performs differential analysis on the infringing content and the original copyright content, generating a structured, traceable, and timestamped chain of evidence. Specifically, this includes the following:

[0076] The system automatically captures and records full-dimensional metadata related to infringing content and integrates it into an electronic evidence archive associated with a unique identifier for the infringing content. This electronic evidence archive includes basic identification information, time and identity information, content comparison information, and core feature data. The basic identification information includes the unique identifier of the identified infringing content and the specific network link when the infringing content was discovered. The time and identity information includes the publication time, last modification time, and the publisher's / creator's private network identity information. The content comparison information includes a semantic similarity score between the infringing content and the copyright statement; for image and video content, it records the coordinates of local areas highly similar to the original content; for text content, it records the start and end positions of similar paragraphs. The core feature data includes a unified semantic representation vector of the infringing content and preliminary feature vectors for each modality.

[0077] The differential analysis compares the subtle differences between infringing content and original copyrighted content within their respective modalities, identifying semantic substitution, sentence restructuring, and paragraph deletion in text; detecting watermark removal, partial occlusion, element addition / deletion, and color adjustment in images; analyzing audio cutting, splicing, and track separation; and combining network traffic logs and access control lists within the private network to analyze the potential propagation path of infringing content within the private network and identify secondary propagation nodes. When different versions of infringing content related to the same copyrighted content exist, a content derivation relationship graph is constructed by comparing their unified semantic representations and preliminary feature vectors, visually demonstrating the evolution process of infringing content and the kinship between different variants.

[0078] All the above information is automatically combined into a standardized electronic evidence chain report, and blockchain technology is used to embed an immutable digital timestamp for each key event and data record in the evidence chain.

[0079] In this embodiment, the visualization and alarm module needs to be specifically described. This module provides a user interface to display real-time copyright monitoring results, generate infringement alarm information, present complete evidence chain details, and allow users to manage copyright statements and query historical monitoring data. Specifically, it includes the following:

[0080] The newly identified infringing content is displayed in a list format. Each record includes a thumbnail of the infringing content, the type of infringement, the similarity score, the discovery time, and the initial source. It also supports quick jump to the detailed evidence chain page and uses charts to show the trend of the number of infringements over time, the distribution of infringements by different content types and platforms, to help users understand the dynamics and patterns of infringement.

[0081] Based on user-defined alarm rules, when the similarity exceeds a preset threshold, an infringement alarm is automatically generated. The alarm is promptly notified to authorized users via in-site messages, emails, SMS, and API callbacks. When a user clicks on an infringement alarm record, they are navigated to a detailed page that fully presents the electronic evidence archive generated by the infringement evidence collection module. All metadata in the electronic evidence archive is displayed in a clear and easy-to-read format, including: basic identification information, time and identity information, content comparison information, and core feature data. The page also embeds a propagation path diagram and a content derivative relationship diagram, allowing users to interact with the graph, click on nodes to view details, and intuitively understand the spread path and variant relationships of the infringing content.

[0082] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0083] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the functional modules of an AI-based and large-scale model-based private network content copyright monitoring and evidence collection system as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0084] Example 2

[0085] The following is another embodiment of the present invention, which provides a private network content copyright monitoring and evidence collection system based on AI and large models. In order to verify the beneficial effects of the present invention, a simulation experiment is conducted for scientific demonstration.

[0086] This experiment aims to verify the effectiveness of a dedicated network content copyright monitoring and evidence collection system based on AI and large models. Through technical means such as content acquisition, multimodal analysis, copyright knowledge base construction, zero-shot identification, infringement evidence collection, and visual alerts, the system improves the accuracy and efficiency of identifying copyright infringement of multimedia content within the dedicated network. The experiment uses simulated and actual collected multimedia content data from the dedicated network, including different types of content features, semantic similarity scores, infringement confidence levels, and multi-dimensional evidence information. By analyzing the consistency between the system's infringement identification results and actual infringement labels, the accuracy and robustness of the system in identifying and collecting copyright infringement evidence are verified.

[0087] The simulation experiment steps are implemented based on the content of the AI-based and large-scale model-based private network content copyright monitoring and evidence collection system provided in Example 1. The specific steps include:

[0088] The system collects content from different sources and modalities, including text, images, videos, and audio, in a simulated private network environment. It simulates user upload and sharing behaviors, records the original copyright content and its metadata provided by the copyright holder, as well as the copyright statement in natural language, and builds a copyright content knowledge base. The collection frequency is set to combine real-time monitoring and timed scanning.

[0089] The acquired copyrighted content and the content to be monitored are input into the multimodal content analysis module, and combined with the large model processor, deep semantic understanding of various types of content is performed through the following steps:

[0090] a. Perform modal recognition and parsing on the acquired multimedia content, decomposing it into text, image, audio, and video sub-content;

[0091] b. Generate preliminary feature vectors for each modality using the corresponding encoder;

[0092] c. Utilize cross-modal large models to map features of different modalities to a unified semantic space, generate a unified semantic representation, and store it in a copyright content knowledge base as the basis for subsequent comparisons;

[0093] The simulated copyright holder enters a new natural language copyright statement through the visualization and alarm module. The zero-sample copyright recognition module receives the statement, uses the large model processor to convert it into a semantic query vector, and performs semantic similarity calculation and matching between the semantic query vector and the unified semantic representation of all content in the copyright content knowledge base. Based on the dynamic threshold, a preliminary judgment of infringement is made.

[0094] When the zero-sample copyright identification module identifies infringing content, the infringement evidence collection module is triggered, recording the infringement evidence and generating a structured evidence chain, including:

[0095] 1) Automatically capture the metadata and core characteristics of infringing content;

[0096] 2) Analyze the dissemination paths and derivative relationships of infringing content;

[0097] 3) Generate a credible chain of evidence containing timestamps.

[0098] The specific data from the above simulation experiment are as follows:

[0099] Content type Content Description Original Copyright Notice Monitoring results semantic similarity Differentiation analysis Secondary transmission Chain of evidence status text Internal Report A This report is for internal circulation only and is strictly prohibited from being leaked or modified. No infringement N / A N / A N / A N / A picture Product design drawings "Copyright reserved. Unauthorized use or distribution is prohibited." No infringement N / A N / A N / A N / A text Internal Report A (Minor Modifications) Same as above Infringement 0.95 Minor textual modifications, semantics unchanged User A uploads -> User B downloads Generated picture Product design drawings (with watermark) Same as above Infringement 0.88 Remove the original watermark and add a new watermark. User C uploads -> User D forwards Generated video Training Video X This video may not be extracted or distributed. No infringement N / A N / A N / A N / A video Training Video X (Edited Clips) Same as above Infringement 0.75 Edit key segments and remove the beginning. Uploaded by User E Generated text Internal Report A (Extensively revised) Same as above Infringement 0.60 Extensive rewriting, but core viewpoints retained. User F uploaded Generated picture Product design drawing (blurred) Same as above Infringement 0.70 Blur processing reduces sharpness User G uploads -> User H distributes Generated text New document Y "all rights reserved." No infringement N / A N / A N / A N / A

[0100] Table 1

[0101] Experimental Analysis:

[0102] By comparing the infringing content identified by the system with preset infringing scenarios, the accuracy of the system in identifying infringement under different modalities and degrees of modification was verified. The comprehensiveness and effectiveness of the system in recording infringing metadata, conducting differential analysis, tracking dissemination paths, and constructing evidence chains were evaluated. The experiment shows that the AI-based private network content copyright monitoring and evidence collection system has high infringement identification accuracy, comprehensive evidence collection capabilities, and a reliable evidence chain generation mechanism, which can effectively improve the level of protection of private network content copyright.

[0103] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. An AI and large model-based private network content copyright monitoring and forensics system, characterized in that: Specifically, it includes a content acquisition module, a multi-modal content analysis module, a copyright content knowledge base, a zero-shot copyright identification module, a copyright infringement evidence collection module, and a visualization and alarm module. The content acquisition module is used to acquire multimedia content to be monitored from a private network environment, including text, images, audio, video and any mixed form thereof. The multi-modal content analysis module uses a large model processor to perform deep semantic understanding and cross-modal feature extraction on the acquired multimedia content to generate a unified semantic representation of the multimedia content and store it in the copyright content knowledge base. The copyright content knowledge base is used to store the metadata of the original content protected by copyright, the unified semantic representation generated by the multi-modal content analysis module, and the natural language copyright declaration provided by the copyright party. The zero-shot copyright identification module receives the natural language copyright declaration from the copyright party, uses the large model processor to perform semantic understanding on the natural language copyright declaration, converts it into a semantic query vector, and performs semantic similarity calculation and matching between the semantic query vector and the unified semantic representation generated by the multi-modal content analysis module to identify infringing content. The copyright infringement evidence collection module records the original source, publication time, publisher identity information, modification history, semantic similarity score between the infringing content and the original copyright content, and matching area of the infringing content when the zero-shot copyright identification module identifies the infringing content, and generates an evidence chain through differential analysis of the infringing content and the original copyright content by the large model processor. The visualization and alarm module provides a user interface to display real-time copyright monitoring results, generate infringement alarm information, present complete evidence chain details, and allow users to manage copyright declarations and query historical monitoring data.

2. The AI and large model-based private network content copyright monitoring and evidence collection system according to claim 1, characterized in that: The content acquisition module is used to acquire multimedia content to be monitored from a private network environment, including text, images, audio, video and any mixed form thereof, specifically including the following content: For file servers, databases and internal websites deployed in the private network, use a directed crawler program to perform content crawling according to a preset strategy to obtain multimedia files and data contained therein. Deploy traffic mirroring devices at key network nodes of the private network and import the mirrored traffic into a sniffer, which can perform deep packet analysis on data packets circulating in the private network to reassemble multimedia content in files, email attachments and instant messaging messages that have been encapsulated during transmission.

3. The AI and large model-based private network content copyright monitoring and evidence collection system according to claim 1, characterized in that: The multi-modal content analysis module uses a large model processor to perform deep semantic understanding and cross-modal feature extraction on the acquired multimedia content to generate a unified semantic representation of the multimedia content and store it in the copyright content knowledge base, specifically including the following content: modal recognition and preliminary analysis are performed on the obtained multimedia content, and the multimedia content is decomposed into text modal sub-content , image modal sub-content , audio modal sub-content , and video modal sub-content For each modal sub-content identified, a preliminary modal feature vector is generated through a corresponding encoder 、 、 、 , respectively representing the modal feature vectors of the text modal, image modal, audio modal, and video modal The large model processor is a large-scale pre-training model based on the Transformer architecture, which uses self-attention mechanisms and feedforward neural networks to pre-train on massive amounts of text, images, audio and video data through self-supervised learning. According to the preliminary modal feature vector, a unified semantic representation is generated, which further includes: Preliminary feature vectors of different modalities , , , It maps to the embedding space shared by the large model through a linear projection layer, and adds positional encoding P to form a token sequence for the model input. ,in, This represents the operation of mapping the initial feature vectors of each modality to a shared embedding space through a linear projection layer; The projected token sequence is fed into the Transformer encoder of the model, which dynamically computes the correlation weight between any two tokens in the sequence through its self-attention mechanism, with the specific formula being where Q, K, and V are the query, key, and value matrices obtained by different linear transformation matrices on the input token sequence , is the dimension of the key vector, used to scale the dot product result; After being processed by the multi-layer Transformer encoder, the large model processor outputs a final hidden state sequence, generates a unified semantic representation by performing a pooling operation on the sequence, and stores the generated unified semantic representation and metadata of the original multimedia content in a copyright content knowledge base. For the original content protected by copyright, the unified semantic representation is marked as , and the copyright content knowledge base also stores the copyright declaration in natural language form provided by the copyright party.

4. The AI and large model-based private network content copyright monitoring and evidence collection system according to claim 3, characterized in that: The preliminary modal feature vectors are generated by corresponding encoders , , , , respectively represent the modal feature vectors of the text modal, the image modal, the audio modal, and the video modal, further comprising: Text modality: perform word segmentation, word vector embedding, use a pre-trained word embedding model to convert the text sequence into an embedding sequence and generate a context-aware preliminary modality feature vector through a shallow Transformer encoder, denoted as ; Image modalities: size normalization, pixel value standardization, scaling the image I to a fixed size and normalization, using a lightweight convolutional neural network to extract its visual features, and representing the preliminary modal feature vector as where H, W, D represent height, width, and depth, respectively. Audio modality: do sampling rate unification, short-time Fourier transform to generate spectrogram, take the spectrogram as an image, use a lightweight CNN similar to the image modality to extract its features, and express the preliminary modal feature vector as ; Video modality: the video is decomposed into a sequence of frames and the accompanying audio is extracted For each frame a preliminary feature vector is generated The accompanying audio is processed in the audio modality to obtain its features The preliminary feature vector of the video modality is represented as a set of frame feature vectors where denotes the audio feature vector extracted from the audio track of the video file.

5. The AI and large model-based private network content copyright monitoring and evidence collection system according to claim 1, characterized in that: The zero-shot copyright identification module receives natural language copyright statements from copyright owners, uses the large model processor to perform semantic understanding on the natural language copyright statements, converts them into semantic query vectors, and performs semantic similarity calculation and matching with the unified semantic representation generated by the multi-modal content analysis module to identify infringing content, including the following: representing a natural language form copyright statement provided by a copyright party as a word sequence wherein, is a word in the copyright statement, and text preprocessing is performed on the copyright statement, including word segmentation, stop word removal, and spelling normalization operations, to form a standardized text sequence; The preprocessed copyright notice is subjected to deep semantic encoding using a large model processor. The text sequence is taken as input and fed into the text input terminal of the large model processor. Through the self-attention mechanism in its Transformer architecture, it captures the contextual semantic information of the copyright statement and outputs a fixed-dimensional text. semantic query vector ; generated semantic query vector unified semantic representation generated by the multi-modal content analysis module in the copyright content knowledge base performing a comparison, calculating the unified semantic representation with the semantic query vector in the shared semantic similarity in the shared semantic space, specifically calculated as: ,in, Represents the semantic similarity function. It is the i-th unified semantic representation The j-th dimension component, It is a semantic query vector The j-th dimension component is calculated, and the calculated similarity is compared with a preset similarity threshold. When comparing, The multimodal content analysis module determined that the content to be monitored was highly semantically related to the original content described in the copyright statement, thus constituting infringement. According to the obtained matching results, all content to be monitored with high semantic similarity to the copyright statement is identified, and the identified infringing content list is taken as the output, including the unified semantic representation, the similarity score with the copyright statement.

6. The AI and large model-based private network content copyright monitoring and evidence collection system according to claim 1, characterized in that: The infringement evidence module records the original source, publication time, publisher identity information, modification history, semantic similarity score with the original copyright content, and matching area of the infringing content when the zero-shot copyright identification module identifies the infringing content, performs differential analysis on the infringing content and the original copyright content through the large model processor, and generates an evidence chain, including the following: Automatically capture and record all-dimensional metadata related to the infringing content and integrate them into an electronic evidence archive associated with a unique identifier of the infringing content, including basic identification information, time and identity information, content comparison information, and core feature data, wherein the content comparison information includes the semantic similarity score between the infringing content and the copyright statement, and for image and video content, records the coordinates of the local area highly similar to the original content, and for text content, records the start and end positions of similar paragraphs; the core feature data includes the unified semantic representation vector of the infringing content and its preliminary feature vector of each modality; The differential analysis is to compare the subtle differences between the infringing content and the original copyright content in their respective modalities, identify semantic replacement, sentence reorganization, and paragraph deletion of text, detect watermark removal, local occlusion, element addition and deletion, color adjustment in images, analyze audio cutting, splicing, and audio track separation, and analyze the propagation path of the infringing content in the private network by combining the network traffic logs and access control lists in the private network to identify secondary propagation nodes; when there are different versions of infringing content related to the same copyright content, a content derivation relationship graph is constructed by comparing their unified semantic representations and preliminary feature vectors to visually display the evolution process of the infringing content; All the above information is automatically integrated into a standardized electronic evidence chain report, and a non-tamperable digital timestamp is embedded for each key event and data record in the evidence chain using blockchain technology.

7. The AI and large model-based private network content copyright monitoring and forensic system according to claim 1, characterized in that: The visualization and alarm module provides a user interface to display real-time copyright monitoring results, generate infringing alarm information, present complete evidence chain details, and allow users to manage copyright statements and query historical monitoring data, including the following: Display the latest identified infringement content in a list format, each record including infringement content thumbnail, infringement type, similarity score, discovery time, preliminary source, and support quick jump to detailed evidence chain page, show the trend of infringement quantity over time, different content types, and platform distribution through charts, help users understand the infringement dynamics and patterns; According to the user's preset alarm rules, when the similarity is higher than the preset threshold, automatically generate infringement alarm information, and timely notify the authorized user through the intranet letter, email, short message and API callback mode, when the user clicks the infringement alarm record, navigate to a detailed page, complete the presentation of the electronic evidence archive generated by the infringement behavior evidence module, display all metadata in the electronic evidence archive, including: basic identification information, time and identity information, content comparison information, core feature data, and embed the propagation path graph and content derivative relationship graph in the page, users can interact with the graph, click the node to view details.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The processor executes the computer program to realize the function modules of the AI and large model based private network content copyright monitoring and evidence system of any one of claims 1-7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the function modules of the AI and large model based private network content copyright monitoring and evidence system of any one of claims 1-7.

Images