A large model content security protection method and system based on pre-small model shunt, multi-model arbitration and hierarchical risk threshold scoring
Patent Information
- Application Number
- CN202611159630.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-02
- Publication Date
- 2026-10-09
AI Technical Summary
其一，单一护栏模型存在系统性偏置与检测盲区，对不同风险类别的检测能力参差不齐；而朴素的多模型投票或平均集成对所有输入一视同仁，无法按输入择优，又徒增开销
[0014]与现有技术相比，本发明的上述技术方案具有如下有益的技术效果：将前置小模型分流、多模型仲裁、分级风险输出与风险等级引导的阈值化打分耦合为一套完整的内容安全防护系统，前置小模型以极低成本消化绝大多数简单流量、仅放行少量疑难样本进入重型管线，大幅降低整体算力与延迟；多模型并行预测加仲裁择优突破单模型盲区、提升鲁棒性；分级风险输出区分违规与疑似以支撑差异化处置；风险等级引导的阈值化打分在保持多标签能力的前提下提供跨样本可比、可按类别灵活设阈的连续分数。由此在大规模线上流量场景中同时实现高判别精度、强鲁棒性、精细化处置与算力经济性的统一。
Smart Images
Figure CN122885932A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence security and large language model content security protection technology, specifically to a method and system for large model content security protection based on pre-positioned small model diversion, multi-model arbitration, and graded risk threshold scoring. Background Technology
[0002] With the large-scale deployment of large language models in scenarios such as intelligent dialogue, intelligent agents, and content generation, the industry generally deploys content security protection models (guardrail models) on both the input and output sides to detect and block illegal content. Generative large model guardrails, represented by LlamaGuard, ShieldGemma, Qwen3Guard, and WildGuard, have strong generalization capabilities and support multi-label discrimination, and have become mainstream solutions. However, existing solutions still have significant limitations in terms of multi-model collaboration, computational overhead, risk classification, and score adjustability. First, single-barrier models suffer from systematic biases and detection blind spots, exhibiting inconsistent detection capabilities across different risk categories. Meanwhile, naive multi-model voting or averaging treats all inputs equally, failing to select the best based on input and unnecessarily increasing overhead. Second, calling multiple large-model barriers for parallel prediction and arbitration for each input incurs high computational and latency costs. Since the vast majority of real-world traffic consists of simple samples that are clearly safe or clearly violate regulations, calling heavy multi-model pipelines for them wastes computational resources, lacking a cost-effective pre-processing step based on difficulty. Third, existing barrier outputs are mostly hard labels, lacking tiered expression of risk levels and unable to support differentiated handling. Fourth, the token-by-token decoding characteristic of generative models makes it difficult to directly output comparable, flexibly configurable continuous discrimination scores across samples; the few models that provide thresholding capabilities (such as Qwen3Guard) can only score for a single label, unable to provide adjustable thresholds for each risk category in multi-label scenarios. Summary of the Invention
[0003] The purpose of this invention is to address the problems existing in the background technology by proposing a method and system for large model content security protection based on front-end small model diversion, multi-model arbitration, and hierarchical risk threshold scoring.
[0004] The technical solution of this invention: A method for protecting the content security of large models based on pre-processing small model traffic splitting, multi-model arbitration, and graded risk threshold scoring, comprising: S1. Obtain the content to be reviewed, use the pre-built small model to judge and classify it, let obviously safe content pass directly, and directly block obviously illegal content. Send the difficult samples along with the corresponding labels and judgment reasons to the multi-model prediction process. S2. Based on difficult samples, multiple heterogeneous guardrail models are called in parallel to make predictions and output candidate prediction results; S3. Based on the candidate prediction results, the arbitration model is used to combine the original input and the discrimination reasons of each model to perform arbitration and generate the final discrimination result. S4. Based on the final judgment result, output the corresponding graded risk level according to each preset risk category, and generate the thresholded continuous risk score corresponding to each risk category by combining the generation probability of the corresponding risk level. S5. Compare the generated continuous risk score with the corresponding category preset threshold, and perform differentiated treatment based on the risk level to output structured results.
[0005] As a further improvement to this technical solution, in step S1, the content to be reviewed is obtained, and a pre-processing small model is used for discrimination and triage. Clearly safe content is directly allowed, while clearly illegal content is directly blocked. Difficult samples, along with their corresponding tags and discrimination reasons, are then sent to the multi-model prediction process. This includes the following steps: Based on all the content to be reviewed, a lightweight discriminative pre-model is used for preliminary discrimination, and the corresponding discrimination score is output. The content is then split into two streams based on two pre-set confidence thresholds. Specifically, when the discrimination score is higher than the upper confidence threshold, the content to be reviewed is judged as obviously violating and directly blocked; when the discrimination score is lower than the lower confidence threshold, the content to be reviewed is judged as obviously safe and directly released; when the discrimination score is between the upper and lower confidence thresholds, the content to be reviewed is judged as a difficult sample and a multi-model prediction process is executed.
[0006] As a further improvement to this technical solution, in step S2, multiple heterogeneous guardrail models are called in parallel to perform predictions based on difficult samples, and candidate prediction results are output, including the following steps: Difficult samples are sent to multiple heterogeneous guardrail models for parallel prediction. Each heterogeneous guardrail model independently performs content security judgment based on the same difficult sample, generates corresponding judgment results, and outputs corresponding candidate prediction results, including judgment labels, risk levels, and judgment reasons.
[0007] As a further improvement to this technical solution, in step S3, arbitration is performed based on the candidate prediction results using an arbitration model that combines the original input and the discrimination reasons of each model to generate the final discrimination result, including the following steps: Acquire candidate prediction results and organize the original input and each candidate prediction result into a unified context. Input the context into the arbitration big model and compare each candidate prediction result according to the discrimination reason corresponding to each candidate prediction result. Analyze the sufficiency and consistency of each discrimination reason. Based on the comparison results, output the fused final discrimination result through the arbitration big model.
[0008] As a further improvement to this technical solution, in step S4, based on the final judgment result, the corresponding graded risk level is output according to each preset risk category, including the following steps: Based on the final judgment result, and reading the preset risk categories corresponding to the final judgment result, a corresponding risk level is generated for each preset risk category. The risk level is taken from the ordered set {safe, low, medium, high}. A single discrete token is used to output the risk level corresponding to each preset risk category so that the generation probability of the corresponding risk level token can be read later.
[0009] As a further improvement to this technical solution, the step of using a single discrete token to output the risk level corresponding to each preset risk category includes the following steps: Obtain the risk level corresponding to each preset risk category, and establish the correspondence between each risk level and a single discrete token. Each risk level corresponds to a unique discrete token. For each preset risk category, map the corresponding risk level to the corresponding single discrete token, and output the single discrete token corresponding to each preset risk category.
[0010] As a further improvement to this technical solution, in step S4, the thresholded continuous risk score corresponding to each risk category is generated by combining the generation probability of the corresponding risk level, including the following steps: Obtain the risk level corresponding to each preset risk category, and establish a corresponding risk interval based on each risk level, dividing the continuous score space into multiple fixed and non-overlapping risk intervals; read the generation probability of the risk level token corresponding to each preset risk category based on the risk interval; generate a thresholded continuous risk score corresponding to each preset risk category based on the risk interval and the generation probability.
[0011] As a further improvement to this technical solution, the step of generating thresholded continuous risk scores corresponding to each preset risk category based on risk intervals and generation probabilities includes the following steps: The width of the corresponding risk interval is calculated based on the risk interval. The generation probability of the corresponding risk level token is mapped to the 0-1 interval using a mapping function. The mapped probability value is scaled proportionally according to the width of the risk interval and superimposed on the lower limit of the corresponding risk interval to generate a thresholded continuous risk score for each preset risk category.
[0012] As a further improvement to this technical solution, in step S5, the generated continuous risk score is compared with the corresponding category preset threshold, and differentiated treatment is performed based on the risk level to output a structured result, including the following steps: Obtain the continuous risk score corresponding to each preset risk category, and read the preset discrimination threshold corresponding to each preset risk category. Compare each continuous risk score with the preset discrimination threshold of the corresponding category to identify the discrimination result corresponding to each preset risk category. Based on the discrimination result and the risk level of the corresponding preset risk category, perform differentiated safety measures. Summarize the discrimination label, risk level, continuous risk score, discrimination reason and handling action corresponding to each preset risk category, generate the corresponding structured results and output them.
[0013] On the other hand, the present invention provides a large model content security protection system based on front-end small model traffic splitting, multi-model arbitration and graded risk thresholding scoring, including a memory, a processor and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the above-mentioned large model content security protection method based on front-end small model traffic splitting, multi-model arbitration and graded risk thresholding scoring.
[0014] Compared with existing technologies, the technical solution of this invention has the following beneficial technical effects: it couples pre-processing small model traffic splitting, multi-model arbitration, hierarchical risk output, and risk level-guided threshold scoring into a complete content security protection system. The pre-processing small model digests the vast majority of simple traffic at extremely low cost, allowing only a small number of difficult samples to enter the heavy pipeline, significantly reducing overall computing power and latency; multi-model parallel prediction plus arbitration selection overcomes the blind spots of single models and improves robustness; hierarchical risk output distinguishes between violations and suspected violations to support differentiated handling; risk level-guided threshold scoring provides continuous scores that are comparable across samples and can be flexibly set by category while maintaining multi-label capabilities. Thus, it simultaneously achieves a unity of high discrimination accuracy, strong robustness, refined handling, and computing power economy in large-scale online traffic scenarios. Attached Figure Description
[0015] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is an overall flowchart of the content security protection system in this embodiment; Figure 3 This is a schematic diagram illustrating the risk level-guided threshold scoring in this embodiment. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0017] Example 1: Please refer to Figure 1-3 As shown, this embodiment provides a method for protecting the content security of large models based on pre-processing small model traffic splitting, multi-model arbitration, and graded risk threshold scoring, including the following steps: S1. Obtain the content to be reviewed, use the pre-built small model to judge and classify it, let obviously safe content pass directly, and directly block obviously illegal content. Send the difficult samples along with the corresponding labels and judgment reasons to the multi-model prediction process. In this embodiment, a pre-processing small model is used for discrimination and traffic separation, allowing obviously safe content to pass directly and blocking obviously illegal content. Difficult samples, along with their corresponding tags and discrimination reasons, are then sent to the multi-model prediction process, including the following steps: Based on all content awaiting review, a lightweight discriminative pre-model is used for initial discrimination, outputting corresponding discrimination scores. Then, based on pre-set upper and lower confidence thresholds, the content is further segmented (e.g., ...). Figure 2 (as shown) Specifically, when the discrimination score is higher than the upper confidence threshold, the content to be reviewed is judged as obviously non-compliant and directly blocked or enters the subsequent processing procedure; when the discrimination score is lower than the lower confidence threshold, the content to be reviewed is judged as obviously safe and directly released; when the discrimination score is between the upper and lower confidence thresholds, the content to be reviewed is judged as a difficult sample and a multi-model prediction process is executed; for difficult samples that require discrimination reasons (such as audits of blocked samples, or preliminary judgment reasons attached when submitting for arbitration), since the front-end small model only outputs discrimination labels, the assistant prefix continuation method is adopted, using the label output by the small model as the assistant response prefix, and calling the large model on the label. Based on the prefix, generate a discrimination reason consistent with the label, and send the difficult samples along with the corresponding label and discrimination reason to the multi-model prediction process (to absorb the majority of simple traffic locally at extremely low cost, and only hand over a small number of difficult samples to the heavy multi-model arbitration pipeline, while maintaining discrimination accuracy and significantly reducing overall computing power and latency overhead; and through prefix continuation, even small models that only output labels can provide auditable reasons for subsequent arbitration and auditing, to absorb the majority of simple traffic locally at extremely low cost, and only hand over a small number of difficult samples to the heavy multi-model arbitration pipeline, while maintaining discrimination accuracy and significantly reducing overall computing power and latency overhead; and through prefix continuation, even small models that only output labels can provide auditable reasons).
[0018] S2. Based on difficult samples, multiple heterogeneous guardrail models are called in parallel to make predictions and output candidate prediction results; In this embodiment, the output of candidate prediction results includes the following steps: Difficult samples are simultaneously sent to multiple heterogeneous guardrail models for parallel prediction. Each guardrail model has a different backbone network, training data, or risk categories it is good at detecting. Each heterogeneous guardrail model independently performs content security judgment based on the same difficult sample, generates corresponding judgment results, and outputs corresponding candidate prediction results, including judgment labels, risk levels, and judgment reasons. The risk level is taken from the ordered set {safe, low, medium, high} to represent the risk level of the current input corresponding to the risk category. In multi-label scenarios, each guardrail model outputs the corresponding label, risk level, and judgment reason for each risk category, and finally aggregates to form multiple candidate prediction results. This approach overcomes the detection blind spots of a single model by dynamically selecting the model best suited for that type of risk for each difficult input, significantly improving judgment robustness. Compared to naive voting, arbitration using reasons makes the selection decision more reliable and interpretable.
[0019] S3. Based on the candidate prediction results, the arbitration model is used to combine the original input and the discrimination reasons of each model to perform arbitration and generate the final discrimination result. In this embodiment, generating the final discrimination result includes the following steps: The arbitration process is selected according to the preset operation mode. When the post-arbitration mode is adopted, the arbitration process is entered after obtaining the candidate prediction results output by all heterogeneous guardrail models. When the pre-routing mode is adopted, the lightweight routing model selects the heterogeneous guardrail model that is most likely to be good at the current difficult input based on the input features to complete the prediction. When multiple candidate prediction results are highly consistent, the corresponding candidate prediction result is directly adopted as the final judgment result. Only when multiple candidate prediction results are inconsistent will the corresponding candidate prediction result be sent to the arbitration big model for arbitration. Obtain candidate prediction results and organize the original input (i.e. the original difficult sample input) and each candidate prediction result into a unified context. Input the context into the arbitration big model and compare each candidate prediction result according to the discrimination reason corresponding to each candidate prediction result. Analyze the sufficiency and consistency of each discrimination reason. Based on the comparison results, output the identifier of the adopted model or the final discrimination result after fusion through the arbitration big model.
[0020] S4. Based on the final judgment result, output the corresponding graded risk level according to each preset risk category, and generate the thresholded continuous risk score corresponding to each risk category by combining the generation probability of the corresponding risk level. In this embodiment, based on the final judgment result, the corresponding graded risk level is output according to each preset risk category, including the following steps: Based on the final judgment result, and reading the corresponding preset risk categories, a corresponding risk level is generated for each preset risk category. The risk level is taken from the ordered set {safe, low, medium, high} (safe means no risk, low / medium correspond to different degrees of "suspected", and high corresponds to "violation"). A single discrete token is used to output the risk level corresponding to each preset risk category so that the generation probability of the corresponding risk level token can be read later. The guardrail judgment is upgraded from a "black and white" hard label to a graded risk expression, which directly supports the differentiated handling of "interception / review / release" and provides a graded basis for the threshold score of key point 4. The method of using a single discrete token to output the risk level corresponding to each preset risk category includes the following steps: Obtain the risk level corresponding to each preset risk category, and establish the correspondence between each risk level and a single discrete token. Each risk level corresponds to a unique discrete token. For each preset risk category, map the corresponding risk level to the corresponding single discrete token, and output the single discrete token corresponding to each preset risk category. This provides input for subsequently reading the generation probability of the corresponding risk level token and generating a continuous risk score.
[0021] Furthermore, the thresholded continuous risk scores corresponding to each risk category are generated, including the following steps: The process involves obtaining the risk level corresponding to each preset risk category and establishing a corresponding risk interval based on each risk level. The continuous score space is divided into multiple fixed and non-overlapping risk intervals (e.g., safe corresponds to [0, 0.25], low to [0.25, 0.5], medium to [0.5, 0.75], and high to [0.75, 1.0]). Based on the risk intervals, the generation probability of the risk level token corresponding to each preset risk category is read (when the model outputs the original predicted values (logits) corresponding to multiple risk level tokens, the original predicted values corresponding to each risk level token are subjected to Softmax renormalization to obtain the normalized generation probability of the corresponding risk level token, and the risk interval and normalized generation probability corresponding to each preset risk category are output). Based on the risk intervals and generation probabilities, a thresholded continuous risk score corresponding to each preset risk category is generated, thus obtaining... It consistently falls within the interval corresponding to the discrete risk level, ensuring cross-sample comparability and calibration stability, while providing fine-grained ranking within the band based on token probability. Each category is calculated independently. This naturally supports multi-label scoring; downstream users can set thresholds for each category (e.g., ...). Figure 3(As shown) (The obtained scores are constrained to fall within the interval band corresponding to the discrete discrimination, ensuring cross-sample comparability and calibration stability, and providing fine-grained ranking within the band through token probability; the threshold can be set arbitrarily on [0,1], and naturally supports setting the threshold separately for each risk category, making up for the shortcomings of existing large model guardrails that cannot achieve both "adjustable threshold" and "multi-label"). The process of generating thresholded continuous risk scores for each preset risk category, based on risk intervals and generation probabilities, includes the following steps: Based on the risk interval, the interval width of the corresponding risk interval is calculated. A mapping function is used to map the generation probability of the corresponding risk level token to the 0-1 interval. The mapped probability value is then scaled proportionally according to the risk interval width and superimposed on the lower limit of the corresponding risk interval to generate a thresholded continuous risk score for each preset risk category. (In the formula, To map the token generation probability to an interval The monotonic mapping function, for example, taking , For the preset risk category index, For category The corresponding discrete risk level takes one of the values from the ordered set {safe, low, medium, high}. Risk level The lower limit of the corresponding risk range, Risk level The upper limit of the corresponding risk range is set so that the generated thresholded continuous risk score always lies within the risk range of the corresponding risk level. At the same time, the generation probability of the corresponding risk level token is used to achieve fine-grained sorting within the same risk range.
[0022] S5. Compare the generated continuous risk score with the corresponding category preset threshold, and perform differentiated treatment based on the risk level, outputting structured results; In this embodiment, the output of structured results includes the following steps: Obtain the continuous risk score corresponding to each preset risk category, and read the preset discrimination threshold corresponding to each preset risk category. Compare each continuous risk score with the preset discrimination threshold of the corresponding category to identify the discrimination result corresponding to each preset risk category. Based on the discrimination result and the risk level of the corresponding preset risk category, perform differentiated security actions (content judged as "violation" (such as high level or score higher than the interception threshold) is directly intercepted; content judged as "suspected" (such as low / medium level or score in the review range) is transferred to manual review; the rest are allowed). Summarize the discrimination label, risk level, continuous risk score, discrimination reason and handling action corresponding to each preset risk category, generate the corresponding structured results and output them for subsequent auditing and data feedback.
[0023] This embodiment utilizes a lightweight front-end small model to inexpensively handle simple traffic, sending only difficult samples to the heavyweight multi-model pipeline, thereby significantly reducing overall computing power. Secondly, it performs multi-model parallel prediction on the samples sent to the pipeline, and the arbitration large model selects the most suitable sub-model result based on the input content, overcoming the blind spot of a single model. Finally, it outputs a hierarchical expression of labels and risk levels to support differentiated handling strategies such as interception, review, and release, and generates thresholded continuous scores based on risk levels. While retaining the ability to discriminate multiple labels, it provides users with flexible thresholds that can be independently adjusted according to risk categories, thereby achieving a fine balance between precision and recall.
[0024] Example 2: This example provides a large model content security protection system based on front-end small model traffic splitting, multi-model arbitration, and graded risk thresholding scoring. It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the large model content security protection method based on front-end small model traffic splitting, multi-model arbitration, and graded risk thresholding scoring described in Example 1.
[0025] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.
Claims
1. A method for protecting the content security of large models based on pre-emptive small-model traffic splitting, multi-model arbitration, and graded risk threshold scoring, characterized in that: include: S1. Obtain the content to be reviewed, use the pre-built small model to judge and classify it, let obviously safe content pass directly, and directly block obviously illegal content. Send the difficult samples along with the corresponding labels and judgment reasons to the multi-model prediction process. S2. Based on difficult samples, multiple heterogeneous guardrail models are called in parallel to make predictions and output candidate prediction results; S3. Based on the candidate prediction results, the arbitration model is used to combine the original input and the discrimination reasons of each model to perform arbitration and generate the final discrimination result. S4. Based on the final judgment result, output the corresponding graded risk level according to each preset risk category, and generate the thresholded continuous risk score corresponding to each risk category by combining the generation probability of the corresponding risk level. S5. Compare the generated continuous risk score with the corresponding category preset threshold, and perform differentiated treatment based on the risk level to output structured results.
2. The method for large model content security protection based on pre-processing small model traffic splitting, multi-model arbitration, and graded risk threshold scoring as described in claim 1, is characterized in that, In step S1, the content to be reviewed is obtained, and a pre-processing small model is used for discrimination and triage. Clearly safe content is directly allowed, while clearly illegal content is directly blocked. Difficult samples, along with their corresponding labels and discrimination reasons, are then sent to the multi-model prediction process, including the following steps: Based on all the content to be reviewed, a lightweight discriminative pre-model is used for preliminary discrimination, and the corresponding discrimination score is output. The content is then split into two streams based on two pre-set confidence thresholds. Specifically, when the discrimination score is higher than the upper confidence threshold, the content to be reviewed is judged as obviously violating and directly blocked; when the discrimination score is lower than the lower confidence threshold, the content to be reviewed is judged as obviously safe and directly released; when the discrimination score is between the upper and lower confidence thresholds, the content to be reviewed is judged as a difficult sample and a multi-model prediction process is executed.
3. The method for large model content security protection based on pre-processing small model traffic splitting, multi-model arbitration, and graded risk thresholding scoring as described in claim 1, is characterized in that... In step S2, multiple heterogeneous guardrail models are called in parallel to perform predictions based on difficult samples, and candidate prediction results are output, including the following steps: Difficult samples are sent to multiple heterogeneous guardrail models for parallel prediction. Each heterogeneous guardrail model independently performs content security judgment based on the same difficult sample, generates corresponding judgment results, and outputs corresponding candidate prediction results, including judgment labels, risk levels, and judgment reasons.
4. The method for large model content security protection based on pre-processing small model traffic splitting, multi-model arbitration, and graded risk thresholding scoring as described in claim 1, characterized in that, In step S3, based on the candidate prediction results, an arbitration model is used to combine the original input and the discrimination reasons of each model to perform arbitration, generating the final discrimination result, including the following steps: Acquire candidate prediction results and organize the original input and each candidate prediction result into a unified context. Input the context into the arbitration big model and compare each candidate prediction result according to the discrimination reason corresponding to each candidate prediction result. Analyze the sufficiency and consistency of each discrimination reason. Based on the comparison results, output the fused final discrimination result through the arbitration big model.
5. The method for large model content security protection based on pre-processing small model traffic splitting, multi-model arbitration, and graded risk thresholding scoring as described in claim 1, characterized in that, In step S4, based on the final judgment result, the corresponding graded risk level is output according to each preset risk category, including the following steps: Based on the final judgment result, and reading the preset risk categories corresponding to the final judgment result, a corresponding risk level is generated for each preset risk category. The risk level is taken from the ordered set {safe, low, medium, high}. A single discrete token is used to output the risk level corresponding to each preset risk category so that the generation probability of the corresponding risk level token can be read later.
6. The method for large model content security protection based on pre-processing small model traffic splitting, multi-model arbitration, and graded risk thresholding scoring as described in claim 1, characterized in that, The method of using a single discrete token to output the risk level corresponding to each preset risk category includes the following steps: Obtain the risk level corresponding to each preset risk category, and establish the correspondence between each risk level and a single discrete token. Each risk level corresponds to a unique discrete token. For each preset risk category, map the corresponding risk level to the corresponding single discrete token, and output the single discrete token corresponding to each preset risk category.
7. The method for large model content security protection based on pre-processing small model traffic splitting, multi-model arbitration, and graded risk threshold scoring as described in claim 1, characterized in that, In step S4, the thresholded continuous risk score for each risk category is generated by combining the generation probability of the corresponding risk level, including the following steps: Obtain the risk level corresponding to each preset risk category, and establish a corresponding risk interval based on each risk level, dividing the continuous score space into multiple fixed and non-overlapping risk intervals; read the generation probability of the risk level token corresponding to each preset risk category based on the risk interval; generate a thresholded continuous risk score corresponding to each preset risk category based on the risk interval and the generation probability.
8. The method for large model content security protection based on pre-processing small model traffic splitting, multi-model arbitration, and graded risk thresholding scoring as described in claim 1, characterized in that, The process of generating thresholded continuous risk scores for each preset risk category based on risk intervals and generation probabilities includes the following steps: The width of the corresponding risk interval is calculated based on the risk interval. The generation probability of the corresponding risk level token is mapped to the 0-1 interval using a mapping function. The mapped probability value is scaled proportionally according to the width of the risk interval and superimposed on the lower limit of the corresponding risk interval to generate a thresholded continuous risk score for each preset risk category.
9. The method for large model content security protection based on pre-processing small model traffic splitting, multi-model arbitration, and graded risk thresholding scoring as described in claim 1, characterized in that, In step S5, the generated continuous risk score is compared with the corresponding category preset threshold, and differentiated treatment is performed based on the risk level to output a structured result, including the following steps: Obtain the continuous risk score corresponding to each preset risk category, and read the preset discrimination threshold corresponding to each preset risk category. Compare each continuous risk score with the preset discrimination threshold of the corresponding category to identify the discrimination result corresponding to each preset risk category. Based on the discrimination result and the risk level of the corresponding preset risk category, perform differentiated safety measures. Summarize the discrimination label, risk level, continuous risk score, discrimination reason and handling action corresponding to each preset risk category, generate the corresponding structured results and output them.
10. A large model content security protection system based on pre-processing small model traffic splitting, multi-model arbitration, and graded risk threshold scoring, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the large model content security protection method as described in any one of claims 1-9, which is based on front-end small model traffic splitting, multi-model arbitration, and graded risk thresholding scoring.