Method and device for verifying the security of traffic data content
By collecting multi-source heterogeneous data and generating a multimodal feature set, and using gradient boosting trees and large language models to verify the authenticity and credibility of the identification information, the consistency check between content and identification is performed, which solves the problem of low efficiency of manual review in existing technologies and realizes automated and accurate content security verification.
Patent Information
- Application Number
- CN202611270812.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing content security verification technologies rely excessively on manual review, resulting in low efficiency and insufficient accuracy. They are unable to identify label abuse and accurately manage boundary-sensitive content, leading to regulatory loopholes and compliance risks.
Collect heterogeneous data from multiple sources, generate a multimodal feature set, verify the authenticity and credibility of the identification information through gradient boosting trees and large language models, perform consistency checks between content and identification, and build a multi-layer intelligent verification model for automated analysis.
Significantly improves verification efficiency and accuracy, identifies label abuse, achieves consistency judgment between content and label, and provides structured reports to support precise management.
Smart Images

Figure CN123122902A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of content security, specifically to a method and apparatus for verifying the security of traffic data content. Background Technology
[0002] With the rapid development of internet technology, the amount of online data has exploded, and the content formats on various online platforms have become increasingly diverse, including articles, images, short videos, long videos, live streams, and advertising materials. To facilitate content classification, recommendation, distribution, and compliance review, platforms typically assign specific labels (tags) to various types of content, such as "News," "Education," "Entertainment," "Finance," "Advertisement," and "Official Release." Regulatory authorities and platforms need to conduct security checks on this labeled content to ensure it complies with laws, regulations, and platform management standards, preventing malicious actors from using false labels (tag abuse) to spread harmful information and evade regulatory review. However, existing content security verification technologies have the following significant shortcomings: (1) Over-reliance on manual review: The current mainstream content security verification method still relies mainly on manual review. Faced with massive network traffic data, manual review suffers from serious manpower shortages, low review efficiency, inconsistent review standards, susceptibility to subjective factors, and the omission of a large amount of illegal content, forming serious regulatory loopholes and failing to meet the needs of real-time and large-scale verification.
[0003] (2) Outdated and limited technical methods: Existing auxiliary review tools mostly use simple keyword filtering or image feature comparison techniques. For example, text content review only triggers an alarm by matching a preset prohibited word database, and image or video review only intercepts content by comparing feature similarity with a known prohibited image database. Such methods are mechanical and rigid, and can only achieve surface feature matching. They cannot understand the contextual semantics, deeper meaning, and emotional tendency of the content, leading to a large number of misjudgments (such as normal text being mistakenly blocked because it contains prohibited word fragments) and missed judgments (such as prohibited content circumventing keyword filtering by replacing synonyms or rewriting homophones). (3) Inability to handle complex scenarios such as label abuse: In order to evade review, criminals often use "label forgery" to affix legal and compliant labels to illegal content. Existing technology can only check the content itself or the label itself separately, and cannot establish a correlation verification mechanism between the label and the content. It is difficult to identify deceptive behavior such as "label does not match the actual content", resulting in a large amount of such illegal content flowing into the network.
[0004] (4) Lack of precision in verification results: Manual review or simple tool review usually only outputs a binary judgment result of "violation" or "non-violation". However, there are a lot of "borderline" content in online content (such as marginal marketing advertisements, mildly inappropriate content, etc.), which cannot be accurately managed through binary judgment. Existing technology cannot quantify and classify the violation risk of such content, which leads to regulators either "one-size-fits-all" letting it go (causing compliance risks) or "one-size-fits-all" blocking it (affecting the dissemination of legitimate content), making it impossible to achieve differentiated and precise supervision.
[0005] Therefore, there is an urgent need for an automated, intelligent, and precise technological approach to build a fully automated verification system that covers the entire process from identifier verification to content review, and from consistency determination to risk classification, thereby improving verification efficiency, coverage, and accuracy, and addressing many shortcomings of existing technologies. Summary of the Invention
[0006] This invention provides a method and apparatus for verifying the security of traffic data content, in order to solve the problems of low efficiency and low accuracy caused by excessive reliance on manpower and insufficient technical means in the existing technology.
[0007] On one hand, the present invention provides a method for verifying the security of traffic data content, the method comprising: Collect multi-source heterogeneous data for each content unit, assign a unique ID to each content unit, and generate a multimodal feature set associated with each content unit ID; the multi-source heterogeneous data includes any one or more of the following formats: text data, image data, video data, audio data, and identification information; Verify the authenticity of the identification information; If the verification passes, a consistency check between the content and the identifier is performed based on the multimodal feature set to determine whether there is any tag abuse.
[0008] Optionally, the method further includes: storing and managing the collected data of each content unit using a three-level architecture of raw layer, preprocessing layer, and feature layer, wherein the raw layer stores the raw data, the preprocessing layer stores the intermediate data after cleaning the raw data, and the feature layer stores the structured features extracted from the intermediate data.
[0009] Optionally, the structured features corresponding to the text data include any one or more of the following: basic statistical features, semantic features, and entity information; the basic statistical features include: word-level statistical features and syntactic-level statistical features; the semantic features corresponding to the image data, video data, and audio data include any one or more of the following: face ID, OCR text, temporal dynamics, and content semantics.
[0010] Optionally, verifying the authenticity of the identification information includes: Verify the compliance of the identification information; If the verification fails, the identification information will be marked as high risk; If the verification is successful, the credibility of the identification information is evaluated using the gradient boosting tree and the relevant features of the identification information to obtain the credibility score of the identification information. If the credibility score of the identification information is greater than the set credibility threshold, the authenticity verification is successful.
[0011] Optionally, verifying the compliance of the identification information includes: Use a compliance identifier library to check whether the identifier information is legal and meets the specification requirements and hierarchical relationships; For legal and compliant identification information, a large language model is used to identify homophonic spoofing to determine whether the identification information is compliant.
[0012] Optionally, the method of using a large language model for homophonic spoofing recognition includes: Using a large language model for contextual semantic understanding, determine whether the identification information is a homophonic spoof and / or provide prompts to the large language model to determine whether the identification information is a homophonic spoof.
[0013] Optionally, the relevant features of the identification information include: the subject features, behavioral features, and context features of the identification information; the gradient boosting tree adopts the XGBoost model or the LightGBM model.
[0014] Optionally, the method further includes: if the credibility score of the identification information is less than or equal to the set credibility threshold, then triggering a deep verification process for the identification information.
[0015] Optionally, the multimodal feature set includes: a content summary and a comprehensive content semantic vector; the step of performing a consistency check between the content and the identifier based on the multimodal feature set to determine whether tag abuse exists includes: The text information is encoded into a standard semantic vector to obtain the tag semantic vector; Calculate the similarity between the integrated content semantic vector and the tag semantic vector to obtain a similarity score; The semantic matching score between the content summary and the identifier information is determined using a large language model. The similarity score and the semantic matching score are weighted and calculated to obtain the consistency score; If the consistency score is greater than the set consistency threshold, it is determined that there is no label abuse.
[0016] Optionally, the method further includes: performing security checks on the content ontology associated with each content unit to determine the degree of violation score of the content unit; the security checks include any one or more of the following: locating explicit violating words and named entities; identifying homophones, metaphors, or ironic expressions; and identifying abnormal non-speech expressions.
[0017] Optionally, the method further includes: Based on the verification and inspection results at different levels, conduct risk assessments and determine corresponding handling recommendations; Generate and output a structured inspection report, which includes: content unit ID, score for each stage, risk level, handling recommendations, and violation location information.
[0018] On the other hand, the present invention also provides a device for verifying the security of traffic data content, the device comprising: The data acquisition module is used to collect multi-source heterogeneous data of each content unit, assign a unique ID to each content unit, and generate a multimodal feature set associated with each content unit ID; the multi-source heterogeneous data includes any one or more of the following formats: text data, image data, video data, audio data, and identification information; The identifier verification module is used to verify the authenticity of the identifier information; The consistency check module is used to perform a consistency check between the content and the identifier based on the multimodal feature set after the identifier verification is passed, in order to determine whether there is any tag abuse.
[0019] Optionally, the identifier verification module includes: A compliance verification unit is used to verify the compliance of the identification information, and to mark the identification information as high-risk if the verification fails. The credibility assessment unit is used to assess the credibility of the identification information by using a gradient boosting tree and related features of the identification information when the compliance verification unit verifies that the identification information is qualified, and to obtain a credibility score of the identification information; if the credibility score of the identification information is greater than a set credibility threshold, the authenticity verification is determined to be passed.
[0020] Optionally, the device further includes: The content inspection module is used to perform security checks on the content ontology associated with each content unit and determine the degree of violation score of the content unit; the security check includes any one or more of the following: locating explicit violation words and named entities; identifying homophones, metaphors, or ironic expressions; and identifying abnormal non-speech.
[0021] Optionally, the device further includes: The risk assessment module is used to conduct risk assessments and determine corresponding handling recommendations based on the verification and inspection results at different levels. The report generation module is used to generate a structured inspection report, which includes: content unit ID, score for each stage, risk level, handling suggestions, and violation location information.
[0022] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, performs the steps of the identification traffic data content security verification method.
[0023] The method and apparatus for verifying the security of traffic data content provided by this invention collect multi-source heterogeneous data of each content unit, assign a unique ID to each content unit, generate a multimodal feature set associated with each content unit ID, and realize fully automated analysis from verifying the authenticity of the identifier to determining the consistency between the content and the identifier by constructing a multi-layer intelligent verification model. This can significantly improve the verification efficiency, coverage and accuracy.
[0024] Furthermore, when verifying the authenticity of the identification information, not only is its compliance verified, but also its credibility is assessed using a gradient boosting tree and relevant features of the identification information, ensuring the comprehensiveness of the verification and the reliability of the results. Moreover, when the credibility score of the identification information is low, a deeper verification process for that identification information can be triggered.
[0025] Furthermore, when verifying the compliance of identification information, not only is its legality and hierarchical relationship checked based on the compliant identification database, but also a large language model is used to identify homophonic spoofing, effectively preventing unqualified identification information disguised by homophony.
[0026] Furthermore, when performing consistency checks between content and identifiers, a multi-dimensional comprehensive content semantic vector can be generated based on a multimodal feature set. Simultaneously, identifier information is encoded into a standard semantic vector to obtain a tag semantic vector. On one hand, the similarity between the comprehensive content semantic vector and the tag semantic vector is calculated to obtain a similarity score; on the other hand, a large language model is used to determine the semantic matching score between the content summary and the identifier information. Combining the similarity score and the semantic matching score to determine the consistency between content and identifiers ensures the accuracy of the check results.
[0027] Furthermore, security checks can be performed on the content ontology associated with each content unit ID to ensure the comprehensiveness of the check.
[0028] Furthermore, structured inspection reports can be generated based on the inspection results at each stage. These reports can also include recommendations for handling violations and information on the location of violations, providing valuable references for subsequent decision-making and improving efficiency. Attached Figure Description
[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.
[0030] Figure 1 This is a flowchart of a method for verifying the security of traffic data content provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the consistency check between content and identifier in an embodiment of the present invention; Figure 3 This is another flowchart of the method for verifying the security of traffic data content provided in this embodiment of the invention; Figure 4 This is a schematic diagram of a security verification device for identifying traffic data content provided in an embodiment of the present invention; Figure 5 This is another structural schematic diagram of the identification traffic data content security verification device provided in this embodiment of the invention; Figure 6 This is another structural schematic diagram of the identification traffic data content security verification device provided in the embodiment of the present invention. Detailed Implementation
[0031] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0032] 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 of the present invention. 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.
[0033] To address the problems of low efficiency and inaccuracy caused by excessive reliance on manpower and insufficient technical means in existing technologies, this invention provides a method and apparatus for verifying the security of identifier traffic data content. By constructing a multi-layer intelligent verification model, it achieves fully automated analysis from identifier authenticity verification to content consistency determination, significantly improving verification efficiency, coverage, and accuracy.
[0034] like Figure 1 The diagram shown is a flowchart of a method for verifying the security of traffic data content provided in an embodiment of the present invention, which includes the following steps: Step 101: Collect multi-source heterogeneous data for each content unit, assign a unique ID to each content unit, and generate a multimodal feature set associated with each content unit ID.
[0035] The multi-source heterogeneous data includes any one or more of the following formats: text data, image data, video data, audio data, and identification information (also known as tags).
[0036] It should be noted that, in this embodiment of the invention, content units are used as the basic data processing units, and each content unit can contain several modalities of data, including text, images, videos, audio, and identification information. Subsequent feature extraction, consistency checks, and security verifications are all performed based on this content unit.
[0037] Specifically, text, images, videos, audio, and identification information can be collected through compliant interfaces, and a unique ID can be assigned to each content unit to generate a multimodal feature set associated with each content unit ID.
[0038] In some embodiments, for ease of management, a three-tier architecture of raw data layer, preprocessing layer, and feature layer can be used for data storage management. The raw data layer stores the raw data, facilitating data traceability and future reprocessing needs; the preprocessing layer stores intermediate data after cleaning the raw data; and the feature layer stores the structured features extracted from the intermediate data.
[0039] Data cleaning processes can include, but are not limited to, removing invalid, duplicate, and noisy data to provide valid data for subsequent analysis.
[0040] Feature extraction for different modalities can be accomplished using corresponding deep learning models such as BERT and ResNet. For different modalities, targeted deep feature extraction models can be designed to extract specific features, such as: The structured features of the corresponding text data may include, but are not limited to, any one or more of the following: basic statistical features, semantic features, and entity information.
[0041] The basic statistical features can include: word-level statistical features and syntactic statistical features. Word-level statistical features may include, but are not limited to: number of characters, number of words, number of special symbols (such as @, #), etc. Syntactic statistical features may include, but are not limited to: number of sentences, average sentence length, number of exclamation marks, number of question marks, etc.
[0042] The semantic features can be semantic richness, such as including but not limited to: the TTR (Telematics Transition) index and sentiment score (through dictionary matching). These lightweight statistical features can serve as an effective supplement to deep semantic features (BERT vectors).
[0043] The entity information refers to specific named entities identified in the text, such as names of people, places, organizations, brands, and sensitive events. Specifically, this structured information can be extracted from the text using NER (Named Entity Recognition) models (such as BERT-BiLSTM-CRF). For example, in the description "A celebrity attended a brand launch event in Hangzhou," we can extract "Name: XXX," "Location: Hangzhou," and "Organization: A certain brand." These entities are used to help determine the authenticity and risk of the content (such as whether it involves sensitive figures or false advertising) and to perform cross-modal alignment with faces identified in images / videos and OCR (Optical Character Recognition) text.
[0044] Visual features can be extracted into deep representations using models such as ResNet (Residual Network) or ViT (VisionTransformer), and supplemented with OCR and face recognition; video and audio features can take into account both temporal dynamics and content semantics. Accordingly, the semantic features of image data, video data, and audio data can include, but are not limited to, any one or more of the following: face ID, OCR text, temporal dynamics, and content semantics.
[0045] Temporal dynamics refer to the changing patterns of video or audio over time. For example, the continuity of characters' actions in a video (such as fighting or dancing) and the frequency of scene transitions; and the fluctuations in tone of voice, the timing of keyword appearances, and sudden changes in background noise in audio. Specifically, temporal modeling networks such as I3D and TimeSformer, or LSTM / Transformer encoding of frame-by-frame / segment features, can be used to capture these dynamic behaviors. Compared to static keyframes, temporal dynamics can more accurately identify violations (such as violent actions or manipulative speech patterns), improving verification accuracy.
[0046] The semantic content can be obtained by transforming the raw data into high-dimensional semantic vectors using a deep learning model. For text data, semantic embeddings can be generated using models such as BERT; for video / images, visual semantics can be extracted using ViT or ResNet networks; and for audio signals, speech content can be encoded using encoders such as wav2vec 2.0 or Whisper.
[0047] Furthermore, LLM (Large Language Model) can be used to generate natural language summaries from multimodal data, which can then be vectorized to obtain the content summary corresponding to the content unit ID. For example, multimodal information such as the original text, the text recognized by OCR, and the speech content transcribed by ASR (Automatic Speech Recognition) can be concatenated and used as input to LLM (such as Qwen), with the instruction: "Please summarize the core theme of this content in one sentence." Based on this, LLM outputs a concise and coherent natural language summary, such as "The video shows a golden retriever chasing a frisbee in the park." This content summary retains key semantics while filtering out noise, and can be used for subsequent consistency checks with the identification information and manual verification.
[0048] Simultaneously, the semantic vectors of all modalities can be fused to form a unified content semantic representation. For ease of description, this unified content semantic representation is called a comprehensive content semantic vector. This comprehensive content semantic vector reflects "what this content is actually about" and can be used for semantic comparison with the identification information. For example, through agent coordination, the text summary generated by LLM, the image subject identified by ViT, the temporal features of the video captured by I3D, and the text transcribed from the audio can be fused to generate a 1024-dimensional comprehensive content semantic vector.
[0049] After extraction and normalization, the features are stored in both a vector database and a relational database. The relational database is a graph database or relational database used to store the relationships between features and links between content and external knowledge. For example: 1) The association between the facial IDs identified in the video and the names of celebrities mentioned in the text.
[0050] 2) Consistency markers between the text extracted by OCR and the audio-transcribed text.
[0051] 3) The association between the content ID and its associated metadata such as the account it belongs to, the IP address it was published in, and its historical behavior.
[0052] This relational information provides contextual support for subsequent cross-modal verification and credibility assessment.
[0053] Through the above processing, a multimodal feature set with associated unified ID can be output. These structured feature vectors constitute the underlying data support for the entire intelligent verification process, enabling subsequent models to understand and analyze complex content consistently and efficiently.
[0054] Step 102: Verify the authenticity of the identification information.
[0055] Verifying the authenticity of labels is the first line of defense in risk filtering, and its core task is to determine whether the label information itself is compliant and credible.
[0056] Specifically, the compliance of the identification information can be verified first; if the verification fails, the identification information is marked as high risk; if the verification passes, the credibility of the identification information is evaluated using gradient boosting trees and related features of the identification information to obtain a credibility score; if the credibility score of the identification information is greater than the set credibility threshold, the authenticity verification is passed.
[0057] When verifying the compliance of identification information, the following two aspects can be considered: (1) Use the compliance identifier library to check whether the identifier information is legal and meets the regulatory requirements and hierarchical relationships. The compliance identifier library not only contains legal identifier terms, but also defines a tree-like or graph-like hierarchical structure between identifiers. For example, "food > Chinese food > Sichuan cuisine". During verification, check whether the identifier combination submitted by the user conforms to this predefined hierarchical logic. If "Sichuan cuisine" and "French cuisine" are listed together, or if a "self-created identifier" that does not exist in the compliance identifier library is used, it is considered non-compliant.
[0058] (2) For legal and compliant identification information, use a large language model to perform homophonic spoofing identification to determine whether the identification information is compliant. Specifically, one or two of the following methods can be used: One approach is to use LLM (Liquidity Management Model) for contextual semantic understanding to determine whether the identified information is a homophonic spoof. LLM can combine the context of the entire sentence to determine whether a word is a homophonic substitute for a sensitive word. For example, "paratrooper" might be identified as profanity in a non-military context.
[0059] Another approach is to guide the LLM with prompts to determine whether the identified information is misused through homophones. For example, the LLM can be fed a command like "Does the following text contain any expressions that circumvent review by using homophones, symbols, or pinyin variations?" along with a label or descriptive text, allowing it to output a judgment result. Furthermore, a table of common homophones can be constructed as supplementary knowledge, enabling the LLM to perform dynamic matching and reasoning.
[0060] For compliant tagging information, further credibility assessment is required using the Gradient Boosting Tree (GBT) model. This model comprehensively considers the tag's creation entity and time, the posting account's historical behavior, and the correlation between the tag and the content. It also introduces an intelligent agent to perform cross-validation using multi-source databases, ultimately outputting a credibility score from 0 to 100.
[0061] For example, in some implementations, XGBoost (eXtreme Gradient Boosting) or LightGBM (Light Gradient Boosting Machine) can be used as the gradient boosting tree model. Its input features include: 1) Main characteristics: Creator account's credit score, number of historical violations, and registration duration.
[0062] 2) Behavioral characteristics: the account's recent tagging frequency and distribution of commonly used tags.
[0063] 3) Contextual features: TF-IDF similarity between the current tag and the preliminary features of the content (such as keywords).
[0064] Risk can be initially assessed based on credibility scores. For example, the initial risk assessment rule is: if the compliance verification of the identification information itself fails, it is directly classified as high risk; if it is compliant but the credibility score is <40, it is also considered high risk and needs to enter the in-depth verification process.
[0065] By verifying the authenticity of the identification information, risks can be identified and diverted at an early stage, and the overall allocation of computing resources can be optimized.
[0066] Step 103: If the verification passes, perform a consistency check between the content and the identifier based on the multimodal feature set to determine whether there is any tag abuse.
[0067] As mentioned earlier, the multimodal feature set corresponding to each content unit ID can include a content summary and a comprehensive content semantic vector. When performing consistency checks between content and identifiers, these two types of information can be compared with the identifier information to determine if any tag abuse has occurred.
[0068] like Figure 2 The diagram shown is a flowchart of a content and identifier consistency check in an embodiment of the present invention, including the following steps: Step 201: Encode the identification information into a standard semantic vector to obtain the label semantic vector.
[0069] It should be noted that the identification information can be a single label or a combination of labels, and this embodiment of the invention does not limit this.
[0070] For example, by sending the user-submitted identification information into the same LLM embedding layer, its semantic vector can be obtained directly.
[0071] Step 202: Calculate the similarity between the comprehensive content semantic vector in the multimodal feature set and the tag semantic vector to obtain a similarity score.
[0072] For example, the cosine similarity between the comprehensive content semantic vector and the tag semantic vector can be calculated. The cosine value of the angle between these two vectors can be used to obtain a similarity score.
[0073] Step 203: Use a large language model to determine the semantic matching score of the content summary and identifier information in the multimodal feature set.
[0074] For example, one can utilize the APIs (Application Programming Interfaces) of open-source or commercial large language models such as GPT (Generative Pre-trained Transformer) or deepseek, or use an internally fine-tuned private deployment version. The choice of large language model can be determined based on business requirements regarding cost, latency, and data privacy, and this embodiment of the invention does not impose any limitations on this.
[0075] Step 204: Perform a weighted calculation on the similarity score and the semantic matching score to obtain a consistency score.
[0076] For example, in a non-limiting embodiment, the weight of the similarity score can be set to 60%, and the weight of the semantic matching score can be set to 40%. The consistency score is obtained by weighted calculation based on the set weights.
[0077] Step 205: If the consistency score is greater than the set consistency threshold, it is determined that there is no tag abuse.
[0078] For example, in a non-limiting embodiment, a consistency threshold can be set to 0.5. If the value is lower than this threshold, it is judged as "label abuse" and an alert is triggered.
[0079] The identification traffic data content security verification method provided in this embodiment of the invention collects multi-source heterogeneous data of each content unit. Based on this multi-source heterogeneous data, it not only verifies the authenticity of the identification information of the content unit, but also performs consistency checks between the content and the identification. It can effectively identify the behavior of maliciously labeling irrelevant tags in order to obtain traffic.
[0080] like Figure 3 The diagram shown is another flowchart of the method for verifying the security of traffic data content provided in this embodiment of the invention, including the following steps: Step 301: Collect multi-source heterogeneous data for each content unit, assign a unique ID to each content unit, and generate a multimodal feature set associated with each content unit ID.
[0081] The multi-source heterogeneous data includes any one or more of the following formats: text data, image data, video data, audio data, and identification information.
[0082] Step 302: Verify the authenticity of the identification information.
[0083] The verification of the authenticity of the identification information includes both compliance and credibility verification. For non-compliant identification information or compliant identification information with a credibility score lower than the set threshold, the identification information can be marked as "high risk". For compliant identification information with a credibility score higher than the set threshold, further consistency checks between the content and the identification are required, i.e., step 303 is executed.
[0084] Step 303: If the verification passes, perform a consistency check between the content and the identifier based on the multimodal feature set to determine whether there is any tag abuse.
[0085] For detailed inspection procedures, please refer to... Figure 2 The flowchart shown is not repeated here.
[0086] If the consistency check fails, the identifier information can be marked as "tag abuse" and an alert can be triggered to initiate other related checks. If the consistency check passes, all content corresponding to the content unit ID can be further checked to ensure the security of the relevant content.
[0087] Step 304: After the consistency check is passed, a security check is performed on the content ontology associated with each content unit to determine the violation level score of the content unit.
[0088] The security checks may include, but are not limited to, any one or more of the following: locating explicit violations of words and named entities; identifying homophones, metaphors, or ironic expressions; and identifying abnormal non-vocal expressions.
[0089] In some embodiments, text verification can combine BERT (Bidirectional Encoder Representations from Transformers) + CRF (Conditional Random Field) models with the semantic understanding capabilities of LLM to identify disguised violations such as homophones and metaphors. Image and video verification can use deep learning models to identify sensitive elements and perform cross-modal verification. Audio verification can analyze anomalies through speech-to-text conversion and acoustic models. For example: The BERT+CRF model is used to accurately locate explicit violation words and named entities (such as sensitive names and locations) with a high recall rate. Using LLM to understand contextual semantics, identify homophones, metaphors, or ironic expressions such as "paratrooper" (a slang term) and "censorship" (a censorship term).
[0090] For example, when the text contains the phrase "This operation is really like a paratrooper," the BERT+CRF model might fail to identify the keyword, but the LLM model can determine it as offensive language based on the context. The results of both models are combined: if either model determines it to be in violation, it enters the high-risk queue, balancing rule coverage and semantic depth.
[0091] Audio verification can be divided into two steps: First, use ASR (such as Whisper) to convert speech to text, and then use a text security model to detect illegal content; second, directly use acoustic models (such as PANNs (Pre-trained Audio Neural Networks) or wav2vec (wav-to-vector, speech waveform to vector representation learning)) to analyze the original audio signal and identify non-speech anomalies, such as screams, fighting sounds, and sensitive background music. Even if ASR transcription fails (e.g., dialects, low-quality recordings), acoustic models can still capture risk signals. Cross-validation of the results from both methods can improve the detection capability of pure audio violations (such as leading speech, illegal calls).
[0092] Each model outputs a violation score from 0 to 100. The violation scores from each model are combined to obtain the final violation score for each content unit.
[0093] In some embodiments, a risk assessment indicator system can be further constructed, such as a risk assessment indicator system and handling strategy. The assessment indicators include identification risk, consistency risk, and content violation risk, and their weights are determined using the analytic hierarchy process (AHP). The total risk score is calculated by weighting the scores of each indicator.
[0094] Different risk levels require different response strategies, for example: Level 1 risk (total risk score ≥ 80 points): Immediately intercept and report. Level 2 risk (total risk score 60-79 points) requires mandatory manual review; Level 3 risk (total risk score 40-59 points), traffic restriction monitoring; Level 4 risk (total risk score < 40 points) allows normal release.
[0095] The aforementioned quantitative grading mechanism enables the refinement and automation of disposal strategies.
[0096] In some embodiments, a structured inspection report may be further output, which may include, but is not limited to: content unit ID, scores for each stage, risk level, handling suggestions, and violation location information.
[0097] The inspection report can be in JSON format and can be pushed to the relevant management platform via API interface.
[0098] Furthermore, to ensure reliability, a human-machine collaboration mechanism can be established, with all high-risk content entering a manual review queue and providing visual annotation tools for reviewers to correct the results.
[0099] Furthermore, the high-quality feedback data generated by annotation can be used to drive continuous system optimization. The agent automatically collects review data, and when it accumulates to a certain scale, it triggers incremental training of core models such as credibility assessment, consistency determination, and violation identification.
[0100] Misjudgment samples in LLM (such as misjudging compliant content as non-compliant, or omitting disguised non-compliant samples) can be collected and organized separately. For example, misjudgment samples can be organized into "instruction-input-correct output" triples for instruction fine-tuning to improve semantic reasoning accuracy.
[0101] For example, given a text containing homophones, the original LLM outputs "No violation," but a human labels it as "Violation." The solution is to construct an instruction: "Please determine whether the following text contains obscure violations," and assign the correct label. These samples form a fine-tuning dataset, and the LLM is retrained using supervised fine-tuning (SFT). In this way, the LLM can gradually learn to more accurately understand semantic traps, cultural contexts, and novel avoidance methods in business scenarios, thereby improving its inference accuracy in tasks such as label matching, content summarization, and violation identification.
[0102] Furthermore, the compliance identifier library and prohibited word list can be updated regularly based on business feedback, and the weights of risk assessment indicators can be recalculated using historical data, enabling dynamic iteration of rules and models. This closed-loop optimization mechanism ensures that the system can adapt to the continuous changes in online content and violations, maintaining long-term accuracy and efficiency.
[0103] The identification traffic data content security verification method provided in this embodiment leverages the strong semantic understanding and reasoning capabilities of LLM (Limited Language Management) and embeds them into key processes such as feature extraction, semantic correction, fuzzy judgment, and matching degree calculation, acting as the "brain" to enhance the ability to identify complex and hidden violations. Simultaneously, it uses traditional machine learning and deep learning models as efficient "sensors and limbs" to handle large-scale feature extraction and pattern recognition. Both are integrated through an intelligent agent for task scheduling and information fusion. This invention overcomes the limitations of single content review or identification verification, constructing a three-tiered progressive verification logic of "identifier itself -> identification and content relationship -> content ontology," covering the entire risk chain. Furthermore, this invention employs a quantitative scoring model based on the Analytic Hierarchy Process (AHP) to condense multi-dimensional and multi-modal verification results into a single risk score, with corresponding clear operational guidelines, making the automated decision-making process transparent, explainable, and easily intervened.
[0104] Accordingly, embodiments of the present invention also provide a device for verifying the security of traffic data content, such as... Figure 4 The diagram shown is a structural schematic of the device.
[0105] The traffic data content security verification device 400 includes the following modules: The data acquisition module 401 is used to acquire multi-source heterogeneous data of each content unit, assign a unique ID to each content unit, and generate a multimodal feature set 40 associated with each content unit ID; the multi-source heterogeneous data includes any one or more of the following formats: text data, image data, video data, audio data, and identification information; The identifier verification module 402 is used to verify the authenticity of the identifier information; The consistency check module 403 is used to perform a consistency check between the content and the identifier based on the multimodal feature set after the identifier verification is passed, in order to determine whether there is any label abuse.
[0106] In practical implementation, the data acquisition module 401 can manage the acquired data of each content unit using a three-level architecture: raw layer, preprocessing layer, and feature layer. The raw layer stores the raw data, the preprocessing layer stores the intermediate data after cleaning the raw data, and the feature layer stores the structured features extracted from the intermediate data.
[0107] The cleaning and feature extraction of data can be referred to the description in the previous embodiments of the present invention, and will not be repeated here.
[0108] In one non-limiting embodiment, the identifier verification module 402 may include: a compliance verification unit and a credibility assessment unit. Wherein: The compliance verification unit is used to verify the compliance of the identification information and, if the verification fails, marks the identification information as high-risk. For example, the identification information can be verified based on a rule base and an LLM (Local Language Model). Specifically, the compliance identification base is used to check whether the identification information is legal and meets the specification requirements and hierarchical relationships; for legal identification information that meets the specification requirements, a large language model is used to perform homophonic spoofing recognition to determine whether the identification information is compliant.
[0109] The credibility assessment unit, after the compliance verification unit verifies the identity information as qualified, uses a gradient boosting tree and relevant features of the identity information to assess its credibility and obtain a credibility score. If the credibility score of the identity information is greater than a set credibility threshold, the authenticity verification is deemed successful. The gradient boosting tree can be, for example, an XGBoost model or a LightGBM model. The relevant features of the identity information may include, but are not limited to, the subject features, behavioral features, and contextual features of the identity information.
[0110] The consistency check module 403 can specifically check whether the identifier information and content are consistent for each content unit. For example, text information can be encoded into a standard semantic vector to obtain a tag semantic vector. The similarity between the comprehensive content semantic vector and the tag semantic vector of the content unit can be calculated to obtain a similarity score. At the same time, LLM can be used to determine the semantic matching score between the content summary and the identifier information of the content unit. By combining the similarity score and the semantic matching score, it can be determined whether there is any tag abuse.
[0111] like Figure 5 The diagram shown is another structural schematic of the identification traffic data content security verification device provided in an embodiment of the present invention.
[0112] and Figure 4 The difference between the embodiments shown is that, Figure 5 The identification traffic data content security verification device 400 shown may further include: The content inspection module 404 is used to perform security checks on the content ontology associated with each content unit and determine the degree of violation score of the content unit. The security check may include, for example, any one or more of the following: locating explicit violating words and named entities; identifying homophones, metaphors, or ironic expressions; and identifying abnormal non-speech elements. The specific inspection process can be referred to the description in the previous embodiments of the method of the present invention, and will not be repeated here.
[0113] like Figure 6 The diagram shown is another structural schematic of the identification traffic data content security verification device provided in an embodiment of the present invention.
[0114] and Figure 5 The difference between the embodiments shown is that, Figure 6 The identification traffic data content security verification device 400 shown may further include: a risk assessment module 405 and a report generation module 406. Wherein: The risk assessment module 405 is used to conduct risk assessments and determine corresponding handling recommendations based on the verification and inspection results at different levels.
[0115] The report generation module 406 is used to generate a structured inspection report, which includes: content unit ID, score of each stage, risk level, handling suggestions, and violation location information.
[0116] The method and apparatus for verifying the security of traffic data content provided by this invention collect multi-source heterogeneous data of each content unit, assign a unique ID to each content unit, generate a multimodal feature set associated with each content unit ID, and realize fully automated analysis from verifying the authenticity of the identifier to determining the consistency between the content and the identifier by constructing a multi-layer intelligent verification model. This can significantly improve the verification efficiency, coverage and accuracy.
[0117] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.
[0118] The present invention also provides a storage medium, which is a computer-readable storage medium storing a computer program thereon, the computer program being executable when it runs. Figure 1 or Figure 2 or Figure 3 The method shown may include some or all of the steps. The storage medium may include read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc. The storage medium may also include non-volatile memory or non-transitory memory, etc.
[0119] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data provider to another website, computer, server, or data provider via wired or wireless means.
[0120] The embodiments of the present invention have been described in detail above. Specific implementation methods have been used to illustrate the present invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and apparatus of the present invention, and are only a part of the embodiments of the present invention, not all of them. 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, and the content of this specification should not be construed as a limitation of the present invention. Therefore, 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. A method for verifying the security of traffic data content, characterized in that, The method includes: Collect multi-source heterogeneous data for each content unit, assign a unique ID to each content unit, and generate a multimodal feature set associated with each content unit ID; the multi-source heterogeneous data includes any one or more of the following formats: text data, image data, video data, audio data, and identification information; Verify the authenticity of the identification information; If the verification passes, a consistency check between the content and the identifier is performed based on the multimodal feature set to determine whether there is any tag abuse.
2. The method for verifying the security of traffic data content according to claim 1, characterized in that, The method further includes: The collected data units are stored and managed using a three-tier architecture consisting of an original layer, a preprocessing layer, and a feature layer. The original layer stores the raw data, the preprocessing layer stores the intermediate data after cleaning the original data, and the feature layer stores the structured features extracted from the intermediate data.
3. The method for verifying the security of traffic data content according to claim 2, characterized in that, The structured features corresponding to the text data include any one or more of the following: basic statistical features, semantic features, and entity information; The basic statistical features include: word-level statistical features and syntax-level statistical features; The semantic features corresponding to the image data, video data, and audio data include any one or more of the following: face ID, OCR text, temporal dynamics, and content semantics.
4. The method for verifying the security of traffic data content according to claim 1, characterized in that, The verification of the authenticity of the identification information includes: Verify the compliance of the identification information; If the verification fails, the identification information will be marked as high risk; If the verification is successful, the credibility of the identification information is evaluated using the gradient boosting tree and the relevant features of the identification information to obtain the credibility score of the identification information. If the credibility score of the identification information is greater than the set credibility threshold, the authenticity verification is successful.
5. The method for verifying the security of identifier traffic data content according to claim 4, characterized in that, The verification of the compliance of the identification information includes: Use a compliance identifier library to check whether the identifier information is legal and meets the specification requirements and hierarchical relationships; For legal and compliant identification information, a large language model is used to identify homophonic spoofing to determine whether the identification information is compliant.
6. The method for verifying the security of identifier traffic data content according to claim 5, characterized in that, The method of using a large language model for homophonic spoofing recognition includes: Using a large language model for contextual semantic understanding, determine whether the identification information is a homophonic spoof and / or The large language model is guided by prompt words to determine whether the identification information is a homophonic spoofing.
7. The method for verifying the security of identifier traffic data content according to claim 4, characterized in that, The relevant features of the identification information include: the subject features, behavioral features, and contextual features of the identification information; the gradient boosting tree adopts the XGBoost model or the LightGBM model.
8. The method for verifying the security of identifier traffic data content according to claim 4, characterized in that, The method further includes: If the credibility score of the identification information is less than or equal to the set credibility threshold, a deep verification process for the identification information is triggered.
9. The method for verifying the security of identifier traffic data content according to claim 1, characterized in that, The multimodal feature set includes: content summary and comprehensive content semantic vector; the consistency check between content and identifier based on the multimodal feature set to determine whether there is tag abuse includes: The text information is encoded into a standard semantic vector to obtain the tag semantic vector; Calculate the similarity between the integrated content semantic vector and the tag semantic vector to obtain a similarity score; The semantic matching score between the content summary and the identifier information is determined using a large language model. The similarity score and the semantic matching score are weighted and calculated to obtain the consistency score; If the consistency score is greater than the set consistency threshold, it is determined that there is no label abuse.
10. The method for verifying the security of identified traffic data content according to any one of claims 1 to 9, characterized in that, The method further includes: A security check is performed on the content ontology associated with each content unit to determine the degree of violation of the content unit; the security check includes any one or more of the following: locating explicit violation words and named entities; identifying homophones, metaphors, or ironic expressions; and identifying abnormal non-speech words.
11. The method for verifying the security of identifier traffic data content according to claim 10, characterized in that, The method further includes: Based on the verification and inspection results at different levels, conduct risk assessments and determine corresponding handling recommendations; Generate and output a structured inspection report, which includes: content unit ID, score for each stage, risk level, handling recommendations, and violation location information.
12. A device for verifying the security of traffic data content, characterized in that, The device includes: The data acquisition module is used to collect multi-source heterogeneous data of each content unit, assign a unique ID to each content unit, and generate a multimodal feature set associated with each content unit ID; the multi-source heterogeneous data includes any one or more of the following formats: text data, image data, video data, audio data, and identification information; The identifier verification module is used to verify the authenticity of the identifier information; The consistency check module is used to perform a consistency check between the content and the identifier based on the multimodal feature set after the identifier verification is passed, in order to determine whether there is any tag abuse.
13. The device for verifying the security of traffic data content according to claim 12, characterized in that, The identifier verification module includes: A compliance verification unit is used to verify the compliance of the identification information, and to mark the identification information as high-risk if the verification fails. The credibility assessment unit is used to assess the credibility of the identification information by using a gradient boosting tree and related features of the identification information when the compliance verification unit verifies that the identification information is qualified, and to obtain a credibility score of the identification information; if the credibility score of the identification information is greater than a set credibility threshold, the authenticity verification is determined to be passed.
14. The device for verifying the security of traffic data content according to claim 12 or 13, characterized in that, The device further includes: The content inspection module is used to perform security checks on the content ontology associated with each content unit and determine the degree of violation score of the content unit; the security check includes any one or more of the following: locating explicit violation words and named entities; identifying homophones, metaphors, or ironic expressions; and identifying abnormal non-speech.
15. The device for verifying the security of traffic data content according to claim 14, characterized in that, The device further includes: The risk assessment module is used to conduct risk assessments and determine corresponding handling recommendations based on the verification and inspection results at different levels. The report generation module is used to generate a structured inspection report, which includes: content unit ID, score for each stage, risk level, handling suggestions, and violation location information.
16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it performs the steps of the identification traffic data content security verification method according to any one of claims 1 to 11.
