Internet content security filtering method and system based on edge computing

Abstract

The invention provides an internet content security filtering method and system based on edge computing, and the method comprises the steps: obtaining a real-time content data flow from user access flow through a preset module disposed at a network edge device, and obtaining a preliminarily classified content fragment sequence; carrying out feature extraction by adopting an edge filtering algorithm according to the preliminarily classified content fragment sequence, and determining a distribution mode of potential harmful information; if the distribution mode of the potential harmful information exceeds a preset threshold value, local optimization of a content recognition algorithm is executed for the distribution mode, and a refined risk assessment score is obtained; obtaining a high-risk fragment subset from the refined risk assessment score, and judging whether the subset meets a blocking condition or not; performing data isolation processing on the high-risk fragment subset through the judged blocking condition to obtain an isolated security content stream; and comparing and fusing the isolated secure content stream with the original flow, and determining a final filtering output sequence.

Description

Technical Field

[0001] This invention belongs to the field of information processing technology, specifically relating to an internet content security filtering method and system based on edge computing. Background Technology

[0002] Internet content security is a crucial component of cyberspace governance, its core purpose being to protect users from harmful information and maintain a healthy and stable online environment. With the widespread adoption of internet applications, content security issues have become increasingly prominent, serving as a vital link in safeguarding user rights and social harmony. Both individual users and organizations urgently need effective technological means to address harmful information and potential threats online.

[0003] However, current content security filtering methods have significant shortcomings when dealing with complex network environments. Many traditional solutions rely on centralized processing, which is prone to response latency and resource bottlenecks when facing large-scale, high-concurrency network traffic. Especially in the context of extremely fast network content dissemination, centralized processing struggles to meet real-time requirements, resulting in harmful content impacting users before it is blocked. This limitation makes the research into more efficient and user-friendly technologies particularly urgent.

[0004] In this field, the main technical challenge lies in achieving rapid and accurate content identification and blocking close to the user. A core issue is the limitation of computing resources in network edge environments. Since edge devices typically lack powerful computing capabilities, they struggle to perform in-depth analysis of massive amounts of content in a short time. This directly leads to another related problem: how to balance the accuracy and real-time performance of filtering under resource constraints. If this problem cannot be solved, content filtering in edge environments may result in misjudgments or delays. For example, when a user visits a webpage, the system may fail to identify hidden malicious links in time due to insufficient computing power, ultimately exposing the user to risks.

[0005] Therefore, how to achieve both high efficiency and accuracy in content filtering in network edge environments with limited computing resources has become a key problem that this research urgently needs to solve. Solving this problem is not only about technological breakthroughs, but also directly impacts the user's security experience in daily network use. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides an internet content security filtering method and system based on edge computing.

[0007] To achieve the above objectives, the present invention provides the following solution: An internet content security filtering method based on edge computing includes: Real-time content data streams are obtained from user access traffic to obtain a sequence of content fragments after preliminary classification; Based on the sequence of content fragments after preliminary classification, the distribution pattern of potentially harmful information is determined; If the distribution pattern of potentially harmful information exceeds a preset threshold, a refined risk assessment score is obtained. A subset of high-risk segments is obtained from the refined risk assessment score, and it is determined whether the subset meets the blocking conditions. Data isolation processing is performed on a subset of high-risk segments to obtain a secure content stream after isolation. The final filtered output sequence is determined by comparing and fusing the isolated secure content stream with the original traffic.

[0008] As a preferred approach, feature extraction is performed using an edge filtering algorithm based on the pre-classified content fragment sequence to determine the distribution pattern of potentially harmful information.

[0009] Preferably, if the distribution pattern of potentially harmful information exceeds a preset threshold, then local optimization of the content recognition algorithm is performed on the distribution pattern to obtain a refined risk assessment score.

[0010] As a preferred approach, the high-risk segment subset is isolated based on the determined blocking conditions to obtain the isolated secure content stream.

[0011] This invention also provides an internet content security filtering system based on edge computing, comprising: The first processing module is used to obtain real-time content data streams from user access traffic and obtain a sequence of content fragments after preliminary classification. The second processing module is used to determine the distribution pattern of potentially harmful information based on the sequence of content fragments after preliminary classification. The third processing module is used to obtain a refined risk assessment score if the distribution pattern of potentially harmful information exceeds a preset threshold. The fourth processing module is used to obtain a subset of high-risk segments from the refined risk assessment score and determine whether the subset meets the blocking conditions. The fifth processing module is used to perform data isolation processing on a subset of high-risk segments to obtain a secure content stream after isolation. The sixth processing module is used to compare and fuse the isolated secure content stream with the original traffic to determine the final filtered output sequence.

[0012] As a preferred option, the second processing module uses an edge filtering algorithm to extract features based on the content fragment sequence after preliminary classification, and determines the distribution pattern of potentially harmful information.

[0013] Preferably, if the distribution pattern of potentially harmful information exceeds a preset threshold, the third processing module performs local optimization of the content recognition algorithm for the distribution pattern to obtain a refined risk assessment score.

[0014] As a preferred approach, the fifth processing module performs data isolation processing on the high-risk segment subset based on the determined blocking conditions, thereby obtaining the isolated secure content stream.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the inefficiencies, insufficient accuracy, and risks of harmful information spread inherent in traditional content filtering methods when handling massive real-time traffic. These problems stem from the interconnected challenges of limited edge computing resources, dynamically changing potentially harmful patterns, and the need to ensure continuous secure output. This invention acquires real-time content data streams from user access traffic using pre-defined modules and performs initial classification. Then, an edge filtering algorithm extracts features to determine harmful distribution patterns. When a pattern exceeds a threshold, a local optimization of the content recognition algorithm is performed to obtain a refined risk assessment score. A subset of high-risk segments is extracted, and blocking conditions are determined. Data isolation is implemented on the subsets that meet the conditions. Finally, the isolated secure content stream is compared and fused with the original traffic to generate a filtered output sequence, effectively solving the aforementioned problems. This invention highlights the core inventive points of edge real-time optimization and isolation fusion, achieving highly efficient and accurate blocking of harmful information, improving the overall response speed and reliability of network content security, and is suitable for information protection in high-concurrency access scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of an Internet content security filtering method based on edge computing, according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] Example 1 like Figure 1 As shown, this invention provides an internet content security filtering method based on edge computing, comprising: Step S101: Obtain real-time content data streams from user access traffic through preset modules deployed on network edge devices to obtain a pre-classified sequence of content fragments.

[0021] By deploying devices at the network edge, a pre-defined module is used to acquire content data from user access traffic, generating a preliminary sequence of content fragments. The captured access traffic is processed in real-time using this module, decomposing the streaming structure of the content data to obtain initially segmented content fragments. Based on these fragments, a support vector machine (SVM) algorithm is used to extract features and determine the category label for each fragment. If the category label does not match a pre-defined threshold range, the content fragment undergoes secondary segmentation to obtain finer-grained fragment units, and their category assignment is determined. The category assignment results for these fine-grained fragment units are used to construct the contextual relationships of the fragment sequence, resulting in a serialized content structure. For the serialized content structure, the logical dependencies between fragments are analyzed to determine the completeness order of the fragment sequence. Based on the completeness order of the fragment sequence, the final content data stream classification result is generated, outputting a structured content fragment sequence.

[0022] For example, in the deployment of network edge devices, a specific scenario can be envisioned where a pre-defined module extracts content data from user access traffic: an edge node of a content delivery network processes video stream data requested by users in real time. The pre-defined module first captures the user access traffic, decomposes it into a streaming structure, and initially divides it into multiple content segments. Assuming a video stream is divided into 10 segments, each approximately 5 seconds long, the initial classification might identify 6 segments as "sports events" and 4 segments as "advertisements."

[0023] For example, when using the Support Vector Machine (SVM) algorithm for feature extraction, features can be extracted from dimensions such as audio frequency and image brightness variations in a segment, and then a category label can be assigned to each segment. Suppose that the audio frequency of a "sports event" segment fluctuates significantly and the image changes frequently, while the audio of a "commercial break" segment is stable and the image changes are infrequent; the algorithm determines the category based on this. If the confidence score of a segment's category label is lower than a preset threshold of 0.8, a secondary segmentation process is triggered, further dividing the segment into smaller units, such as splitting a 5-second segment into two 2.5-second sub-segments, and re-analyzing their category assignment.

[0024] For example, when constructing contextual relationships for a sequence of segments, the logical relationships between segments can be analyzed based on the category results of fine-grained segments. Suppose a "commercial break" segment is preceded and followed by "sports event" segments; the system can infer that it is interstitial content within a sports event, thus forming a sequential content structure. This contextual relationship helps improve classification accuracy and avoids errors caused by isolated judgments.

[0025] For example, when analyzing the logical dependencies and integrity ordering between segments, the order can be determined by timestamps and the continuity of content themes. Suppose that "sports event" segments in a video stream should be arranged chronologically, while "advertisement" segments are inserted as independent units; the system can generate a complete sequence ordering, ensuring the logical consistency of the content data stream.

[0026] For example, when generating a final structured sequence of content fragments, the system outputs the classification results as an ordered content stream for subsequent applications, such as targeted content delivery or content moderation. This method not only improves classification efficiency but also optimizes user experience and reduces interference from misclassification. By forming a closed loop from initial segmentation to fine-grained analysis, and then to contextual association and integrity sorting, the entire process ensures high-quality output of the content data stream. Its benefits include enhancing the content processing capabilities of edge devices, reducing the burden on central servers, and providing reliable support for real-time application scenarios.

[0027] Step S102: Based on the preliminarily classified content fragment sequence, feature extraction is performed using an edge filtering algorithm to determine the distribution pattern of potentially harmful information.

[0028] Step 1: Obtain categorized sequence data from content fragments and perform preliminary stratification of the categorized sequences using a pre-established rule base to obtain a stratified sequence set. Step 2: For the stratified sequence set, apply an edge filtering algorithm to extract features from each stratum, identifying feature points related to harmful information in each stratum. Step 3: Based on the data distribution of feature points, construct a structured representation of information distribution to obtain the distribution range of potential risks and identify high-risk areas. Step 4: If the feature point density in a high-risk area exceeds a preset threshold, perform deep information analysis on the content fragments in that area to obtain a detailed risk distribution description. Step 5: Using the risk distribution description and pattern recognition technology, classify and label the distribution patterns of harmful information to determine specific pattern categories. Step 6: Based on the pattern categories, generate targeted filtering rules and apply them to subsequent data filtering stages to obtain an optimized content fragment set. Step 7: For the optimized content fragment set, perform a secondary verification of the information distribution to determine if any potential risks have been missed, obtaining the final analysis results.

[0029] For example, in scenarios where network edge devices process user traffic for content classification, a series of methods can be used to stratify and analyze content fragments to gradually optimize data filtering results. First, for the step of obtaining classification sequence data from content fragments, a pre-established rule base can be used to initially stratify the sequences according to content theme, source, or timestamp. Assuming there are 1000 content fragments, the rule base might divide them into three levels: 300 entertainment fragments, 500 news fragments, and 200 social interaction fragments. This stratification facilitates subsequent targeted analysis.

[0030] For example, in the feature extraction stage using edge filtering algorithms, feature points related to harmful information can be identified for each layer of content fragments. Suppose that in news-related layers, by analyzing the sentiment and keyword frequency of the text, the algorithm discovers that 10% of the fragments contain negative emotional words such as "crisis" or "conflict," and these feature points are marked as potential risks.

[0031] For example, when constructing a structured representation of information distribution, feature points can be mapped onto a two-dimensional matrix according to their density distribution. The horizontal axis represents the sentiment value of the content, and the vertical axis represents the frequency of keyword occurrence. If the sentiment value of a certain area is below -0.5 and the keyword frequency is above 5 times per 100 words, it is judged as a high-risk area. This method helps to intuitively locate problematic content. In the stage of deep information analysis of high-risk areas, if the density of feature points in a certain area exceeds a preset threshold, such as 20 out of every 100 segments being marked as high-risk, these segments are further broken down to analyze their contextual semantics. For example, if the phrase "emergency event" appears frequently in a news segment, it may be a true report or a rumor based on the context, requiring more detailed judgment. The step of classifying and labeling the risk distribution description using pattern recognition technology can categorize the distribution patterns of harmful information into rumor, misleading, and malicious categories. For example, if a model is trained using historical data and the semantic pattern of a segment matches the rumor category with 80% accuracy, it is labeled as a rumor. This classification provides a precise basis for subsequent processing. When generating targeted filtering rules, rules can be set according to pattern categories. For example, a keyword blacklist can be set for rumor-related segments to automatically block content segments containing specific words. This type of rule can effectively reduce the spread of harmful information. Finally, when performing secondary verification and optimization of the content segment set, random sampling can be used to check for any omissions. For example, 10% of the remaining 900 segments can be manually reviewed to ensure that no high-risk content is ignored. This approach further improves the reliability of content filtering and ensures the overall security of the data stream.

[0032] Step S103: If the distribution pattern of potentially harmful information exceeds a preset threshold, then perform local optimization of the content recognition algorithm for the distribution pattern to obtain a refined risk assessment score.

[0033] If the distribution pattern of harmful information exceeds a preset threshold, the pattern analysis module performs a preliminary decomposition of the distribution pattern to obtain preliminary analysis data. Based on the preliminary analysis data, content recognition technology is used to perform a deep scan of the decomposed distribution pattern to identify potential risk areas. If the distribution density of a potential risk area is higher than a preset standard, a local optimization strategy is implemented for that area to obtain optimized distribution characteristics. Using the optimized distribution characteristics, a risk assessment model is used to quantify the potential harmful information and obtain a risk assessment score. If the risk assessment score exceeds the safe range, the relevant information processing flow is dynamically adjusted to determine the adjusted risk level. Based on the adjusted risk level, a corresponding information processing strategy is generated to determine the final risk control measures. Through the final risk control measures, the distribution pattern of harmful information is continuously monitored to obtain real-time monitoring data.

[0034] For example, when dealing with potentially harmful information in online content, a series of technical means can be used to analyze and control distribution patterns. The initial decomposition of the pattern analysis module can be understood as breaking down complex distribution patterns into multiple sub-patterns to more precisely identify anomalies. Suppose that on a content platform, the spread of a certain type of information is detected to be abnormally fast, reaching 100,000 reposts per day. Through initial decomposition, this can be broken down into sub-patterns such as geographical distribution, temporal distribution, and user group distribution, and the characteristics of each sub-pattern can be analyzed separately.

[0035] For example, deep scanning for content recognition technology can be viewed as a further refinement of the decomposed sub-patterns.

[0036] In one possible implementation, for a specific geographic distribution sub-pattern, the system scans for high-frequency keywords within a particular region. For example, if the frequency of "rumor"-related words is as high as 30% in a certain region, that region is marked as a potential risk area. This method can help quickly pinpoint the source of the problem, providing accurate evidence for subsequent processing.

[0037] For example, when implementing a local optimization strategy, content in high-density risk areas can be prioritized. Suppose a certain area has a harmful information density of 500 items per hour, exceeding a preset standard of 200 items. The system will automatically adjust the content review frequency for that area from once per hour to once every 30 minutes, thereby reducing the possibility of risk spread. The benefit of this strategy is that it can address problems specifically and reduce resource waste.

[0038] For example, when using risk assessment models for quantification, risk factors can be converted into specific scores. Suppose the distribution characteristics of a certain area show that harmful information accounts for 40%. Combining the speed of dissemination and user reach, the system assesses a risk score of 85, exceeding the safe range by 60 points. This quantification method helps to intuitively determine the risk level and provides data support for subsequent decision-making.

[0039] For example, when dynamically adjusting information processing workflows, if the risk score is too high, the system will automatically switch to a stricter review mode. If the risk level drops from high to medium after adjustment, the system will generate a corresponding handling strategy, such as increasing the proportion of manual review to 50% to ensure content security. The advantage of this dynamic adjustment is its flexibility in responding to different risk scenarios.

[0040] For example, after generating an information processing strategy, the distribution of harmful information can be continuously tracked through a real-time monitoring system to ensure the implementation of final risk control measures. Suppose monitoring data shows that the spread of a certain type of information has decreased to 1,000 times per day after adjustments, far lower than the 5,000 times before the adjustments, proving the control measures are effective. The benefit of this continuous monitoring is that it enables timely detection of problems and optimization of strategies, ensuring long-term security.

[0041] Step S104: Obtain a subset of high-risk segments from the refined risk assessment score, and determine whether the subset meets the blocking conditions.

[0042] The system retrieves refined assessment results from stored risk data, performs preliminary classification on each data set to obtain a preliminary set of high-risk segments. For these high-risk segments, a subset is extracted using pre-defined filtering rules to determine the subsets that meet the risk level criteria. Based on the extracted subsets, risk characteristic information is obtained. If the characteristic information matches pre-defined blocking conditions, the subset is determined to require blocking. Subset analysis is performed on the determined subsets to obtain a detailed risk level distribution and determine if any potential hidden risks exist. Based on the risk level distribution, corresponding condition matching logic is obtained. If the matching logic indicates that the blocking condition is met, the blocking priority of the subset is determined. For subsets with high blocking priority, a logical decision mechanism combined with a support vector machine algorithm is used for final confirmation to obtain the final decision on blocking execution. Based on the final decision data, a blocking command is generated and transmitted to the relevant systems, completing the processing flow for the high-risk segment subsets.

[0043] For example, when processing stored risk data, one can first extract refined assessment results from the database. Suppose a batch of data contains 1000 records, each with a risk score ranging from 0 to 100. After refinement, records with scores higher than 80 are selected as an initial set of high-risk segments. This initial classification can quickly identify segments requiring focused attention, reducing the complexity of subsequent processing.

[0044] For example, when extracting subsets from high-risk segments in the initial set using pre-defined screening rules, the rule could be set to "segments with a risk score higher than 85 for three consecutive days." Assuming 50 segments meet this condition, these segments are categorized as a subset that meets the risk level criteria. This screening method helps to accurately identify segments with persistently high risk, providing a reliable basis for subsequent analysis.

[0045] For example, when acquiring risk characteristics of a subset of data points, features such as "information spread speed" or "number of users involved" can be considered. If a subset shows a daily increase of more than 30% in the spread speed and involves more than 5,000 users, matching the preset blocking conditions, then blocking action is deemed necessary. This feature matching mechanism can promptly identify potential threats and ensure that handling measures are targeted and effective.

[0046] For example, by conducting a detailed analysis of subsets of data to obtain the risk level distribution, it can be discovered that 80% of the data within a subset are concentrated in the high-risk range, while another 20% have potential hidden risks. For these hidden risks, their sources can be further analyzed, such as whether they are related to a specific time period or region. This distribution analysis helps to comprehensively understand the risk situation and avoid overlooking key issues.

[0047] For example, when obtaining conditional matching logic based on risk level distribution results, if the high-risk proportion of a subset exceeds 75% and the propagation trend continues to rise, then the blocking condition is met, and the priority is set to the highest. This logical judgment method can reasonably allocate resources and ensure that high-risk content is processed first.

[0048] For example, when finalizing a subset with high blocking priority using a support vector machine (SVM) algorithm, the model can be trained using historical data to identify the correlation between segment features and blocking requirements. If, after algorithmic analysis, a subset is confirmed to have a blocking necessity of over 90%, a final decision is generated. This combination of logic and algorithms improves the accuracy of the decision.

[0049] For example, when generating blocking instructions and transmitting them to relevant systems, the instructions can be refined to "restrict the spread of specific content" or "suspend related account operations," and transmitted in real time through system interfaces to ensure that high-risk segments are handled promptly. This process design can effectively reduce the possibility of risk spread and ensure overall security.

[0050] Step S105: Based on the determined blocking conditions, data isolation processing is performed on the high-risk segment subset to obtain the isolated secure content stream.

[0051] Using a pre-established risk assessment model, the input content data stream is analyzed segment by segment to identify potential high-risk segments, resulting in a preliminary high-risk segment subset. Based on this subset, pre-defined blocking conditions are applied for secondary filtering. If a segment meets these conditions, it is marked as requiring isolation, thus determining the final high-risk segment set. For this final high-risk segment set, data isolation technology is used to process it hierarchically, separating it from the original content stream to obtain isolated independent data units. These isolated independent data units are then encrypted and their access permissions are restricted, generating protected isolated data modules. Based on the remaining non-high-risk segments in the original content stream, the data sequence is reconstructed to generate a continuous secure content stream. For the generated secure content stream, a consistency verification tool is used for integrity checks. If data loss or anomalies are detected, a repair mechanism is triggered, obtaining a repaired secure content stream. Finally, the repaired secure content stream is standardized to output the final compliant content stream.

[0052] For example, when analyzing content data streams segment by segment using a pre-established risk assessment model, it can be understood as a rule-based and feature-matching identification process. Assuming a content data stream contains multiple text segments, the system analyzes the keywords, semantic tendencies, and other features of each segment to identify segments that may involve sensitive information or illegal content. Assuming there are 100 segments, 5 segments are marked as potentially high-risk segments because they contain specific high-risk words, forming an initial subset. This approach can quickly pinpoint problem areas, providing a foundation for subsequent processing.

[0053] For example, when applying preset blocking conditions for secondary filtering, the initial subset can be further refined using multi-dimensional criteria. Suppose that among these five high-risk segments, two segments not only contain high-risk words but also involve specific context, such as descriptions of privacy-related data; these would be marked as objects requiring isolation. This secondary filtering, by combining context and rules, ensures that the selected segments truly require processing, avoiding misjudgments.

[0054] For example, when using data isolation techniques for the final set of high-risk segments, they can be separated from the original content stream to form independent data units. Assume that after these two segments are isolated, they are stored in a separate temporary database, completely isolated from the original data stream. This layered processing approach effectively prevents high-risk content from interfering with the overall data stream.

[0055] For example, when encrypting and restricting access to isolated data units, symmetric encryption technology can be used to protect the data, and access can be restricted to specific authorized roles. Assuming that the encrypted data modules can only be viewed by the security audit team, and other roles have no access, data security is ensured.

[0056] For example, when reconstructing the data sequence for the remaining non-high-risk segments, the remaining 98 segments can be reassembled in their original order to form a continuous stream of secure content. This reassembly process ensures that the logic and integrity of the data stream are not affected.

[0057] For example, when using a consistency verification tool to perform integrity checks, if a segment of metadata is found to be missing in the reconstructed data stream, the system will trigger a repair mechanism to extract the corresponding data from the backup and complete the data. This detection and repair mechanism ensures the availability of the data stream.

[0058] For example, when standardizing the format of the repaired secure content stream, the text encoding format and structure can be unified to ensure that the output content meets compliance requirements. Assume the final output compliant content stream is adjusted to a unified encoding format, making it easier for downstream systems to process it directly. This standardization process improves the compatibility and application efficiency of the data stream.

[0059] Step S106: Compare and fuse the isolated secure content stream with the original traffic to determine the final filtered output sequence.

[0060] Based on the aforementioned business content and extracted relevant attributes, the following business solution is generated. It revolves around the goal of comparing and fusing secure data with original data to generate a filtered output sequence. The technical process is designed based on interrelated attributes, discarding isolated attributes that cannot form a logical connection. The solution is implemented through the following steps: First, an initial data stream is obtained from the original data. This stream is then categorized according to the various data types it contains, resulting in a pre-classified data set. Second, the data stream is isolated using preset rules based on the pre-classified data set, generating an isolated secure data stream. Third, a content comparison operation is performed between the isolated secure data stream and the original data stream to determine if any discrepancies exist. If discrepancies are detected, the discrepancies are recorded, resulting in a comparison discrepancy dataset. Fourth, a content verification mechanism is used on the comparison discrepancy dataset to filter out data portions that meet security standards, generating a verified secure data subset. Fifth, a data fusion operation is performed on the verified secure data subset, integrating it with the corresponding portions of the original data stream to obtain a fused data stream. Finally, streaming processing technology is applied to the fused data stream to generate filtered output data segment by segment, determining the final output sequence. By performing sequence generation processing on the final output sequence and using data filtering methods to remove content that does not meet the format requirements, a standardized output result is obtained.

[0061] For example, when addressing security issues related to content data streams, a rigorous comparison and fusion process can be gradually built starting from acquiring the raw data stream. Regarding the initial data stream acquisition, let's assume a data stream is extracted from a content platform containing text, images, and metadata, with an initial data volume of 10 million records per day. Through classification processing, different types of data, such as text and images, are categorized to form a preliminary classified dataset, for example, 60% text data, 30% image data, and 10% other metadata, laying the foundation for subsequent processing.

[0062] For example, in the isolation processing stage, the preset rules for filtering the initially categorized dataset might be based on content sensitivity. Suppose that fragments containing specific keywords in the text data are marked as sensitive content, accounting for approximately 5%, and these fragments will be isolated into a separate safe data stream, while the remaining 95% of the content remains in the original data stream. This isolation method ensures that sensitive content does not directly affect the processing of the overall data stream.

[0063] For example, when comparing a secure data stream with the original data stream, differences can be identified by comparing each segment sequentially. Suppose that during the comparison, 2% of the text segments in the secure data stream are marked as sensitive, while the corresponding positions in the original data stream are missing these segments. These differences are then recorded, forming a difference dataset. This type of comparison helps to clarify which content requires further processing.

[0064] For example, in the application of content verification mechanisms, security standards can be set for the differential dataset, such as allowing only non-sensitive content that meets the integrity standards to pass verification. Suppose that 1% of the content in the differential dataset is removed due to integrity issues, the final safe data subset will only contain the parts that meet the standards. This verification ensures the quality of the data subsequently merged.

[0065] For example, in data fusion operations, a verified secure subset of data is integrated with the original data stream. Assuming that the proportion of sensitive content in the fused data stream is reduced to 0.5%, with the remainder being compliant content, this fusion method maintains the integrity of the data stream while mitigating risk.

[0066] For example, for the merged data stream, streaming processing technology is used to generate filtered output data segment by segment. The amount of data processed in each segment can be set to 100,000 records to ensure real-time performance. The final output sequence is filtered segment by segment to ensure the continuity and compliance of the content.

[0067] For example, in sequence generation processing, data filtering methods are used to remove non-conforming content from the final output sequence, such as removing incomplete records. Assuming a removal rate of 0.2%, a standardized output result is obtained. This approach ensures the standardization of the output content and improves the convenience of subsequent use.

[0068] Example 2 This invention also provides an internet content security filtering system based on edge computing, comprising: The first processing module is used to obtain real-time content data streams from user access traffic and obtain a sequence of content fragments after preliminary classification. The second processing module is used to determine the distribution pattern of potentially harmful information based on the sequence of content fragments after preliminary classification. The third processing module is used to obtain a refined risk assessment score if the distribution pattern of potentially harmful information exceeds a preset threshold. The fourth processing module is used to obtain a subset of high-risk segments from the refined risk assessment score and determine whether the subset meets the blocking conditions. The fifth processing module is used to perform data isolation processing on a subset of high-risk segments to obtain a secure content stream after isolation. The sixth processing module is used to compare and fuse the isolated secure content stream with the original traffic to determine the final filtered output sequence.

[0069] As one embodiment of the present invention, the second processing module uses an edge filtering algorithm to extract features based on the preliminarily classified content fragment sequence to determine the distribution pattern of potentially harmful information.

[0070] As one embodiment of the present invention, if the distribution pattern of potentially harmful information exceeds a preset threshold, the third processing module performs local optimization of the content recognition algorithm for the distribution pattern to obtain a refined risk assessment score.

[0071] As one embodiment of the present invention, the fifth processing module performs data isolation processing on the high-risk segment subset based on the determined blocking conditions to obtain the isolated secure content stream.

[0072] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for internet content security filtering based on edge computing, characterized in that, include: Real-time content data streams are obtained from user access traffic to obtain a sequence of content fragments after preliminary classification; Based on the sequence of content fragments after preliminary classification, the distribution pattern of potentially harmful information is determined; If the distribution pattern of potentially harmful information exceeds a preset threshold, a refined risk assessment score is obtained. A subset of high-risk segments is obtained from the refined risk assessment score, and it is determined whether the subset meets the blocking conditions. Data isolation processing is performed on a subset of high-risk segments to obtain a secure content stream after isolation. The final filtered output sequence is determined by comparing and fusing the isolated secure content stream with the original traffic.

2. The Internet content security filtering method based on edge computing as described in claim 1, characterized in that, Based on the preliminarily classified content fragment sequences, an edge filtering algorithm is used to extract features and determine the distribution pattern of potentially harmful information.

3. The Internet content security filtering method based on edge computing as described in claim 2, characterized in that, If the distribution pattern of potentially harmful information exceeds a preset threshold, then the content recognition algorithm is locally optimized for the distribution pattern to obtain a refined risk assessment score.

4. The Internet content security filtering method based on edge computing as described in claim 3, characterized in that, By applying the determined blocking conditions, high-risk segment subsets are isolated to obtain the isolated secure content stream.

5. An internet content security filtering system based on edge computing, characterized in that, include: The first processing module is used to obtain real-time content data streams from user access traffic and obtain a sequence of content fragments after preliminary classification. The second processing module is used to determine the distribution pattern of potentially harmful information based on the sequence of content fragments after preliminary classification. The third processing module is used to obtain a refined risk assessment score if the distribution pattern of potentially harmful information exceeds a preset threshold. The fourth processing module is used to obtain a subset of high-risk segments from the refined risk assessment score and determine whether the subset meets the blocking conditions. The fifth processing module is used to perform data isolation processing on a subset of high-risk segments to obtain a secure content stream after isolation. The sixth processing module is used to compare and fuse the isolated secure content stream with the original traffic to determine the final filtered output sequence.

6. The Internet content security filtering method based on edge computing as described in claim 5, characterized in that, The second processing module uses an edge filtering algorithm to extract features based on the pre-classified content fragment sequence to determine the distribution pattern of potentially harmful information.

7. The Internet content security filtering system based on edge computing as described in claim 6, characterized in that, If the distribution pattern of potentially harmful information exceeds a preset threshold, the third processing module performs local optimization of the content recognition algorithm for the distribution pattern to obtain a refined risk assessment score.

8. The Internet content security filtering system based on edge computing as described in claim 7, characterized in that, The fifth processing module performs data isolation processing on the high-risk segment subset based on the determined blocking conditions, and obtains the isolated secure content stream.

Images