A PCDN service identification method

By obtaining the terminal device category and manufacturer bias score, calculating adaptive weights, and using a gradient boosting tree model and strong rule verification, the problems of misjudgment and unstable judgment in PCDN service identification are solved, and traceable and stable identification results are achieved.

CN123123041APending Publication Date: 2026-11-17SUZHOU MAXNET NETWORK SECURITY TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611636005.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-22
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for identifying PCDN services of home broadband users have problems such as misjudgment, inability to trace exclusion criteria, lack of interaction between device type and network behavior in the judgment, and unstable judgment results due to fixed weights.

Method used

By acquiring terminal device category, manufacturer bias score, and numerical behavior features, adaptive weights are calculated. A gradient boosting tree model combined with strong rule verification is used to generate non-PCDN explanatory labels with evidence, thus comprehensively identifying PCDN services.

Benefits of technology

It achieves traceable PCDN service identification, avoids misjudgment, maintains the stability of the judgment criteria, adapts to changes in service distribution, and provides interpretable identification results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123041A_ABST
    Figure CN123123041A_ABST
Patent Text Reader

Abstract

This invention discloses a PCDN service identification method. It obtains the terminal device category, terminal manufacturer, terminal type bias score, and numerical behavioral features of the downstream device. The method uses the absolute value of the correlation coefficient between the numerical behavioral features and the PCDN behavior anchor point, normalized, to obtain a daily adaptive weight. This weighted weight is then added to the terminal type bias score to determine the daily behavior score. Downstream devices whose comprehensive quantile scores reach the rejection quantile threshold generate a non-PCDN explanation label with evidence while maintaining the daily behavior score. Numerical behavioral features, terminal manufacturer, and terminal device category are used to construct a categorical feature using a gradient boosting tree to obtain a gateway PCDN prediction flag. Strong rule verification fails, and this flag is removed. The identification result is obtained by combining the non-PCDN explanation label and the PCDN prediction flag. This method achieves terminal fingerprint and behavioral data fusion judgment, excludes actions without modifying the behavior score, ensures traceability, and limits false alarms.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of computer network security and big data processing and analysis, and in particular to a PCDN service identification method. Background Technology

[0002] As content delivery network (CDN) technology extends to the network edge, home broadband users' terminal devices are widely used as edge nodes for content distribution. This results in PCDN services operating continuously online and occupying long-term uplink bandwidth, altering the composition of home broadband uplink traffic. Such services impact resource planning in operator access networks and metropolitan area networks, user service quality, and network management order, making it increasingly urgent for operators to identify and classify PCDN services at the broadband gateway level.

[0003] In operator network management practices, deep packet inspection systems deployed on the gateway side can record application flow details of connected devices, while traffic acquisition systems can obtain traffic data from the gateway's WAN and LAN ports. Based on this data, existing identification methods typically operate on two levels: first, by setting thresholds or rules based on uplink traffic, application identifiers, destination ports, and domain name access records to filter and label gateways or accounts; second, by training a classification model using historical labeled samples to perform binary classification prediction on the gateway's traffic statistical characteristics. Furthermore, information such as terminal manufacturer, model, and category obtained from the associated connected device identifiers is generally used for result display or auxiliary statistics in the existing process.

[0004] The above identification method is constrained by the following issues in actual deployment:

[0005] First, the screening and labeling actions directly affect the sample set involved in statistical calculations. When misjudged objects are removed, zeroed out, or exempted, the overall sample population used to calculate the distribution of business characteristics and the model input data is modified by the handling criteria. The statistical criteria and the handling criteria are superimposed on the same data set, forming a feedback loop in which the strategy actions and statistical results influence each other, causing the judgment criteria for subsequent business days to shift.

[0006] Secondly, the exclusion process lacks traceable criteria. Live streaming, peer-to-peer downloads, and camera backhaul services share similar uplink traffic patterns with PCDN services. Current methods only provide a conclusion for excluding these services, without recording the basis for that conclusion. Therefore, the exclusion conclusion cannot be verified, nor can it be corrected in case of subsequent changes in conditions.

[0007] Third, terminal information representing the category of the downstream device and application data representing the network behavior of the downstream device do not participate in the judgment link, and the behavioral differences undertaken by different categories of devices within the same gateway cannot enter the judgment process.

[0008] Fourth, the importance of each behavioral dimension is determined by fixed weights or human experience, which cannot be adjusted with changes in the sample distribution of regions and time periods, resulting in decreased stability of the judgment results when business distribution drifts. Summary of the Invention

[0009] In view of the above problems, a PCDN service identification method is proposed to overcome or at least partially solve the above problems, including:

[0010] Obtain the terminal device category, terminal manufacturer, terminal type bias score, and numerical behavioral characteristics of the connected devices to the gateway;

[0011] The original weights are obtained by summing the absolute values ​​of the correlation coefficients between numerical behavioral features and each PCDN behavioral anchor point. Each original weight is then normalized to obtain the daily adaptive weight corresponding to each numerical behavioral feature.

[0012] The participation score of each numerical behavior feature is weighted and summed with the corresponding adaptive weight for the day, and then the terminal type bias score is added to obtain the behavior score of the downstream device for the day.

[0013] Based on numerical behavioral characteristics, calculate the non-PCDN comprehensive percentile score of the connected device. For connected devices whose comprehensive percentile score reaches the rejection percentile threshold, generate a non-PCDN explanatory label with evidence and retain the behavioral score of the day.

[0014] The numerical behavioral features and model input information are fed into the gradient boosting tree to obtain the gateway PCDN prediction label. The model input information includes categorical features composed of terminal manufacturers and terminal device categories.

[0015] Clear the PCDN prediction flag of the gateway that fails strong rule verification;

[0016] Read the main device category in the terminal device category, and for the subordinate devices whose main device category is outside the preset exemption category set, remove the live broadcast category non-PCDN interpretation tags;

[0017] The non-PCDN interpretation label and PCDN prediction flag of the downstream device with the largest uplink traffic in the integrated gateway are used to obtain the gateway's PCDN service identification result.

[0018] Optionally, obtaining numerical behavioral characteristics includes:

[0019] Load the deep packet inspection application flow details for the specified business day; use a hash function to map the gateway identifier to several parallel computing buckets; in each bucket, obtain the application identifier, destination IP, destination port, and flow count of each downstream device from the deep packet inspection application flow details; calculate the numerical behavioral characteristics of each downstream device based on the application identifier, destination IP, destination port, and flow count;

[0020] Among them, numerical behavioral characteristics include application-side behavioral characteristics, destination IP-side behavioral characteristics, and port-side behavioral characteristics;

[0021] Application-side behavioral characteristics include the highest value of application traffic share, the sum of traffic shares of the three applications with the highest traffic share, application distribution entropy, PCDN characteristic application share, and negative application share; negative application share includes the traffic share of live streaming application set, the traffic share of peer-to-peer download application set, and the traffic share of camera application set;

[0022] The behavioral characteristics of the destination IP side include the highest value of the destination IP traffic share, the sum of the traffic shares of the three destination IPs with the highest traffic share, the destination IP distribution entropy, and the destination IP concentration index;

[0023] Port-side behavioral characteristics include the highest percentage of destination port traffic, the percentage of traffic from live streaming port sets, and the percentage of traffic from general web port sets; application traffic percentage is the ratio of traffic from a single application to the total traffic of the downstream device; destination IP traffic percentage is the ratio of traffic from a single destination IP to the total traffic of the downstream device; and destination port traffic percentage is the ratio of traffic from a single destination port to the total traffic of the downstream device.

[0024] Optionally, obtaining the terminal device category, terminal manufacturer, and terminal type bias score includes: selecting the device identifier of the downstream device from the application flow details of deep packet inspection, associating the device identifier with the terminal fingerprint database to obtain the terminal device category and terminal manufacturer; assigning the terminal type bias score according to the terminal device category, assigning a positive bias score to the terminal device category that tends to run PCDN services, assigning a negative bias score to the camera-type terminal device category, and assigning a zero bias score to the remaining terminal device categories.

[0025] Optionally, calculating the adaptive weights for the day includes:

[0026] The calculation window is defined as a time window consisting of a preset number of historical backtracking days; the value sequence of each numerical behavioral feature within the calculation window is taken; the value sequence of each PCDN behavioral anchor point within the calculation window is taken; the daily adaptive weight is calculated using the following formula:

[0027] W i =|r(f i ,A1)|+|r(f i ,A2)|+|r(f i A3)|

[0028] w i =W i ÷ΣW i

[0029] In the formula, f iFor the i-th numerical class behavioral feature, r(f) i A1) represents the numerical behavioral feature f. i The correlation coefficient between the value sequence of point A1 and the value sequence of anchor point A1, r(f i A2) represents the numerical behavioral feature f. i The correlation coefficient between the value sequence of point A1 and the value sequence of anchor point A2, r(f i A3) represents the numerical behavioral feature f. i The correlation coefficient W between the value sequence of point A3 and the value sequence of anchor point A3. i For numerical behavior features f i The original weights, w i For numerical behavior features f i The adaptive weights of the day, ΣW i It is the sum of the original weights of all numerical behavioral features; Anchor point A1 is the application ratio of PCDN features, anchor point A2 is the degree of application homogeneity, anchor point A3 is the highest value of the destination IP traffic ratio, the degree of application homogeneity is equal to 1 minus the application distribution entropy, and the application distribution entropy is the normalized information entropy of the application distribution.

[0030] Optionally, the daily behavior score of the attached device is calculated using the following formula:

[0031] S=Σ[w i ×s(f i )]+b,

[0032] When numerical behavior characteristics f i When considering entropy-based behavioral features and negative application proportion-based behavioral features,

[0033] s(f i ) = 1 - p i ,

[0034] When numerical behavior characteristics f i When considering numerical behavioral features other than entropy-based behavioral features and negative application proportion-based behavioral features,

[0035] s(f i ) = p i ,

[0036] In the formula, S represents the daily behavioral score of the attached device, and w i For numerical behavior features f i The adaptive weights of the day, p i For numerical behavior features f i The daily quantile, s(f i ) represents the numerical behavioral feature f iThe participation score is b, the terminal type bias score is b, and Σ[wi×s(fi)] is the sum of the weighted participation scores of all numerical behavioral features.

[0037] The daily quantile refers to the relative ranking position of numerical behavioral features within all samples on that day; all samples on that day refer to the numerical behavioral features of all connected devices on that day; entropy-based behavioral features include application distribution entropy and destination IP distribution entropy; negative application proportion behavioral features include the traffic proportion of live streaming application sets, the traffic proportion of peer-to-peer download application sets, and the traffic proportion of camera application sets.

[0038] Optionally, generating non-PCDN explanatory tags with evidence includes:

[0039] Calculate the overall percentile score for live streaming applications based on their traffic share; calculate the overall percentile score for peer-to-peer download applications based on their traffic share; calculate the overall percentile score for camera applications based on their traffic share; generate a non-PCDN explanation tag when the highest value among the three overall percentile scores reaches the rejection percentile threshold.

[0040] Candidate evidence is obtained, including application evidence, terminal evidence, and port evidence. Application evidence is the application identifier of the downstream device, terminal evidence is the terminal device category of the downstream device, and port evidence is the destination port of the downstream device. The main evidence is selected from the candidate evidence in the order of application evidence, terminal evidence, and port evidence, with the selection order being: first, application evidence is selected; if application evidence is missing, terminal evidence is selected; if terminal evidence is missing, port evidence is selected. The main evidence is associated with a non-PCDN interpretation tag to obtain a non-PCDN interpretation tag with evidence.

[0041] Optionally, the gateway PCDN prediction flags include:

[0042] Load gateway WAN port traffic data and LAN port traffic data; extract the gateway daily traffic sequence from the gateway WAN port traffic data and LAN port traffic data;

[0043] The inter-day correlation coefficient, root mean square error, coefficient of variation, and burstiness of the daily traffic series of adjacent business days of the gateway are used as the extracted features.

[0044] The characteristics obtained from the deep packet inspection application flow details are: the proportion of home broadband access to the gateway, the application flow recognition rate of the gateway, the number of times and types of access to specific domain names of the gateway, the port traffic distribution of the gateway, and the traffic summary of the gateway's preset application set.

[0045] The extracted features and statistically obtained features are included in the model input information; the model input information and numerical behavioral features are input into the gradient boosting tree, and the gateway PCDN prediction flag is output.

[0046] Optionally, clearing gateway PCDN prediction flags that fail strong rule validation includes:

[0047] For gateways carrying the PCDN prediction flag, the pass conditions are verified sequentially. The pass conditions are: the gateway's uplink traffic on the current day is higher than the preset uplink threshold for the current day; the gateway's average uplink traffic over the past few days is higher than the preset average uplink threshold; the gateway is not on the list of gateways with sudden characteristics; the gateway's home broadband access ratio is higher than the preset home broadband access ratio threshold; the PCDN prediction flag is retained for gateways that meet all the pass conditions; the PCDN prediction flag is cleared for gateways that do not meet the pass conditions.

[0048] Optionally, clearing non-PCDN explanation tags for live streaming includes:

[0049] Non-PCDN explanation tags for live streaming that meet the misidentification filtering criteria based on the proportion of home broadband access and the application stream recognition rate will be removed.

[0050] Optionally, the non-PCDN interpretation tag and PCDN prediction flag of the downstream device with the largest uplink traffic within the integrated gateway are used to obtain the gateway PCDN service identification result, including:

[0051] Based on the uplink traffic, select the downstream device with the largest uplink traffic within the gateway;

[0052] Calculate the proportion of uplink traffic of the gateway to the total uplink traffic of the connected devices (cameras) within the gateway.

[0053] Gateways whose uplink traffic accounts for a preset threshold of their total uplink traffic are classified as high-traffic non-PCDN users. Gateways whose downstream device terminal manufacturers belong to a preset set of camera-determined manufacturers are also classified as high-traffic non-PCDN users. The gateway with the evidence-based non-PCDN interpretation tag of the downstream device with the largest uplink traffic is read. Gateways carrying the non-PCDN interpretation tag are classified as high-traffic non-PCDN users. Gateways that are not high-traffic non-PCDN users but carry the PCDN prediction flag are classified as PCDN users. Gateways that are not high-traffic non-PCDN users and do not carry the PCDN prediction flag are classified as normal users. The gateway PCDN service identification result is obtained.

[0054] This invention uses gateway-connected devices as the discrimination unit, incorporating terminal device category and terminal manufacturer into the judgment chain in two ways: the terminal type bias score is superimposed on the daily behavior score, and the terminal manufacturer and terminal device category constitute categorical features that enter the model input information of the gradient boosting tree model. Thus, information characterizing device attributes and data characterizing network behavior work together in the same judgment process, allowing the behavioral differences undertaken by different categories of devices within the same gateway to be included in the discrimination process, thereby constraining the sources of misjudgment at the device category level.

[0055] In this invention, exclusion actions for non-PCDN services are expressed in the form of non-PCDN explanation tags with evidence, while maintaining the daily behavior score when generating tags. The daily adaptive weight is calculated by the correlation coefficient and normalization between numerical behavior features and each PCDN behavior anchor point, depending on the sample set participating in the statistical calculation. Exclusion actions are expressed in the form of tags without changing the sample set participating in the statistical calculation. The statistical caliber and the handling caliber are independent of each other, and the strategy actions no longer affect the judgment criteria of subsequent business days through sample distribution, thus avoiding the problem of the judgment criteria accumulating and shifting with the handling actions. At the same time, the explanation tags carry the judgment criteria, and the exclusion conclusions are traceable and verifiable. For downstream devices whose main device category is outside the preset exemption category set, the live broadcast non-PCDN explanation tags are cleared, and the exclusion conclusions can be corrected when conditions change. The false negatives caused by the exclusion actions do not accumulate with the business days.

[0056] The daily adaptive weights for each numerical behavioral feature are obtained by normalizing the sum of the absolute values ​​of the correlation coefficients between the feature and each PCDN behavioral anchor point, and the weights are adjusted according to the sample distribution. Compared with the fixed weight method, the behavioral score can adjust with changes in the sample distribution of region and time period, maintaining the effectiveness of the judgment criteria when the business distribution drifts.

[0057] Based on the above mechanism, the gradient boosting tree model uses numerical behavioral characteristics and model input information to provide a gateway PCDN prediction flag. Gateway PCDN prediction flags that fail strong rule verification are removed. Finally, the non-PCDN interpretation label of the downstream device with the highest uplink traffic within the gateway is combined with the PCDN prediction flag to obtain the gateway PCDN service identification result. Each step is interconnected with traceable evidence, and the identification process is interpretable, providing a basis for PCDN service identification and classification on the operator's broadband gateway side. Attached Figure Description

[0058] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the present invention will be briefly introduced below. Obviously, the accompanying 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.

[0059] Figure 1 This is a flowchart of a PCDN service identification method provided in an embodiment of the present invention;

[0060] Figure 2 This is a flowchart of numerical behavior feature acquisition provided in an embodiment of the present invention;

[0061] Figure 3 This is a flowchart of terminal fingerprint association and terminal type bias score determination provided in an embodiment of the present invention;

[0062] Figure 4 This is a flowchart of the generation and removal of non-PCDN interpretation tags with evidence provided in an embodiment of the present invention;

[0063] Figure 5 This is a flowchart of gradient boosting tree prediction and strong rule verification provided in an embodiment of the present invention;

[0064] Figure 6 This is a flowchart of the comprehensive determination of gateway PCDN service identification results provided in an embodiment of the present invention. Detailed Implementation

[0065] 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. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0066] Reference Figures 1 to 6 This invention provides a PCDN service identification method, which may specifically include:

[0067] Obtain the terminal device category, terminal manufacturer, terminal type bias score, and numerical behavioral characteristics of the connected devices to the gateway;

[0068] The original weights are obtained by summing the absolute values ​​of the correlation coefficients between numerical behavioral features and each PCDN behavioral anchor point. Each original weight is then normalized to obtain the daily adaptive weight corresponding to each numerical behavioral feature.

[0069] The participation score of each numerical behavior feature is weighted and summed with the corresponding adaptive weight for the day, and then the terminal type bias score is added to obtain the behavior score of the downstream device for the day.

[0070] Based on numerical behavioral characteristics, calculate the non-PCDN comprehensive percentile score of the connected device. For connected devices whose comprehensive percentile score reaches the rejection percentile threshold, generate a non-PCDN explanatory label with evidence and retain the behavioral score of the day.

[0071] The numerical behavioral features and model input information are fed into the gradient boosting tree to obtain the gateway PCDN prediction label. The model input information includes categorical features composed of terminal manufacturers and terminal device categories.

[0072] Clear the PCDN prediction flag of the gateway that fails strong rule verification;

[0073] Read the main device category in the terminal device category, and for the subordinate devices whose main device category is outside the preset exemption category set, remove the live broadcast category non-PCDN interpretation tags;

[0074] The non-PCDN interpretation label and PCDN prediction flag of the downstream device with the largest uplink traffic in the integrated gateway are used to obtain the gateway's PCDN service identification result.

[0075] In this embodiment of the invention, PCDN refers to Content Delivery Network Edge Node Service, i.e., network behavior where user-side devices participate in content distribution and generate abnormal uplink traffic. The method in this embodiment is executed by the operator-side computing device, with a processing cycle of business days. A business day refers to a data processing cycle divided by calendar days, and the currently processed business log is the current day. The method in this embodiment uses gateway-connected devices as the discrimination unit. A gateway-connected device refers to a terminal accessing the local area network through a gateway, and the gateway-represented terminal refers to the gateway-connected device with the highest uplink traffic within the gateway.

[0076] The computing device acquires the terminal device category, terminal manufacturer, terminal type bias score, and numerical behavioral characteristics of the downstream devices. It reads the device identifier of each downstream device from the daily deep packet inspection application flow details, queries the terminal fingerprint database using the device identifier as the key to obtain the terminal device category, terminal manufacturer, and main device category for each downstream device; it then queries a preset bias configuration table based on the terminal device category to obtain the terminal type bias score for each downstream device; finally, it groups and statistically analyzes the daily deep packet inspection application flow details by device identifier to obtain the numerical behavioral characteristics of each downstream device. This acquisition process generates two intermediate records: one is a terminal fingerprint association record, recording the device identifier, terminal device category, terminal manufacturer, main device category, and terminal type bias score; the other is a behavioral characteristic record, recording the values ​​of the device identifier and the numerical behavioral characteristics.

[0077] The computing device calculates the sum of the absolute values ​​of the correlation coefficients between each numerical behavior feature and each PCDN behavior anchor point to obtain the original weight. It then normalizes each original weight to obtain the daily adaptive weight corresponding to each numerical behavior feature. Within a time window consisting of a preset historical backtracking period, the computing device takes the window statistics of each numerical behavior feature on a business day basis to form a feature value sequence, and takes the window statistics of each PCDN behavior anchor point on a business day basis to form an anchor point value sequence. It calculates the correlation coefficient between the feature value sequence and each anchor point value sequence, sums the absolute values ​​of the correlation coefficients, and obtains the original weight corresponding to that numerical behavior feature. Using the sum of the original weights of all numerical behavior features as the base, it normalizes each original weight to obtain the daily adaptive weight corresponding to each numerical behavior feature. The daily adaptive weight is recalculated for each business day, and all connected devices on the same business day share the same set of daily adaptive weights.

[0078] The computing device calculates the daily behavior score of each sub-device by multiplying the participation score of each numerical behavior feature with its corresponding daily adaptive weight and then adding the terminal type bias score. The computing device first aggregates the numerical behavior features of all sub-devices for the day to form the total sample for that day. Then, for each sub-device, it multiplies the participation score of each numerical behavior feature with its corresponding daily adaptive weight, sums the results, and then adds the terminal type bias score of that sub-device to obtain its daily behavior score. The daily behavior score is written into the rolling summary of behavior scores, which is maintained on a daily basis, recording the daily behavior score for each sub-device.

[0079] In this embodiment, the rolling summary result of behavior scores adds three rolling statistical fields to the daily behavior score field: intra-group quantile, rolling mean, and stability index. Intra-group quantile refers to the relative ranking position of the downstream device's daily behavior score among all downstream devices within the gateway. The calculation method is the same as for the daily quantile, and the ranking object is the daily behavior score of the downstream devices within the gateway. The rolling mean refers to the arithmetic mean of the behavior scores of the downstream device over a preset number of consecutive rolling days prior to the current day. The stability index refers to the standard deviation of the behavior scores of the downstream device over a preset number of consecutive rolling days prior to the current day. An increase in the stability index indicates increased fluctuation in the daily behavior score of the downstream device within the preset rolling days. The preset rolling days is a positive integer parameter set during deployment. The preset rolling days and the preset historical backtracking days are two independent parameters. The preset historical backtracking days are used for daily adaptive weight calculation, while the preset rolling days are used for rolling statistics of behavior scores. The three rolling statistical fields, along with the daily behavior score, are written into the rolling summary of behavior scores for manual review and subsequent judgment.

[0080] The computing device calculates the non-PCDN comprehensive percentile score of the connected devices based on numerical behavioral characteristics. For connected devices whose comprehensive percentile score reaches the rejection percentile threshold, an evidence-based non-PCDN explanatory label is generated. The label generation action retains the behavioral score for that day. The label generation action does not delete the connected device sample, modify the numerical behavioral characteristics, change the participating scores, or modify the daily behavioral score. The evidence-based non-PCDN explanatory label is written as an independent field in the rolling summary result of the behavioral score and stored alongside the score record. The evidence-based non-PCDN explanatory label can be further subdivided into live streaming non-PCDN explanatory labels, peer-to-peer download non-PCDN explanatory labels, and camera non-PCDN explanatory labels.

[0081] The computing device inputs numerical behavioral features and model input information into a gradient boosting tree to obtain the gateway PCDN prediction label. The model input information includes categorical features composed of terminal manufacturers and terminal device categories. The gradient boosting tree outputs at the gateway granularity. The model input data is aggregated by gateway. The aggregation method is to take the numerical behavioral features of the gateway's representative terminal, along with the categorical features composed of the terminal manufacturer and terminal device category of the gateway's representative terminal, and include them in the gateway's input feature vector. The terminal fingerprint information of the devices connected to the gateway is indexed by the gateway identifier, recording the terminal fingerprint association fields and uplink traffic of each device connected to the gateway, and registering the device identifier of the gateway's representative terminal. The terminal manufacturer and terminal device category of the gateway's representative terminal are read from the terminal fingerprint information of the devices connected to the gateway. The gradient boosting tree is trained using PCDN samples and non-PCDN samples verified and labeled by historical business days. The same feature caliber is used for training and prediction. During training, sample class weights are set for PCDN samples and non-PCDN samples. The sample class weights are used to weight the sample imbalance in the training loss. The sample class weights are parameters set during deployment. The gradient boosting tree output gateway PCDN prediction flag can also output prediction probabilities.

[0082] The computing device clears the PCDN prediction flag for gateways that fail strong rule verification. The computing device performs strong rule verification on gateways carrying the PCDN prediction flag, checking each preset verification condition. For gateways that fail any preset verification conditions, the PCDN prediction flag is cleared, and the reason for clearing is recorded. After clearing the PCDN prediction flag, the gateway is treated as not carrying the PCDN prediction flag in subsequent comprehensive judgments.

[0083] The computing device reads the main device category from the terminal device category. For downstream devices whose main device category is outside the preset exemption category set, the live streaming non-PCDN interpretation tag is cleared. For each downstream device carrying a live streaming non-PCDN interpretation tag, the computing device reads the main device category of the downstream device from the terminal fingerprint association record and compares the main device category with the preset exemption category set. If the main device category is within the preset exemption category set, the live streaming non-PCDN interpretation tag is retained; if the main device category is outside the preset exemption category set, the live streaming non-PCDN interpretation tag is cleared. The clearing action is implemented by updating the tag validity status, and the clearing operation is completed before the comprehensive judgment reads the tag.

[0084] The computing device integrates the non-PCDN interpretation tag and PCDN prediction flag of the gateway representative terminal to obtain the gateway PCDN service identification result. The computing device reads the evidence-bearing non-PCDN interpretation tag from the behavioral score rolling summary result of the gateway representative terminal, combines the read result with the gateway PCDN prediction flag to make a comprehensive judgment, and outputs the gateway PCDN service identification result. The gateway PCDN service identification result identifies the service category to which the gateway belongs, for use by the operator's broadband gateway side for service classification and management.

[0085] The terminal device category refers to the device category to which the downstream device belongs; the terminal manufacturer refers to the manufacturer of the downstream device; and the main device category refers to the parent category group to which the terminal device category belongs. For example, if the terminal device category is a downstream device for a home network camera, the main device category is "camera"; if the terminal device category is a downstream device for an internet TV set-top box, the main device category is "set-top box." The terminal fingerprint database is a pre-defined mapping database that records the correspondence between device identifiers and terminal device categories, terminal manufacturers, and main device categories. The terminal type bias score is a pre-set score based on the terminal device category's tendency to run PCDN services. The terminal type bias score is used to apply prior terminal category information to the daily behavior score.

[0086] Numerical behavioral characteristics refer to numerical indicators that characterize the distribution of network behavior of connected devices. These include application-side behavioral characteristics, destination IP-side behavioral characteristics, and port-side behavioral characteristics. Specific indicators and statistical methods for each dimension are described in the corresponding embodiments below. Numerical behavioral characteristics, participation scores, and daily behavioral scores are three distinct quantities: numerical behavioral characteristics are the raw statistics; participation scores are calculated from the relative ranking position of the raw statistics within the entire sample for that day; and the daily behavioral score is obtained by weighted summation of participation scores and superimposing a terminal type bias score. The entire sample for that day refers to the sample set composed of the numerical behavioral characteristics of all connected devices on that day.

[0087] PCDN behavior anchors are referential behavioral indicators used to measure the similarity between the behavior of downstream devices and typical PCDN business behaviors. PCDN behavior anchors include multiple anchors, the semantics and value methods of each anchor are described in the corresponding embodiments below. The window statistical value is the statistical measure of numerical behavior features across all samples on a single business day. It can use the mean, median, or preset quantile values. The statistical caliber is set during deployment and remains consistent across all business days. The correlation coefficient uses the Pearson correlation coefficient. The daily adaptive weight reflects the overall correlation strength between numerical behavior features and all PCDN behavior anchors within the time window.

[0088] The non-PCDN comprehensive quantile score refers to a quantity calculated from the quantiles of numerical behavioral features, measuring the similarity between the behavior of the connected device and the business models of live streaming, peer-to-peer downloading, and camera services. The specific calculation method is described in the corresponding embodiments below. The rejection quantile threshold refers to a preset threshold parameter for generating a non-PCDN explanation label with evidence, with a value ranging from 0 to 1. The non-PCDN explanation label with evidence indicates that the connected device is more likely to belong to non-PCDN services such as live streaming, peer-to-peer downloading, and camera services. This label carries evidence information from the application dimension, terminal dimension, and port dimension. The method of selecting and writing evidence is described in the corresponding embodiments below.

[0089] Categorical features refer to classification features with values ​​belonging to a finite set. Terminal manufacturers and terminal device categories belong to categorical features. Categorical features can be converted into numerical representations using one-hot encoding or label encoding and then input into a gradient boosting tree, or they can be directly processed by a gradient boosting tree that supports categorical feature processing. The gateway PCDN prediction flag and the gateway PCDN service identification result are two different quantities: the gateway PCDN prediction flag is an intermediate label output by the gradient boosting tree and corrected by strong rule validation, while the gateway PCDN service identification result is the final service category output by comprehensive judgment.

[0090] Strong rule verification refers to a set of preset verification conditions that correct the model prediction results. These conditions are set based on the gateway's uplink traffic volume, traffic burst characteristics, and broadband access attributes. The specific content of each preset verification condition is described in the corresponding embodiments below. The preset exemption category set refers to the set of main device categories that are allowed to retain live streaming non-PCDN interpretation tags. Set-top box-type devices running live streaming services are considered normal business operations, and set-top box-type main device categories can be included in the preset exemption category set.

[0091] The removal of non-PCDN explanation tags for live streaming is aimed at circumventing live streaming scenarios. PCDN services operating in live streaming mode can present service traffic as live streaming traffic. When a non-PCDN explanation tag for live streaming is established on a non-live streaming terminal, the reliability of the tag decreases, and its continued validity will lead to circumvention judgment. After removing the non-PCDN explanation tag for live streaming from a downstream device whose main device category is outside the preset exemption category set, the downstream device continues to participate in PCDN judgment.

[0092] The reason for maintaining the daily behavior score is to separate the statistical caliber from the handling caliber. The calculation of the daily adaptive weight depends on the feature value sequence and anchor point value sequence within the time window, and the calculation of the score depends on the distribution of all samples on that day. If the action of generating a label simultaneously deletes a sample or rewrites the score, the statistical values ​​of all samples on that day and the historical window values ​​of subsequent business days will change with the handling action. The caliber used in the statistical calculation and the handling caliber will become mixed, and the sample set on which the weight calculation and quantile calculation of subsequent business days depend will shift, and the judgment criteria will cumulatively shift accordingly. When generating labels with evidence-based non-PCDN explanations, the daily behavior score is maintained, and the label records are stored in parallel as independent fields. Exclusion opinions are expressed in a bypass form, the statistical caliber and the handling caliber are independent of each other, and the judgment criteria do not shift with the handling action.

[0093] The adaptive weights are recalculated daily as the business day progresses, and the behavioral scores can be adjusted to follow changes in the sample distribution by region and time period. Compared with the fixed weight method, the weight values ​​are synchronized with the business distribution of the day.

[0094] Numerical behavioral features are calculated at the device-level granularity, while gradient boosting trees are output at the gateway granularity. The conversion between the two granularities is completed through the gateway representing the terminal. The numerical behavioral features of the gateway representing the terminal and the categorical features composed of the terminal manufacturer and terminal device category of the gateway representing the terminal are included in the input feature vector of the gateway granularity.

[0095] The preset historical backtracking days, statistical caliber, veto quantile threshold, preset exemption category set, and sample category weights involved in this embodiment are all configurable policy parameters, set by the operator based on network scale and service characteristics during deployment. The steps of the method in this embodiment are executed sequentially on a daily basis, with intermediate results from previous steps passed to subsequent steps in record form. The method in this embodiment can be executed on a single computing device or on a cluster of computing devices.

[0096] In a preferred embodiment of the present invention, obtaining numerical behavior features includes:

[0097] Load the deep packet inspection application flow details for the specified business day; use a hash function to map the gateway identifier to several parallel computing buckets; in each bucket, obtain the application identifier, destination IP, destination port, and flow count of each downstream device from the deep packet inspection application flow details; calculate the numerical behavioral characteristics of each downstream device based on the application identifier, destination IP, destination port, and flow count;

[0098] Among them, numerical behavioral characteristics include application-side behavioral characteristics, destination IP-side behavioral characteristics, and port-side behavioral characteristics;

[0099] Application-side behavioral characteristics include the highest value of application traffic share, the sum of traffic shares of the three applications with the highest traffic share, application distribution entropy, PCDN characteristic application share, and negative application share; negative application share includes the traffic share of live streaming application set, the traffic share of peer-to-peer download application set, and the traffic share of camera application set;

[0100] The behavioral characteristics of the destination IP side include the highest value of the destination IP traffic share, the sum of the traffic shares of the three destination IPs with the highest traffic share, the destination IP distribution entropy, and the destination IP concentration index;

[0101] Port-side behavioral characteristics include the highest percentage of destination port traffic, the percentage of traffic from live streaming port sets, and the percentage of traffic from general web port sets; application traffic percentage is the ratio of traffic from a single application to the total traffic of the downstream device; destination IP traffic percentage is the ratio of traffic from a single destination IP to the total traffic of the downstream device; and destination port traffic percentage is the ratio of traffic from a single destination port to the total traffic of the downstream device.

[0102] In this embodiment, the application flow details for deep packet inspection are stored in a distributed data storage system, partitioned by business day. The computing device loads all application flow records for that day from the distributed data storage system, using a specified business day as the filtering condition. Each application flow record corresponds to one application flow, and the record fields include gateway identifier, device identifier, application identifier, destination IP, destination port, and traffic byte count. The device identifier is used to distinguish different downstream devices under the same gateway; the application identifier refers to the application number output by the deep packet inspection system after identifying the application; the destination IP refers to the destination IP address of the application flow; and the destination port refers to the destination port number of the application flow. The flow count is obtained by aggregating and counting the application flow records by device identifier, representing the number of application flows generated by the downstream device on that day. The traffic refers to the cumulative value of the byte count recorded in the application flow details.

[0103] The computing device uses a hash function to map gateway identifiers to parallel computing buckets. The gateway identifier is input into the hash function to obtain a hash value. The hash value is then modulo the number of buckets to obtain the bucket number. The number of buckets is a configurable parameter, set during deployment based on the gateway size and the number of parallel computing units. The hash function can be the MD5 digest algorithm or other hash algorithms with uniform hashing properties. All application flow records from the same gateway carry the same gateway identifier and, after hash mapping, fall into the same bucket. All devices connected to the gateway complete feature calculations within a single bucket; device-level feature calculations do not cross buckets. Each bucket is processed in parallel by independent computing units, with no data dependency between buckets. The processing results from each bucket are written to non-overlapping records.

[0104] Within each bucket, the computing device groups application flow records by device identifier and aggregates the traffic bytes for each downstream device according to three dimensions: application identifier, destination IP, and destination port. The total traffic of the downstream device refers to the sum of the traffic bytes of all application flows on that device for the day. The application traffic percentage refers to the ratio of the traffic bytes of a single application to the total traffic of the downstream device; the destination IP traffic percentage refers to the ratio of the traffic bytes of a single destination IP to the total traffic of the downstream device; and the destination port traffic percentage refers to the ratio of the traffic bytes of a single destination port to the total traffic of the downstream device. All three percentages are calculated using the same set of total downstream device traffic as the denominator.

[0105] Application-side behavioral characteristics are calculated based on the above proportions. The traffic proportions of each application are sorted in descending order, with the application at the top having the highest traffic proportion. The traffic proportions of the top three applications are summed to obtain the sum of the traffic proportions of the three applications with the highest traffic proportions. If there are fewer than three application types on a given day, the traffic proportions of all existing applications are summed. The application distribution entropy uses normalized information entropy, calculated based on the traffic proportion of each application and then normalized according to the number of application types on that day. The application distribution entropy is 0 when the traffic of the connected device is concentrated in a single application on a given day; it is close to 1 when the traffic of each application is evenly distributed on a given day. An increase in application distribution entropy indicates a more uniform application traffic distribution, while a decrease indicates a more singular application traffic distribution. The PCDN feature application proportion is the sum of the traffic proportions of applications belonging to a preset PCDN feature application set. Set matching is based on the condition that the application identifiers are equal. The negative application share includes the traffic share of live streaming application set, the traffic share of peer-to-peer download application set, and the traffic share of camera application set. The three shares are the sum of the traffic share of applications belonging to the preset live streaming application set, the preset peer-to-peer download application set, and the preset camera application set, respectively. Each set is matched independently, and the content of the set is set during deployment.

[0106] In addition to calculating the traffic share of the three sets, the negative application share also records the application information with the highest traffic share within each set. For the preset live streaming application set, preset peer-to-peer download application set, and preset camera application set, the computing device selects the application with the highest traffic share within each set and records the application identifier, application name, and share of that application. The application name refers to the application name output by the deep packet inspection system. The correspondence between application name and application identifier is stored in the application identifier mapping table, which is maintained synchronously by the deployment party along with the deep packet inspection system, recording the correspondence between application identifier, application name, and application vendor. The share refers to the ratio of the number of bytes of traffic of the application to the total traffic of the downstream devices. The recorded application identifier, application name, and share of the three sets are written into the behavioral feature record for use in reading application evidence when generating non-PCDN interpretation tags with evidence.

[0107] The behavioral characteristics of the destination IP side are calculated based on the above proportions. The traffic proportions of each destination IP are sorted in descending order, with the highest proportion being the first. The sum of the top three destination IPs is obtained as the sum of their traffic proportions. If there are fewer than three types of destination IPs on a given day, the traffic proportions of all existing destination IPs are summed. The destination IP distribution entropy uses normalized information entropy, calculated based on the traffic proportion of each destination IP and then normalized according to the number of destination IP types on that day. When daily traffic is concentrated on a single destination IP, the destination IP distribution entropy is 0. When the traffic is evenly distributed among all destination IPs on that day, the destination IP distribution entropy is close to 1. An increase in destination IP distribution entropy indicates a more uniform distribution of destination IP traffic, while a decrease indicates a more singular distribution. The destination IP concentration index is defined using the Herfindahl-Hirschman Index. When a single destination IP carries all the traffic, the destination IP concentration index is 1. When the traffic is evenly distributed among all destination IPs, the destination IP concentration index decreases. An increase in the destination IP concentration index indicates a more concentrated distribution of destination IP traffic.

[0108] Based on the above proportions, calculate the port-side behavioral characteristics. Sort the traffic proportions of each destination port in descending order, with the first one being the highest-ranking destination port. The traffic proportion of the live streaming port set is the sum of the traffic proportions of destination ports belonging to the preset live streaming port set, and the traffic proportion of the general web port set is the sum of the traffic proportions of destination ports belonging to the preset general web port set. Port matching is based on equal port numbers. The port set content is set during deployment; the general web port set can include port number 80 and port number 443.

[0109] In addition to the three established port-side behavioral characteristics, a single web endpoint concentration score can be added. When the destination port corresponding to the highest percentage of destination port traffic belongs to a preset set of general web ports, the computing device takes the highest percentage of destination IP traffic as the single web endpoint concentration score. The single web endpoint concentration score characterizes the degree of concentration of services carried by the downstream device through a single web endpoint, and is used to identify fixed PCDN service peers disguised as web traffic. When the destination port corresponding to the highest percentage of destination port traffic does not belong to the preset set of general web ports, the single web endpoint concentration score is not taken, and the single web endpoint concentration score of this downstream device is not included in the daily sample. The single web endpoint concentration score is a port-side behavioral characteristic, belonging to the numerical behavioral characteristics other than entropy-based behavioral characteristics and negative application proportion-based behavioral characteristics, and directly uses the daily percentile as the participating score.

[0110] For example, suppose a preset general web port set includes port number 80 and port number 443. The highest destination port traffic percentage for a certain downstream device is 0.8, corresponding to destination port 443. 443 belongs to the preset general web port set. The highest destination IP traffic percentage for this device is 0.5, and the single web endpoint concentration of this device is set to 0.5. The highest destination port traffic percentage for another downstream device corresponds to destination port 8000. 8000 does not belong to the preset general web port set, so the single web endpoint concentration of this device is not taken, and the single web endpoint concentration of this device is not included in the daily sample.

[0111] Each bucket writes the calculated numerical behavioral characteristics into a behavioral characteristic record. The behavioral characteristic record is indexed by the device identifier and records the values ​​of the gateway identifier, device identifier, flow count, application-side behavioral characteristics, destination IP-side behavioral characteristics, and port-side behavioral characteristics. The flow count serves as a record field for the activity level of the downstream device, used for data quality assessment in subsequent stages. The behavioral characteristic record is used for daily adaptive weight calculation, daily behavioral score calculation, and subsequent judgment stages. Downstream devices without application flow records do not generate numerical behavioral characteristics and are not included in the daily overall sample. There is no overlap between records written by each bucket, and parallel execution of buckets will not result in concurrent writes to the same record. All behavioral characteristic records written by all parallel-computed buckets together constitute the complete set of behavioral characteristic records for all downstream devices on that day. Subsequent stages read the behavioral characteristic records by summarizing and reading records from all parallel-computed buckets.

[0112] The aggregation phase of the complete set of behavioral feature records is performed by binning and parallel aggregation according to the aggregation bin number. The aggregation bin number refers to the number of parallel computing bins used in the aggregation phase. The aggregation bin number and the number of bins used for behavioral feature extraction are two independent configurable parameters, both of which are set during deployment.

[0113] In a preferred embodiment of the present invention, obtaining the terminal device category, terminal manufacturer, and terminal type bias score includes: selecting the device identifier of the downstream device from the deep packet inspection application flow details, associating the device identifier with the terminal fingerprint database to obtain the terminal device category and terminal manufacturer; assigning the terminal type bias score according to the terminal device category, assigning a positive bias score to the terminal device category that tends to run PCDN services, assigning a negative bias score to the camera-type terminal device category, and assigning a zero bias score to the remaining terminal device categories.

[0114] In this embodiment, the device identifier field in the deep packet inspection application flow details records the identifier carried by the downstream device in the application flow. The device identifier carried by the same physical device in different application flows can exist in multiple forms. The computing device counts the occurrence frequency of all device identifiers for each downstream device on a given day, and selects the device identifier with the highest occurrence frequency as the representative device identifier. For device identifiers with the same number of occurrence frequencies, the device identifier that appeared earlier on the day is selected as the representative device identifier. The representative device identifier serves as the key for associating the downstream device with the terminal fingerprint database.

[0115] The terminal fingerprint database is constructed by the deployer based on the identification results of the manufacturer prefix in the device identifier, and is maintained by the deployer based on the device identification results of historical business days. The terminal fingerprint database uses the device identifier as an index, and each record registers the device identifier, terminal device category, terminal manufacturer, and main device category. All fields in the same record correspond to the same physical device.

[0116] The computing device searches the terminal fingerprint database using a representative device identifier as the key, employing an exact match. If a match is found, the terminal device category, terminal manufacturer, and host device category are read from the record. If a match is not found, the terminal device category is recorded as an unknown category, and the terminal manufacturer is recorded as an unknown manufacturer. The unknown category and unknown manufacturer are still written into the terminal fingerprint association record.

[0117] The bias configuration table is a preset parameter table, indexed by the terminal device category. Each record registers the terminal device category, bias type, and bias score. Bias types include positive bias, negative bias, and zero bias. The computing device queries the bias configuration table using the associated terminal device category as the key to retrieve the corresponding bias type and bias score. For downstream devices with an unknown category, the terminal type bias score is set to 0.

[0118] Positive bias scores are recorded in the records corresponding to terminal device categories that tend to run PCDN services. These categories can include routers and home network storage. Positive bias scores are positive and, when added to the daily behavior score, increase the daily behavior score. Negative bias scores are recorded in the records corresponding to camera-type terminal device categories, which can include home network cameras. Negative bias scores are negative and, when added to the daily behavior score, decrease the daily behavior score, thus suppressing score increases for devices in this category. Zero bias scores are recorded in the records corresponding to the remaining terminal device categories. Zero bias scores are 0 and, when added to the daily behavior score, keep the daily behavior score unchanged. The bias scores and the sum of the weighted participation scores are on the same dimension, and the specific value of the bias score is set by the deployer.

[0119] The associated terminal device category, terminal manufacturer, and terminal type bias score obtained from the bias query are written into the terminal fingerprint association record. The terminal fingerprint association record is used for the daily behavior score calculation and subsequent judgment process.

[0120] The terminal fingerprint association record adds two binary flag fields: a set-top box identifier and a camera identifier. The set-top box identifier is set to 1 if the downstream device's terminal device category is set-top box, and 0 for other terminal device categories. The camera identifier is set to 1 if the downstream device's terminal device category is camera, and 0 for other terminal device categories. After the terminal fingerprint association is completed, the computing device generates a set-top box identifier and a camera identifier for each downstream device based on the associated terminal device category. These two flags, along with the terminal device category, terminal manufacturer, main device category, and terminal type bias score, are written into the terminal fingerprint association record. The camera identifier can be read when filtering camera-type downstream devices in the comprehensive judgment, while the set-top box identifier can be read when verifying the main device category of live streaming devices that are not interpreted by PCDN. The judgment criteria for the flag fields are consistent with the comparison criteria for the terminal device category.

[0121] After the terminal fingerprint association is completed, the computing device aggregates the terminal fingerprint association fields and uplink traffic of each downstream device within the gateway according to the gateway identifier, forming the terminal fingerprint information of the downstream device. The terminal fingerprint information of the downstream device is indexed by the gateway identifier and registers the terminal fingerprint association fields, uplink traffic, and device identifier representing the terminal of each downstream device within the gateway. The terminal fingerprint information of the downstream device is used for subsequent judgment steps.

[0122] In a preferred embodiment of the present invention, calculating the adaptive weight for the day includes:

[0123] The calculation window is defined as a time window consisting of a preset number of historical backtracking days; the value sequence of each numerical behavioral feature within the calculation window is taken; the value sequence of each PCDN behavioral anchor point within the calculation window is taken; the daily adaptive weight is calculated using the following formula:

[0124] W i =|r(f i ,A1)|+|r(f i ,A2)|+|r(f i A3)|

[0125] w i =W i ÷ΣW i

[0126] In the formula, f i For the i-th numerical class behavioral feature, r(f) i A1) represents the numerical behavioral feature f. i The correlation coefficient between the value sequence of point A1 and the value sequence of anchor point A1, r(f i A2) represents the numerical behavioral feature f. i The correlation coefficient between the value sequence of point A1 and the value sequence of anchor point A2, r(f i A3) represents the numerical behavioral feature f. i The correlation coefficient W between the value sequence of point A3 and the value sequence of anchor point A3. i For numerical behavior features f i The original weights, w i For numerical behavior features f i The adaptive weights of the day, ΣW i It is the sum of the original weights of all numerical behavioral features; Anchor point A1 is the application ratio of PCDN features, anchor point A2 is the degree of application homogeneity, anchor point A3 is the highest value of the destination IP traffic ratio, the degree of application homogeneity is equal to 1 minus the application distribution entropy, and the application distribution entropy is the normalized information entropy of the application distribution.

[0127] In this embodiment, the preset historical backtracking days are positive integers set during deployment, and the calculation window covers the preset number of consecutive historical backtracking days prior to the current day. When the number of available historical business days prior to the current day is less than the preset historical backtracking days, the calculation window covers the actual existing historical business days, and the length of the value sequence is equal to the actual number of existing historical business days.

[0128] For each numerical behavioral feature, the window statistics for that feature on each business day within the calculation window are taken and arranged in chronological order to form a feature value sequence. The window statistics for historical business days are read from the complete set of behavioral feature records stored in the distributed data storage system for historical business days. These records are generated by executing the method of this embodiment on the historical business days. For each PCDN behavioral anchor point, the window statistics for that anchor point on each business day within the calculation window are taken and arranged in chronological order to form an anchor point value sequence. The anchor point value sequence is obtained by reading or transforming the window statistics for numerical behavioral features on historical business days. The value sequence for anchor point A1 is read from the window statistics of the PCDN feature application ratio, the value sequence for anchor point A3 is read from the window statistics of the highest value of the destination IP traffic ratio, and the value sequence for anchor point A2 is obtained by subtracting the application distribution entropy from the window statistics of 1. The statistical caliber of the window statistics follows the definition above, and the same statistical caliber is used for all business days within the calculation window. The feature value sequence and the anchor value sequence have the same length, which is equal to the number of business days within the calculation window.

[0129] PCDN behavior anchors include anchors A1, A2, and A3. Anchor A1 represents the proportion of PCDN characteristic applications, measuring the percentage of traffic handled by PCDN characteristic applications in the downstream device's service traffic. Anchor A2 represents the degree of application specialization, measuring the degree to which downstream device service traffic is concentrated on a single application. Anchor A3 represents the highest value of the destination IP traffic proportion, measuring the degree to which downstream device service traffic is concentrated on a single destination IP. The degree of application specialization equals 1 minus the application distribution entropy, which is the normalized information entropy of the application distribution. The value range of application distribution entropy is 0 to 1, and the value range of application specialization is also 0 to 1. As the application distribution entropy increases, the degree of application specialization decreases. Anchors A1 and A3 correspond to the highest values ​​of the proportion of PCDN characteristic applications and the highest value of the destination IP traffic proportion in the numerical behavior characteristics, respectively. Anchor A2 is obtained by transforming the application distribution entropy in the opposite direction. These three anchors describe the typical behavior patterns of PCDN services from three perspectives: the traffic handling proportion of PCDN characteristic applications, the degree of application specialization, and the degree of destination IP concentration.

[0130] The daily adaptive weight is calculated using the following formula:

[0131] W i =|r(f i ,A1)|+|r(f i ,A2)|+|r(f i A3)|

[0132] w i =W i ÷ΣW i

[0133] In the formula, fi For the i-th numerical class behavioral feature, r(f) i A1) represents the numerical behavioral feature f. i The correlation coefficient between the value sequence of point A1 and the value sequence of anchor point A1, r(f i A2) represents the numerical behavioral feature f. i The correlation coefficient between the value sequence of point A1 and the value sequence of anchor point A2, r(f i A3) represents the numerical behavioral feature f. i The correlation coefficient W between the value sequence of point A3 and the value sequence of anchor point A3. i For numerical behavior features f i The original weights, w i For numerical behavior features f i The adaptive weights of the day, ΣW i This is the sum of the original weights of all numerical behavioral features.

[0134] The correlation coefficient used is the Pearson correlation coefficient, which calculates the degree of linear correlation between the feature value sequence and the anchor value sequence. The reason for using the absolute value of the correlation coefficient is that there can be a positive or negative correlation between numerical behavioral features and PCDN behavioral anchors. Both positive and negative correlations reflect the strength of the association between the feature and the anchor. Taking the absolute value ensures that the strength of the association in different directions is consistently included in the original weights. For example, if the applied distribution entropy is negatively correlated with anchor A2, and anchor A2 is obtained by transforming the applied distribution entropy in the opposite direction, then as the applied distribution entropy increases, the value of anchor A2 decreases. Taking the absolute value, the strength of this negative correlation is consistent with that of a positive correlation. When the numerical behavioral feature and the PCDN behavioral anchor are derived from the same indicator, the feature value sequence and the anchor value sequence are identical, and the correlation coefficient is 1. When the value sequence remains unchanged within the calculation window, the variance of the value sequence is 0, and the Pearson correlation coefficient cannot be calculated; therefore, this correlation coefficient is 0.

[0135] The original weights reflect the overall correlation strength between numerical behavioral features and the three PCDN behavioral anchors within the calculation window. The larger the original weight, the stronger the correlation between the feature and typical PCDN business behaviors. Normalization is performed using the sum of all original weights as the base, resulting in the daily adaptive weight. The sum of all daily adaptive weights equals 1, and the value of a single daily adaptive weight ranges from 0 to 1. When the sum of all original weights is 0, the daily adaptive weights of all numerical behavioral features are equal. The daily adaptive weights correspond one-to-one with the numerical behavioral features, providing data for calculating the daily behavior score.

[0136] For example, suppose the numerical behavioral features used in the calculation on a given day are three items: PCDN feature application ratio, application distribution entropy, and target IP concentration index. The original weights of the three features are 0.9, 0.6, and 0.3, respectively, with a sum of 1.8. After normalization, the daily adaptive weight of the PCDN feature application ratio is 0.5, the daily adaptive weight of the application distribution entropy is 0.33, and the daily adaptive weight of the target IP concentration index is 0.17, with a sum of 1.

[0137] In a preferred embodiment of the present invention, the daily behavior score of the attached device is calculated using the following formula:

[0138] S=Σ[w i ×s(f i )]+b,

[0139] When numerical behavior characteristics f i When considering entropy-based behavioral features and negative application proportion-based behavioral features,

[0140] s(f i ) = 1 - p i ,

[0141] When numerical behavior characteristics f i When considering numerical behavioral features other than entropy-based behavioral features and negative application proportion-based behavioral features,

[0142] s(f i ) = p i ,

[0143] In the formula, S represents the daily behavioral score of the attached device, and w i For numerical behavior features f i The adaptive weights of the day, p i For numerical behavior features f i The daily quantile, s(f i ) represents the numerical behavioral feature f i The participation score is b, the terminal type bias score is b, and Σ[wi×s(fi)] is the sum of the weighted participation scores of all numerical behavioral features.

[0144] The daily quantile refers to the relative ranking position of numerical behavioral features within all samples on that day; all samples on that day refer to the numerical behavioral features of all connected devices on that day; entropy-based behavioral features include application distribution entropy and destination IP distribution entropy; negative application proportion behavioral features include the traffic proportion of live streaming application sets, the traffic proportion of peer-to-peer download application sets, and the traffic proportion of camera application sets.

[0145] In this embodiment, the daily quantile index value represents the relative ranking position of a behavioral feature within all samples on that day. For each numerical behavioral feature, the computing device sorts the feature values ​​in all samples on that day in ascending order, counts the number of samples whose values ​​are not higher than those of the downstream devices, and uses the ratio of this number to the total number of samples as the daily quantile for that feature of the downstream device. When multiple downstream devices have the same value for the feature, samples with the same value are counted in the same number of samples, and downstream devices with the same value obtain the same daily quantile. The daily quantile value is greater than 0 and does not exceed 1; a higher value indicates that the downstream device's relative position for that feature within all samples on that day is further back. The calculation of the daily quantile is based on all samples on that day. After all samples on that day are formed, the daily quantile is calculated for each feature of each downstream device.

[0146] The participation score is calculated from the daily percentile. The participation score for entropy-based behavioral features and negative application proportion-based behavioral features is equal to 1 minus the daily percentile. The participation score for other numerical behavioral features is equal to the daily percentile. Entropy-based behavioral features include application distribution entropy and destination IP distribution entropy. Negative application proportion-based behavioral features include the traffic proportion of live streaming applications, the traffic proportion of peer-to-peer download applications, and the traffic proportion of camera applications. An increase in the value of entropy-based behavioral features indicates a more uniform distribution of behavior. An increase in the value of negative application proportion-based behavioral features indicates an increase in the proportion of traffic carried by live streaming, peer-to-peer download, or camera-related services. As the values ​​of both types of features increase, the similarity between the behavior of downstream devices and PCDN services decreases. The participation score for these two types of features is obtained by subtracting the daily percentile from 1. Higher values ​​for these two types of features result in lower participation scores, and the direction of the participation score aligns with the similarity to PCDN services. In addition to the two categories mentioned above, numerical behavioral characteristics include the highest application traffic share, the sum of the traffic shares of the three applications with the highest traffic share, the PCDN characteristic application share, the highest destination IP traffic share, the sum of the traffic shares of the three destination IPs with the highest traffic share, the destination IP concentration index, the highest destination port traffic share, the traffic share of live streaming port sets, and the traffic share of general web port sets. For these characteristics, the daily percentile is used directly as the participating score. When adding a single web endpoint concentration score, the daily percentile is used directly as the participating score.

[0147] The daily behavior score of the attached device is calculated using the following formula:

[0148] S=Σ[w i ×s(f i )]+b,

[0149] When numerical behavior characteristics f i When considering entropy-based behavioral features and negative application proportion-based behavioral features,

[0150] s(f i ) = 1 - p i ,

[0151] When numerical behavior characteristics f i When considering numerical behavioral features other than entropy-based behavioral features and negative application proportion-based behavioral features,

[0152] s(f i ) = p i ,

[0153] In the formula, S represents the daily behavioral score of the attached device, and w i For numerical behavior features f i The adaptive weights of the day, p i For numerical behavior features f i The daily quantile, s(f i ) represents the numerical behavioral feature f i The participation score, b is the terminal type bias score, Σ[w i ×s(f i [)] is the sum of the weighted participation scores of all numerical behavioral features.

[0154] The summation operation covers all numerical behavioral features of the connected devices. Each numerical behavioral feature is calculated by multiplying its participation score by its daily adaptive weight. The sum of all daily adaptive weights equals 1. The participation score is neither less than 0 nor greater than 1. The weighted sum of participation scores is the weighted average of all participation scores, and the weighted sum of participation scores is neither less than 0 nor greater than 1. 'b' is the terminal type bias score, read from the terminal fingerprint association record and superimposed on the weighted sum of participation scores. When the terminal type bias score is positive, the daily behavioral score increases based on the weighted sum of participation scores; when the terminal type bias score is negative, the daily behavioral score decreases based on the weighted sum of participation scores; when the terminal type bias score is 0, the daily behavioral score is equal to the weighted sum of participation scores.

[0155] For example, suppose the total samples for a given day include 5 downstream devices, and the application percentages of PCDN features are 0.05, 0.12, 0.20, 0.35, and 0.48, respectively. For a downstream device with a value of 0.20, the number of samples with values ​​no higher than 0.20 is 3, the total sample size is 5, and the daily quantile is 0.6. Suppose the daily quantile of the device's application distribution entropy is 0.9. Application distribution entropy belongs to the entropy category of behavioral features, and its score is 1 minus 0.9, which equals 0.1. Suppose the daily adaptive weight of the PCDN feature application percentage is 0.5, the daily adaptive weight of the application distribution entropy is 0.3, the terminal type bias score of the device is 0.1, and other numerical behavioral features are not included in the example calculation. The daily behavioral score of the device is equal to 0.5 multiplied by 0.6 plus 0.3 multiplied by 0.1 plus 0.1, resulting in 0.43.

[0156] In this embodiment, the number of flows is used in data quality assessment. The computing device sets a preset minimum flow count threshold; devices with a daily flow count below this threshold are excluded from daily behavior score calculation. When the flow count is below the threshold, the behavior statistics for these devices are derived from a small number of application flows, leading to decreased reliability. Excluding these devices from the daily behavior score calculation prevents their scores from entering the judgment stage. The preset minimum flow count threshold is a positive integer parameter set during deployment. Based on this threshold, a one-to-one correspondence is established between the daily behavior score and the devices participating in the daily behavior score calculation; devices not participating in the calculation do not generate a daily behavior score.

[0157] In one optional implementation, the daily behavior score is calculated by summing the weighted participating scores and adding the terminal type bias score, then multiplying it by a score scaling factor. The score scaling factor is a positive parameter set during deployment. When the score scaling factor is 1, the daily behavior score retains its value before scaling; when the score scaling factor is any other positive number, the daily behavior score is scaled proportionally according to the scaling factor. The score scaling factor applies uniformly to the daily behavior scores of all connected devices on that day, and scaling does not change the relative order of the daily behavior scores.

[0158] The higher the daily behavior score, the closer the behavior of the connected device is to the PCDN business. The daily behavior score is used for subsequent judgment steps.

[0159] In a preferred embodiment of the present invention, generating a non-PCDN explanatory tag with evidence includes:

[0160] Calculate the overall percentile score for live streaming applications based on their traffic share; calculate the overall percentile score for peer-to-peer download applications based on their traffic share; calculate the overall percentile score for camera applications based on their traffic share; generate a non-PCDN explanation tag when the highest value among the three overall percentile scores reaches the rejection percentile threshold.

[0161] Candidate evidence is obtained, including application evidence, terminal evidence, and port evidence. Application evidence is the application identifier of the downstream device, terminal evidence is the terminal device category of the downstream device, and port evidence is the destination port of the downstream device. The main evidence is selected from the candidate evidence in the order of application evidence, terminal evidence, and port evidence, with the selection order being: first, application evidence is selected; if application evidence is missing, terminal evidence is selected; if terminal evidence is missing, port evidence is selected. The main evidence is associated with a non-PCDN interpretation tag to obtain a non-PCDN interpretation tag with evidence.

[0162] In this embodiment, the calculation of the three comprehensive quantile scores is based on the daily quantile of all samples on that day. The comprehensive quantile score for live streaming is calculated based on the traffic share of live streaming applications. The calculation device takes the daily quantile of the traffic share of live streaming applications within the entire sample on that day, and uses this daily quantile as the comprehensive quantile score for live streaming. This score is used to comprehensively measure the similarity between the behavior of the downstream device and the live streaming business model. The comprehensive quantile scores for peer-to-peer downloads and cameras are calculated based on the traffic share of peer-to-peer download applications and camera applications, respectively. The calculation methods for these two comprehensive quantile scores are the same as for the live streaming score. A higher comprehensive quantile score indicates a closer similarity between the behavior of the downstream device and the corresponding business model. The range of values ​​for the comprehensive quantile scores is consistent with the daily quantile.

[0163] The computing device compares the three composite percentile scores numerically and takes the highest value. If the highest value is not lower than the rejection percentile threshold, a non-PCDN explanatory label is generated for the downstream device. The label sub-type is "Live Streaming" if the highest value is the composite percentile score for live streaming, "Peer-to-Peer Download" if the highest value is the composite percentile score for peer-to-peer download, and "Camera" if the highest value is the composite percentile score for camera. If the three composite percentile scores are tied for the highest, the sub-type is selected in the order of "Live Streaming," "Peer-to-Peer Download," and "Camera." If the highest value is lower than the rejection percentile threshold, no non-PCDN explanatory label is generated for the downstream device. Generating a non-PCDN explanatory label does not change the daily behavior score.

[0164] The computing device acquires candidate evidence. Candidate evidence includes application evidence, terminal evidence, and port evidence. Application evidence refers to the application identifier of the connected device. The computing device takes the application identifier with the highest traffic share within the application set corresponding to the tag sub-type as application evidence. For example, if the tag sub-type is live streaming, it corresponds to the live streaming application set; if the sub-type is peer-to-peer download, it corresponds to the peer-to-peer download application set; and if the sub-type is camera, it corresponds to the camera application set. If there is no application stream record in the corresponding application set, the application evidence is missing. Terminal evidence refers to the terminal device category of the connected device. The computing device reads the terminal device category of the connected device from the terminal fingerprint association record as terminal evidence. If the terminal device category is unknown, the terminal evidence is missing. Port evidence refers to the destination port of the connected device. The computing device takes the destination port with the highest traffic share as port evidence. Reading port evidence requires that the connected device has application stream records, which is established when the tag is generated.

[0165] The computing device selects primary evidence from candidate evidence in the following order: application evidence, terminal evidence, and port evidence. The selection order is: first, application evidence; if application evidence is missing, terminal evidence; and if terminal evidence is missing, port evidence. The selected primary evidence is written to the evidence field of the non-PCDN interpretation tag. The tag record also registers the evidence category corresponding to the primary evidence, which can be one of application evidence, terminal evidence, or port evidence. After writing, the tag becomes a non-PCDN interpretation tag with evidence. The tag record also registers the proportion of primary evidence. When the primary evidence is application evidence, the proportion is the traffic proportion of that application; when the primary evidence is port evidence, the proportion is the traffic proportion of that destination port; when the primary evidence is terminal evidence, the proportion is not set. The calculation method for traffic proportion follows the granularity definition of the downstream device. The non-PCDN interpretation tag with evidence is written alongside the score record in the behavior score rolling summary result for subsequent judgment stages.

[0166] For example, let the veto percentile threshold be 0.7. A certain connected device has a live streaming category composite percentile score of 0.45, a peer-to-peer download category composite percentile score of 0.72, and a camera category composite percentile score of 0.30. The highest value is 0.72, reaching the veto percentile threshold. A non-PCDN explanation tag is generated for this connected device, with the tag subcategorized as peer-to-peer download. Let application A have the highest traffic share within the peer-to-peer download application set. Application A's application identifier and traffic share of 0.6 are written into the tag as primary evidence, categorized as application evidence, with a primary evidence share of 0.6. Before and after generating the non-PCDN explanation tag, the device's daily behavior score remains unchanged.

[0167] In a preferred embodiment of the present invention, obtaining the gateway PCDN prediction flag includes:

[0168] Load gateway WAN port traffic data and LAN port traffic data; extract the gateway daily traffic sequence from the gateway WAN port traffic data and LAN port traffic data;

[0169] The inter-day correlation coefficient, root mean square error, coefficient of variation, and burstiness of the daily traffic series of adjacent business days of the gateway are used as the extracted features.

[0170] The characteristics obtained from the deep packet inspection application flow details are: the proportion of home broadband access to the gateway, the application flow recognition rate of the gateway, the number of times and types of access to specific domain names of the gateway, the port traffic distribution of the gateway, and the traffic summary of the gateway's preset application set.

[0171] The extracted features and statistically obtained features are included in the model input information; the model input information and numerical behavioral features are input into the gradient boosting tree, and the gateway PCDN prediction flag is output.

[0172] In this embodiment, gateway WAN and LAN traffic data are stored in a distributed data storage system, partitioned by business day and recorded hourly. The gateway WAN and LAN traffic data are compiled from bandwidth details, which record the traffic values ​​of each gateway on the WAN and LAN ports at an hourly granularity. The hourly record fields include gateway identifier, business day, hour number, WAN uplink traffic, WAN downlink traffic, LAN uplink traffic, and LAN downlink traffic, with traffic measured in bytes. The computing device loads gateway WAN and LAN traffic data from the distributed data storage system using a specified business day and the preceding business day as filtering conditions.

[0173] The computing device extracts the daily traffic sequence of the gateways from the loaded data. The daily traffic sequence of a gateway refers to the hourly traffic sequence of a gateway within a single business day. The hourly span of the time grid is set by the number of observation hours, a configurable parameter set during deployment. The number of observation hours is set to 24 hours, and the time grid covers all hours of the business day. The hour numbers are arranged from 0:00 to the observation hour minus 1. The computing device constructs a time grid for online gateways on that day. Online gateways refer to gateways that report traffic on that day. For each gateway, the hourly values ​​of the WAN port uplink traffic are taken and arranged in chronological order to form the uplink daily traffic sequence; the hourly values ​​of the WAN port downlink traffic are taken and arranged in chronological order to form the downlink daily traffic sequence. LAN port traffic data is recorded with the same hourly granularity as WAN port traffic data. The hourly traffic values ​​of the LAN port are used for port-side data consistency comparison. Missing hour values ​​are filled with 0 to ensure that the sequence length is consistent with the number of observation hours and that the sequence length of adjacent business days is consistent. Adjacent business days refer to two business days that are immediately adjacent in time. The daily traffic sequence for the previous business day is formed in the same way.

[0174] The computing device uses the inter-day correlation coefficient of the daily traffic series of adjacent business days of the gateway, the root mean square error of the daily traffic series of adjacent business days of the gateway, the coefficient of variation of the daily traffic series of the gateway, and the burstiness of the daily traffic series of the gateway as the extracted features.

[0175] The inter-day correlation coefficient refers to the degree of linear correlation between the daily traffic sequences of a gateway on two adjacent business days, calculated using the Pearson correlation coefficient. The computing device takes the daily traffic sequence of the gateway for the current day and the daily traffic sequence of the previous business day, calculates the correlation coefficient between the two sequences, and uses this as the inter-day correlation coefficient for the gateway. The inter-day correlation coefficient is calculated separately for the uplink and downlink daily traffic sequences, yielding the uplink inter-day correlation coefficient and the downlink inter-day correlation coefficient. The value of the inter-day correlation coefficient ranges from -1 to 1. A value close to 1 indicates that the traffic patterns of adjacent business days are stable, while a value close to -1 indicates that the traffic patterns of adjacent business days are opposite.

[0176] The root mean square error (RMSE) is the square root of the mean of the squares of the hourly traffic differences between two adjacent daily traffic sequences of a gateway. The calculation device takes the values ​​for the same hour number from both sequences, calculates the difference, sums the squares of the differences over all hour numbers, divides by the number of hour numbers, and then takes the square root to obtain the RMSE. The RMSE is calculated separately for the uplink and downlink daily traffic sequences. RMSE and standard deviation are two different statistics. Standard deviation measures the dispersion of values ​​within a single sequence relative to the sequence mean, while RMSE measures the difference in hourly values ​​between two sequences.

[0177] The coefficient of variation (CV) is the ratio of the standard deviation to the mean of a gateway's daily traffic sequence. The calculation device takes all hourly traffic values ​​for the day, calculates the mean, and then takes the square root of the mean of the squares of the differences between each value and the mean to obtain the standard deviation. Dividing the standard deviation by the mean yields the CV. The CV is calculated separately for the uplink and downlink daily traffic sequences. The CV is 0 when the mean of the daily traffic sequence is 0. An increased CV indicates a greater degree of traffic fluctuation relative to the overall traffic level within the business day.

[0178] Burst rate is the ratio of the highest value to the average value of the daily traffic sequence at the gateway. The calculation device takes the highest value from all hourly traffic sequences for the day, divides it by the average value to obtain the burst rate. The burst rate is calculated separately for the uplink and downlink daily traffic sequences. The burst rate is 1 when the average of the daily traffic sequence is 0. An increased burst rate indicates that traffic is concentrated in a small number of hours within a business day. The burst rate of the gateway's daily traffic sequence is used for subsequent verification processes.

[0179] The computing device uses the deep packet inspection application flow details to statistically analyze the gateway's home broadband access ratio, gateway application flow identification rate, number and types of accesses to specific domain names of the gateway, gateway port traffic distribution, and gateway preset application set traffic summaries, and uses the statistical results as statistical features.

[0180] Home broadband access percentage refers to the proportion of all application flows from the gateway whose destination IP belongs to the home broadband network segment. The home broadband network segment is defined by a network segment list maintained by the deployer. The computing device retrieves all application flow records from the gateway for the day, counts the number of flows whose destination IP belongs to the home broadband network segment list, and calculates the ratio of this ratio to the total number of flows from the gateway for the day to obtain the home broadband access percentage. Under one statistical metric, the home broadband access percentage is calculated based on the number of unique IPs, taking the ratio of the number of unique IPs whose destination IP belongs to the home broadband network segment list to the total number of unique IPs from the gateway for the day. One of the two statistical metric methods is selected during deployment and remains consistent across all business days.

[0181] Application flow identification rate refers to the ratio of the number of application flows carrying application identifiers on a given day to the total number of application flows on that day. The deep packet inspection system outputs application identifiers for identifiable application flows and does not output application identifiers for unidentifiable application flows. The higher the proportion of application flows carrying application identifiers among all application flows, the higher the application flow identification rate.

[0182] The number of visits and types of specific domain names are statistically analyzed from the domain name information corresponding to the application flow. The deep packet inspection system obtains the domain name information corresponding to the application flow during protocol parsing, and this information is stored along with the application flow records. The specific domain name set includes large website domains, IP query domains, and speed test domains; the set is maintained by the deployment party. The computing device retrieves records from all application flow records of the gateway for the day whose domain name information belongs to the specific domain name set, counts the number of records to obtain the number of visits to the specific domain name, and after deduplicating the domain name information, counts the number of access types to obtain the number of accesses to the specific domain name.

[0183] Port traffic distribution refers to the proportion of traffic in each destination port range to the total gateway traffic for a given day. The total gateway traffic is the sum of the number of bytes in traffic from all application flows on that day. The port ranges are divided into three ranges: port numbers 0 to 1023 (well-known ports); port numbers 1024 to 49151 (registered ports); and port numbers 49152 to 65535 (dynamic ports). The computing device categorizes all application flows on the gateway for the day into these three ranges based on their destination ports and calculates the ratio of the sum of the number of bytes in traffic from each range to the total gateway traffic. It also calculates the ratios of the sum of the number of bytes in traffic from application flows destined for port 80 and application flows destined for port 53 to the total gateway traffic, thus forming the port traffic distribution.

[0184] The preset application set traffic summary refers to the proportion of application traffic corresponding to each preset application set to the total gateway traffic. The preset application sets include PCDN characteristic application sets, live streaming application sets, peer-to-peer download application sets, and camera application sets. The content of these sets is consistent with the sets used for calculating the behavioral characteristics of downstream devices. The computing device categorizes all application flows of the gateway for the day into each set according to application identifier, and calculates the ratio of the sum of the application traffic bytes in each set to the total gateway traffic to obtain the preset application set traffic summary.

[0185] The computing device incorporates the extracted and statistically derived features into the model input information. The model input information consists of field groups; the extracted and statistically derived features are each written as fields into the model input information. The model input information also includes categorical features composed of the terminal manufacturer and terminal device category represented by the gateway. The computing device extracts the numerical behavioral features of the gateway-represented terminal and merges the model input information with these numerical behavioral features into a gateway-level input feature vector. This input feature vector is then fed into a gradient boosting tree. The gradient boosting tree predicts the input feature vector, outputting a gateway PCDN prediction flag and a prediction probability. The gateway PCDN prediction flag indicates that the model predicts the gateway belongs to a PCDN service user, for subsequent verification and comprehensive judgment.

[0186] In a preferred embodiment of the present invention, clearing the gateway PCDN prediction flag that fails strong rule verification includes:

[0187] For gateways carrying the PCDN prediction flag, the pass conditions are verified sequentially. The pass conditions are: the gateway's uplink traffic on the current day is higher than the preset uplink threshold for the current day; the gateway's average uplink traffic over the past few days is higher than the preset average uplink threshold; the gateway is not on the list of gateways with sudden characteristics; the gateway's home broadband access ratio is higher than the preset home broadband access ratio threshold; the PCDN prediction flag is retained for gateways that meet all the pass conditions; the PCDN prediction flag is cleared for gateways that do not meet the pass conditions.

[0188] In this embodiment, the strong rule verification object is the gateway carrying the gateway PCDN prediction flag. The verification is performed after the gradient boosting tree outputs the gateway PCDN prediction flag and before the comprehensive judgment.

[0189] The passing conditions include four items. The first condition is that the gateway's uplink traffic on that day exceeds the preset uplink threshold for that day. The gateway's uplink traffic on that day refers to the sum of the hourly uplink traffic values ​​of the gateway's WAN port on that day. This is calculated by summing all values ​​of the gateway's daily uplink traffic sequence. The preset uplink threshold is a parameter set during deployment, expressed in bytes. The second condition is that the gateway's average uplink traffic over the past few days exceeds the preset average uplink threshold. The past few days refer to a positive integer number of business days set during deployment. The average uplink traffic is the arithmetic mean of the daily uplink traffic of the gateway over the previous few business days. Daily uplink traffic refers to the sum of the hourly uplink traffic values ​​of the gateway's WAN port on a single business day. The business days covered by the second condition do not include the current day. The business day ranges covered by the first and second conditions do not overlap. The uplink traffic on the current day is verified separately by the first condition. The preset average uplink threshold is a parameter set during deployment, expressed in bytes. The third condition is that the gateway is not included in the list of gateways exhibiting sudden bursts of activity on that day. The fourth condition is that the percentage of gateway home broadband accesses is higher than the preset home broadband percentage threshold. The home broadband access percentage follows the gateway-level statistical caliber. The preset home broadband percentage threshold is a parameter set during deployment, with a value greater than 0 and less than 1.

[0190] The list of gateways exhibiting burst characteristics is generated by the deployer during the preprocessing stage and updated daily. The generation method involves calculating the number of business days for each gateway whose uplink daily traffic sequence burst rate reaches a preset burst rate threshold within a preset statistical window. Gateways with a burst rate not less than the minimum number of occurrence days are added to the list of gateways exhibiting burst characteristics. The burst rate calculation uses the same definition as the burst rate of the gateway's daily traffic sequence. The preset statistical window, preset burst rate threshold, and minimum number of occurrence days are all parameters set during deployment.

[0191] The computing device sequentially verifies the pass conditions for gateways carrying the gateway PCDN prediction flag. Verification is performed in the order of the first, second, third, and fourth conditions; if a previous condition fails, verification of subsequent conditions is stopped. Gateways that meet all four pass conditions retain their gateway PCDN prediction flag. Gateways with failed pass conditions have their gateway PCDN prediction flag cleared. The clearing action is performed by updating the flag field status, setting the gateway PCDN prediction flag to invalid. The reason for clearing can be recorded, using the name of the failed pass condition, for manual review.

[0192] For example, suppose the preset daily uplink threshold is 500,000,000,000 bytes, the preset average uplink threshold is 300,000,000,000 bytes, and the preset home broadband access ratio threshold is 0.5. Gateway A's daily uplink traffic is 600,000,000,000 bytes, exceeding the preset daily uplink threshold; Gateway A's average uplink traffic over the past few days is 400,000,000,000 bytes, exceeding the preset average uplink threshold; Gateway A is not on the daily list of gateways with sudden bursts of traffic; Gateway A's home broadband access ratio is 0.6, exceeding the preset home broadband access ratio threshold. Gateway A passes all four conditions and retains the gateway PCDN prediction flag it carries. Gateway B's home broadband access ratio is 0.3, below the preset home broadband access ratio threshold, and fails the fourth condition. Therefore, Gateway B's gateway PCDN prediction flag is cleared, and the reason for clearing is recorded as home broadband access ratio.

[0193] In a preferred embodiment of the present invention, clearing non-PCDN explanation tags for live streaming includes:

[0194] Non-PCDN explanation tags for live streaming that meet the misidentification filtering criteria based on the proportion of home broadband access and the application stream recognition rate will be removed.

[0195] In this embodiment, the misidentification filtering condition is applied to the live streaming non-PCDN interpretation tags retained after the main device category verification. The main device category verification refers to the aforementioned steps of comparing the main device category with the preset exemption category set and clearing the live streaming non-PCDN interpretation tags of the downstream devices whose main device category is outside the preset exemption category set. The clearing action and the clearing action of the main device category verification are executed sequentially in the same step. Both clearing paths are implemented by updating the tag validity status. The clearing operation is completed before the comprehensive judgment reads the tag.

[0196] The false positive filtering condition consists of two sub-conditions. The first sub-condition is that the gateway's home broadband access ratio reaches a preset home broadband ratio filtering threshold. The second sub-condition is that the gateway's application flow recognition rate is lower than a preset application flow recognition rate filtering threshold. The false positive filtering condition is established when both sub-conditions are met. Both the preset home broadband ratio filtering threshold and the preset application flow recognition rate filtering threshold are parameters set during deployment, and their values ​​are both greater than 0 and less than 1. The preset home broadband ratio filtering threshold and the preset home broadband ratio threshold used for strong rule validation are two different parameters and are set independently.

[0197] Both home broadband access ratio and application flow recognition rate are calculated using gateways as the statistical unit, and the statistical criteria follow the definition above. For each live streaming non-PCDN interpretation tag that is still retained after passing the main device category verification, the home broadband access ratio and application flow recognition rate of the gateway where the device reads the tag is located are calculated, and the two values ​​are compared with the two sub-conditions respectively.

[0198] Application flow identification rate reflects the completeness of deep packet inspection's parsing of gateway services. When the application flow identification rate is lower than the preset application flow identification rate filtering threshold, deep packet inspection only outputs application identifiers for some application flows. The application identifier records on which the statistics of the proportion of traffic from live streaming applications rely are incomplete, and the reliability of the judgment basis for non-PCDN interpretation tags for live streaming is reduced.

[0199] The home broadband access ratio reflects the proportion of home broadband network segments in the gateway's access destination addresses. When the home broadband access ratio reaches the preset home broadband access ratio filtering threshold, the gateway's access peers are concentrated in the home broadband network segments. The access pattern is consistent with the peer characteristics of PCDN services distributing content to home broadband users, but inconsistent with the peer characteristics of live streaming services accessing content platforms.

[0200] When both sub-conditions are met, the live streaming non-PCDN interpretation tag is derived from incomplete application data, and the gateway's traffic pattern is similar to that of PCDN services. The tag is highly likely to be misidentified, and the computing device will clear the corresponding live streaming non-PCDN interpretation tag. The clearing action updates the tag record to an invalid state, and the cleared tag will not be read in the comprehensive judgment.

[0201] In a preferred embodiment of the present invention, the gateway PCDN service identification result is obtained by combining the non-PCDN interpretation tag and PCDN prediction flag of the downstream device with the largest uplink traffic within the integrated gateway, including:

[0202] Based on the uplink traffic, select the downstream device with the largest uplink traffic within the gateway;

[0203] Calculate the proportion of uplink traffic of the gateway to the total uplink traffic of the connected devices (cameras) within the gateway.

[0204] Gateways whose uplink traffic accounts for a preset threshold of their total uplink traffic are classified as high-traffic non-PCDN users. Gateways whose downstream device terminal manufacturers belong to a preset set of camera-determined manufacturers are also classified as high-traffic non-PCDN users. The gateway with the evidence-based non-PCDN interpretation tag of the downstream device with the largest uplink traffic is read. Gateways carrying the non-PCDN interpretation tag are classified as high-traffic non-PCDN users. Gateways that are not high-traffic non-PCDN users but carry the PCDN prediction flag are classified as PCDN users. Gateways that are not high-traffic non-PCDN users and do not carry the PCDN prediction flag are classified as normal users. The gateway PCDN service identification result is obtained.

[0205] In this embodiment, the uplink traffic of a downstream device refers to the sum of the number of uplink traffic bytes in the application flow of that device on that day. The number of traffic bytes in the application flow record is accumulated separately for the uplink and downlink directions, and the accumulated value for the uplink direction constitutes the data source of the uplink traffic of the downstream device. When parsing the application flow, the deep packet inspection system identifies the uplink and downlink directions according to the packet transmission direction, and the number of traffic bytes in each direction is stored with the application flow record. The computing device summarizes the sum of the uplink traffic bytes of each device in the application flow details for all downstream devices in the gateway to obtain the uplink traffic of each downstream device. The computing device compares the uplink traffic of all downstream devices in the gateway by value, and selects the downstream device with the largest uplink traffic as the downstream device with the largest uplink traffic in the gateway, that is, the gateway represents the terminal. When the uplink traffic of multiple downstream devices is tied for the largest, the downstream device with the device identifier first in lexicographical order is selected, and the selection result is unique.

[0206] The computing device reads the terminal device categories of all downstream devices within the gateway from the terminal fingerprint association record, filters downstream devices whose terminal device category is camera, and sums the uplink traffic of the filtered downstream devices to obtain the total uplink traffic of camera-type downstream devices. The ratio of the total uplink traffic of camera-type downstream devices to the gateway's uplink traffic is then calculated to obtain the proportion of uplink traffic to the gateway's total uplink traffic. The gateway's uplink traffic follows the previous definition, taking the sum of the hourly uplink traffic values ​​of the gateway's WAN port on that day, consistent with the statistical caliber of the first condition for passing the strong rule validation.

[0207] In the comprehensive assessment, a gateway is classified as a high-traffic non-PCDN user if it meets any of the following three classification paths.

[0208] The first classification path is the proportion threshold path. The preset threshold refers to the threshold parameter set during deployment for the proportion of uplink traffic from cameras, with a value greater than 0 and less than 1. When the proportion of uplink traffic to gateway uplink traffic reaches the preset threshold, the gateway is classified as a high-traffic non-PCDN user, meaning the proportion is not lower than the preset threshold.

[0209] The second classification path is the vendor set path. The preset camera vendor set refers to the set of terminal manufacturers mainly producing camera-type terminals maintained by the deployer. The contents of the set are set during deployment. The computing device reads the terminal manufacturer represented by the gateway from the fingerprint information of the terminal devices connected to the gateway and compares it item by item with the preset camera vendor set. If the terminal manufacturer belongs to the preset camera vendor set, the gateway classifies it as a high-traffic non-PCDN user.

[0210] The third classification path is the tag path. The computing device reads the non-PCDN explanation tag with evidence from the behavior score rolling summary result of the downstream device with the largest uplink traffic within the gateway. When the tag is in a valid state, the downstream device carries a non-PCDN explanation tag, and the gateway classifies it as a high-traffic non-PCDN user. Tags whose tags have been updated to an invalid state by the clearing action are not included in the reading.

[0211] If a gateway does not meet any of the three classification paths but carries the gateway PCDN prediction flag, it is classified as a PCDN user. A valid gateway PCDN prediction flag is considered as carrying the flag. If a gateway does not meet any of the three classification paths and does not carry the gateway PCDN prediction flag, it is classified as a normal user. The three classification paths are executed in the following order: proportional threshold path, vendor set path, and tag path. Once a gateway meets any of these paths, it will not proceed to the next classification path or category determination.

[0212] For example, let the preset threshold be 0.6. The ratio of the total uplink traffic of the camera-type downstream devices of Gateway C to the gateway's uplink traffic is 0.7, reaching the preset threshold. Gateway C meets the first classification path and is classified as a high-traffic non-PCDN user. Gateway D does not meet all three classification paths, carries the gateway PCDN prediction flag, and is classified as a PCDN user. Gateway E does not meet all three classification paths, does not carry the gateway PCDN prediction flag, and is classified as a normal user.

[0213] The gateway PCDN service identification result is output in the form of a record, which records the gateway identifier and service category. The service category is one of three types: high-traffic non-PCDN users, PCDN users, and normal users. The gateway PCDN service identification result is used by the operator's broadband gateway for service classification and management.

[0214] High-traffic non-PCDN users refer to user categories whose uplink traffic reaches a high level and is determined not to belong to PCDN services. When camera-type downstream devices handle high-traffic uplinks, the service model is mainly camera backhaul, which differs from the content distribution model of PCDN services. Classifying the corresponding gateway as a high-traffic non-PCDN user can avoid mistakenly classifying camera backhaul services as PCDN services.

[0215] In one or more embodiments of the present invention, before the computing device performs the actions of obtaining the terminal device category, terminal manufacturer, terminal type bias score, and numerical behavioral characteristics of the downstream devices, it first performs a preprocessing stage to complete the gateway dimension information. The computing device constructs a unified gateway dimension table according to a multi-source merging strategy. The multi-source merging strategy refers to taking the values ​​of the gateway's broadband account, province / city / district, gateway manufacturer, and gateway model fields in the order of user-imported data, interface query data, and gateway registration data. If the corresponding field value exists in the preceding data source, the preceding data source value is used; if the corresponding field value is missing in the preceding data source, the subsequent data source value is used; if all data sources are missing, the field is left blank. The unified gateway dimension table uses the gateway identifier as an index and records the gateway identifier, broadband account, province / city / district, gateway manufacturer, and gateway model. The unified gateway dimension table is used for daily user profile table output and reading.

[0216] In one or more embodiments of the present invention, after loading the application flow details for deep packet inspection in the preprocessing stage, the uplink traffic of each downstream device and its proportion of the gateway's total uplink traffic are statistically analyzed. The downstream device's uplink traffic refers to the sum of the number of bytes of uplink traffic in the device's application flow on that day, and the statistical caliber follows the definition above. The proportion of the gateway's total uplink traffic refers to the ratio of the downstream device's uplink traffic to the gateway's uplink traffic on that day. The gateway's uplink traffic on that day is the sum of the hourly values ​​of the gateway's WAN port uplink traffic on that day. The statistical results are written to the device traffic record, which is indexed by the device identifier and records the gateway identifier, device identifier, downstream device uplink traffic, and the proportion of the gateway's total uplink traffic, for subsequent judgment stages to read.

[0217] In one or more embodiments of the present invention, the preprocessing stage identifies a list of abnormal gateways. The computing device retrieves the uplink traffic of each gateway for the day. The uplink traffic of a gateway for the day refers to the sum of the hourly uplink traffic values ​​of the gateway's WAN port for the day, and the statistical caliber is consistent with the statistical caliber of the first condition for passing the strong rule verification. Gateways whose uplink traffic for the day exceeds a preset abnormal traffic threshold are added to the abnormal gateway list. The abnormal gateway list is updated daily, and the preset abnormal traffic threshold is a parameter set during deployment, measured in bytes. The abnormal gateway list and the list of gateways with sudden characteristics are two independent lists. The list of gateways with sudden characteristics is identified based on the degree of suddenness and the minimum number of days it occurs, while the abnormal gateway list is identified based on the uplink traffic threshold for the day. The abnormal gateway list is used for classifying abnormal gateway users during multi-layer type conversion.

[0218] In one or more embodiments of the present invention, the gateway PCDN service identification result output by the comprehensive judgment is processed in the time-series buffer pool. The daily identification result refers to the service category corresponding to the gateway PCDN service identification result output by the comprehensive judgment. The computing device maintains a judgment sequence for each gateway on a daily basis, and the judgment sequence records the service category of the gateway on each daily business day. The observation window refers to the number of consecutive preset observation days prior to the current day. The total number of days refers to the number of days within the observation window where the service category is PCDN user. The maximum consecutive days refers to the longest consecutive number of days within the observation window where the service category is PCDN user. The computing device calculates the total number of days and the maximum consecutive days within the observation window. If the total number of days is not lower than the preset total number of days threshold or the maximum consecutive days is not lower than the preset consecutive days threshold, the gateway passes through the time-series buffer pool. The preset observation days, preset total number of days threshold, and preset consecutive days threshold are all positive integer parameters set during deployment. For gateways whose daily identification result is a normal user and which pass through the time-series buffer pool, the service category is corrected to PCDN user. Gateways using a time-series buffer pool retain the PCDN (Public Transaction Domain Name) determination result for a preset number of business days after the determination date. During the retention period, if the gateway's daily identification result is a normal user, the service category remains PCDN user. Gateways with a service category of high-traffic non-PCDN user retain their service category. Gateways with a service category of high-traffic non-PCDN user are downgraded to normal user if their daily uplink traffic falls below a preset high-traffic downgrade threshold. The statistical caliber of daily uplink traffic follows the definition above. The preset retention period is a positive integer parameter set during deployment, and the preset high-traffic downgrade threshold is a parameter set during deployment, in bytes.

[0219] For example, suppose the preset observation period is 7 days, the preset total days threshold is 4 days, and the preset consecutive days threshold is 3 days. If the gateway has observed that the service days for which the service category is PCDN user within the observation window are days 1, 2, 4, and 6, totaling 4 days, the preset total days threshold has been reached, and the gateway has passed the time-series buffer pool. The gateway's identification result for that day is a normal user, and the service category is corrected to PCDN user. If the gateway G has observed that the service days for which the service category is PCDN user within the observation window are days 2 and 5, totaling 2 days, and the maximum consecutive days is 1, the preset total days threshold and the preset consecutive days threshold have not been reached. Therefore, the gateway G has not passed the time-series buffer pool, and the gateway G's identification result for that day is a normal user, and the service category remains normal user.

[0220] In one or more embodiments of the present invention, the service category processed by the time-series buffer pool enters a multi-level type conversion. The multi-level type conversion is performed in the following order: pending verification, high-volume upgrade, blocking exclusion, abnormal gateway, and offline inheritance. If a gateway is involved in a preceding conversion, subsequent conversions are not performed. Pending verification conversion means that when the gateway is on the pending verification list, the service category is classified as a pending verification user. The pending verification list is maintained by the deployer and updated according to operational verification needs. High-volume upgrade conversion means that when the gateway was previously classified as a high-volume non-PCDN user in the observation window, and the identification result for the day is a PCDN user, the service category is upgraded to a high-volume non-PCDN user. Blocking exclusion conversion means that when the gateway has a blocking operation record, the service category is classified as a blocked user; when the gateway has an exclusion operation record, the service category is classified as an excluded user; when the gateway has a cancellation operation record, the service category is classified as a cancelled user. The operation records are returned by the operator's network management system. Abnormal gateway conversion means that when the gateway is on the abnormal gateway list for the day, the service category is classified as an abnormal gateway user. Offline inheritance conversion means that when the gateway does not report data on the current day, the business category inherits the business category of the previous business day and is marked as offline. No data reported means that the gateway has no traffic data and application flow records reported on the current day.

[0221] In one or more embodiments of the present invention, after multi-level type conversion, an identification conclusion and confidence level are generated. The conclusion text generates a natural language description according to the business category, recording the business category name and business day. The confidence level is the predicted probability output by the gradient boosting tree; if the gateway does not enter the gradient boosting tree prediction, the confidence level is left blank. The PCDN application vendor is the name of the application vendor with the most rule hits in the past few days, using the definition from above. The application vendor refers to the vendor name to which the application identifier belongs, read from the application identifier mapping table. Rule hit means the application identifier belongs to a preset PCDN feature application set. The computing device retrieves the application flow records of the gateway for the past few business days before the current day, counts the rule hits for each application vendor, and takes the application vendor name with the most hits as the PCDN application vendor; if the hits are tied, the name of the application vendor name ranked first in lexicographical order is taken. The main terminal category is the terminal device category represented by the gateway, read from the terminal fingerprint association record.

[0222] In one or more embodiments of the present invention, the identification conclusions and confidence levels are aggregated and output as a strategy. The computing device generates a daily user profile table, indexed by the gateway identifier. Each record registers the gateway identifier, broadband account, province / city / district, service category, conclusion text, confidence level, PCDN application vendor, main terminal category, initial identification time, monitoring start and end dates, number of active days, and uplink traffic statistics for the past few days. The broadband account and province / city / district are retrieved from a unified gateway dimension table. The initial identification time refers to the business day on which the gateway's service category is first classified as a PCDN user or a high-traffic non-PCDN user. The monitoring start and end dates refer to the start and end dates of the monitoring period. The number of active days refers to the number of business days on which the gateway reports data within the monitoring period. The uplink traffic statistics for the past few days refer to the total daily uplink traffic and average uplink traffic of the gateway over the previous several consecutive business days. The average uplink traffic refers to the arithmetic mean of the daily uplink traffic, and the definition for the past few days is the same as above. The computing device generates a periodic monitoring table, which summarizes the daily user profile table content according to the monitoring period, which is the business day span set during deployment. When a business category changes, the computing device writes a type change log, which records the gateway identifier, the business category before the change, the business category after the change, and the business day of the change. For blocked users, the computing device calculates the average daily uplink traffic before and after the blocking. The average daily uplink traffic before the blocking refers to the arithmetic average of the daily uplink traffic for the most recent several business days before the business day corresponding to the blocking operation record, and the average daily uplink traffic after the blocking refers to the arithmetic average of the daily uplink traffic for the most recent several business days after the business day corresponding to the blocking operation record. The traffic drop rate is equal to the difference between the average daily uplink traffic before and after the blocking, divided by the average daily uplink traffic before the blocking. If the average daily uplink traffic before the blocking is 0, the traffic drop rate is not calculated. Traffic drop rate is written to the blocking effect table, which records the gateway identifier, the service day of the blocking operation, the average daily uplink traffic before blocking, the average daily uplink traffic after blocking, and the traffic drop rate. The computing device outputs a PCDN user list and a high-traffic non-PCDN user list. The PCDN user list summarizes gateway entries whose service category is PCDN user, and the high-traffic non-PCDN user list summarizes gateway entries whose service category is high-traffic non-PCDN user. The computing device generates blocking suggestions, which are disposal recommendations generated for gateways whose service category is PCDN user and have no blocking operation records. The PCDN user list, the high-traffic non-PCDN user list, and the blocking suggestions are available for use by the operator's network management system.

[0223] This invention uses gateway-connected devices as the discrimination unit, incorporating terminal device category and terminal manufacturer into the judgment chain in two ways: the terminal type bias score is superimposed on the daily behavior score, and the terminal manufacturer and terminal device category constitute categorical features that enter the model input information of the gradient boosting tree model. Thus, information characterizing device attributes and data characterizing network behavior work together in the same judgment process, allowing the behavioral differences undertaken by different categories of devices within the same gateway to be included in the discrimination process, thereby constraining the sources of misjudgment at the device category level.

[0224] In this invention, exclusion actions for non-PCDN services are expressed in the form of non-PCDN explanation tags with evidence, while maintaining the daily behavior score when generating tags. The daily adaptive weight is calculated by the correlation coefficient and normalization between numerical behavior features and each PCDN behavior anchor point, depending on the sample set participating in the statistical calculation. Exclusion actions are expressed in the form of tags without changing the sample set participating in the statistical calculation. The statistical caliber and the handling caliber are independent of each other, and the strategy actions no longer affect the judgment criteria of subsequent business days through sample distribution, thus avoiding the problem of the judgment criteria accumulating and shifting with the handling actions. At the same time, the explanation tags carry the judgment criteria, and the exclusion conclusions are traceable and verifiable. For downstream devices whose main device category is outside the preset exemption category set, the live broadcast non-PCDN explanation tags are cleared, and the exclusion conclusions can be corrected when conditions change. The false negatives caused by the exclusion actions do not accumulate with the business days.

[0225] The daily adaptive weights for each numerical behavioral feature are obtained by normalizing the sum of the absolute values ​​of the correlation coefficients between the feature and each PCDN behavioral anchor point, and the weights are adjusted according to the sample distribution. Compared with the fixed weight method, the behavioral score can adjust with changes in the sample distribution of region and time period, maintaining the effectiveness of the judgment criteria when the business distribution drifts.

[0226] Based on the above mechanism, the gradient boosting tree model uses numerical behavioral characteristics and model input information to provide a gateway PCDN prediction flag. Gateway PCDN prediction flags that fail strong rule verification are removed. Finally, the non-PCDN interpretation label of the downstream device with the highest uplink traffic within the gateway is combined with the PCDN prediction flag to obtain the gateway PCDN service identification result. Each step is interconnected with traceable evidence, and the identification process is interpretable, providing a basis for PCDN service identification and classification on the operator's broadband gateway side.

[0227] The above provides a detailed description of the PCDN service identification method. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, those skilled in the art will recognize that there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A PCDN service identification method, characterized in that, The method includes: Obtain the terminal device category, terminal manufacturer, terminal type bias score, and numerical behavioral characteristics of the connected devices to the gateway; The original weights are obtained by summing the absolute values ​​of the correlation coefficients between numerical behavioral features and each PCDN behavioral anchor point. Each of the original weights is then normalized to obtain the daily adaptive weights corresponding to each numerical behavioral feature. The participation score of each numerical behavior feature is weighted and summed with the corresponding adaptive weight for the day, and then the terminal type bias score is added to obtain the behavior score of the downstream device for the day. Based on the numerical behavioral characteristics, calculate the non-PCDN comprehensive percentile score of the connected device, and generate a non-PCDN explanation label with evidence for the connected device whose comprehensive percentile score reaches the rejection percentile threshold while maintaining the behavioral score of the day. The numerical behavioral features and model input information are input into a gradient boosting tree to obtain the gateway PCDN prediction label. The model input information includes categorical features composed of the terminal manufacturer and the terminal device category. Clear the PCDN prediction flag of the gateway that fails strong rule verification; Read the main device category from the terminal device category, and for the downstream devices whose main device category is outside the preset exemption category set, remove the live broadcast category non-PCDN interpretation tags; The non-PCDN interpretation label and PCDN prediction flag of the downstream device with the largest uplink traffic in the integrated gateway are used to obtain the gateway's PCDN service identification result.

2. The method according to claim 1, characterized in that, Obtaining the numerical behavioral features includes: Load the deep packet inspection application flow details for a specified business day; use a hash function to map the gateway identifier to several parallel computing buckets; in each bucket, obtain the application identifier, destination IP, destination port, and flow count of each downstream device from the deep packet inspection application flow details; calculate the numerical behavior characteristics of each downstream device based on the application identifier, destination IP, destination port, and flow count; The numerical behavioral characteristics include application-side behavioral characteristics, destination IP-side behavioral characteristics, and port-side behavioral characteristics. The application-side behavioral characteristics include the highest value of application traffic share, the sum of the traffic shares of the three applications with the highest traffic share, application distribution entropy, PCDN characteristic application share, and negative application share; the negative application share includes the traffic share of live streaming application set, the traffic share of peer-to-peer download application set, and the traffic share of camera application set. The destination IP-side behavioral characteristics include the highest value of destination IP traffic share, the sum of the traffic shares of the three destination IPs with the highest traffic share, the destination IP distribution entropy, and the destination IP concentration index. The port-side behavioral characteristics include the highest value of destination port traffic percentage, the traffic percentage of live streaming port sets, and the traffic percentage of general web port sets; the application traffic percentage is the ratio of the traffic of a single application to the total traffic of the downstream device; the destination IP traffic percentage is the ratio of the traffic of a single destination IP to the total traffic of the downstream device; and the destination port traffic percentage is the ratio of the traffic of a single destination port to the total traffic of the downstream device.

3. The method according to claim 2, characterized in that, Obtaining the terminal device category, terminal manufacturer, and terminal type bias score includes: selecting the device identifier of the downstream device from the deep packet inspection application flow details, associating the device identifier with the terminal fingerprint database to obtain the terminal device category and terminal manufacturer; assigning the terminal type bias score according to the terminal device category, assigning a positive bias score to terminal device categories that tend to run PCDN services, assigning a negative bias score to camera-type terminal device categories, and assigning a zero bias score to the remaining terminal device categories.

4. The method according to claim 3, characterized in that, The calculation of the daily adaptive weight includes: The calculation window is defined as a time window consisting of a preset number of historical backtracking days; the value sequence of each numerical behavior feature within the calculation window is taken; the value sequence of each PCDN behavior anchor point within the calculation window is taken; the daily adaptive weight is calculated using the following formula: W i =|r(f i ,A1)|+|r(f i ,A2)|+|r(f i ,A3)| In i =In i ÷ΣW i In the formula, f i For the i-th numerical class behavioral feature, r(f) i A1) represents the numerical behavioral feature f. i The correlation coefficient between the value sequence of point A1 and the value sequence of anchor point A1, r(f i A2) represents the numerical behavioral feature f. i The correlation coefficient between the value sequence of point A1 and the value sequence of anchor point A2, r(f i A3) represents the numerical behavioral feature f. i The correlation coefficient W between the value sequence of point A3 and the value sequence of anchor point A3. i For numerical behavior features f i The original weights, w i For numerical behavior features f i The adaptive weights of the day, ΣW i It is the sum of the original weights of all numerical behavioral features; anchor point A1 is the application ratio of PCDN features, anchor point A2 is the degree of application homogeneity, and anchor point A3 is the highest value of the destination IP traffic ratio. The degree of application homogeneity is equal to 1 minus the application distribution entropy, and the application distribution entropy is the normalized information entropy of the application distribution.

5. The method according to claim 4, characterized in that, The daily behavior score of the attached device is calculated using the following formula: S=Σ[w i ×s(f i )]+b, When numerical behavior characteristics f i When considering entropy-based behavioral features and negative application proportion-based behavioral features, s(f i )=1-p i , When numerical behavior characteristics f i When the behavioral features are numerical behavioral features other than entropy-based behavioral features and negative application proportion-based behavioral features, s(f i )=p i , In the formula, S represents the daily behavioral score of the attached device, and w i For numerical behavior features f i The adaptive weights of the day, p i For numerical behavior features f i The daily quantile, s(f i ) represents the numerical behavioral feature f i The participation score is b, the terminal type bias score is b, and Σ[wi×s(fi)] is the sum of the weighted participation scores of all numerical behavioral features. The daily quantile refers to the relative ranking position of the numerical behavioral feature in all samples on that day; the all samples on that day refer to the numerical behavioral features of all connected devices on that day. The entropy-type behavioral characteristics include application distribution entropy and destination IP distribution entropy; the negative application proportion-type behavioral characteristics include the traffic proportion of live streaming application sets, the traffic proportion of peer-to-peer download application sets, and the traffic proportion of camera application sets.

6. The method according to claim 5, characterized in that, The generation of evidence-based non-PCDN explanation tags includes: The comprehensive percentile score for live streaming applications is calculated based on the traffic share of live streaming applications; the comprehensive percentile score for peer-to-peer download applications is calculated based on the traffic share of peer-to-peer download applications; the comprehensive percentile score for camera applications is calculated based on the traffic share of camera applications; when the highest value among the three comprehensive percentile scores reaches the rejection percentile threshold, a non-PCDN explanation tag is generated. Candidate evidence is obtained, including application evidence, terminal evidence, and port evidence. The application evidence is the application identifier of the downstream device, the terminal evidence is the terminal device category of the downstream device, and the port evidence is the destination port of the downstream device. Primary evidence is selected from the candidate evidence in the order of application evidence, terminal evidence, and port evidence, with the selection order being: first, application evidence; if application evidence is missing, terminal evidence; if terminal evidence is missing, port evidence. The primary evidence is then associated with the non-PCDN interpretation tag to obtain a non-PCDN interpretation tag with evidence.

7. The method according to claim 6, characterized in that, The obtained gateway PCDN prediction flag includes: Load gateway WAN port traffic data and LAN port traffic data; extract the gateway daily traffic sequence from the gateway WAN port traffic data and LAN port traffic data; The inter-day correlation coefficient, root mean square error, coefficient of variation, and burstiness of the daily traffic series of adjacent business days of the gateway are used as the extracted features. The deep packet inspection application flow details are used to statistically analyze the gateway's home broadband access ratio, gateway application flow recognition rate, number of accesses and types of specific domain names, gateway port traffic distribution, and gateway preset application set traffic. The gateway's home broadband access ratio, gateway application flow recognition rate, number of accesses and types of specific domain names, gateway port traffic distribution, and gateway preset application set traffic are used as statistical features. The extracted features and the statistically obtained features are included in the model input information; the model input information and the numerical class behavior features are input into the gradient boosting tree, and the gateway PCDN prediction flag is output.

8. The method according to claim 7, characterized in that, The removal of gateway PCDN prediction flags that fail strong rule verification includes: For gateways carrying the PCDN prediction flag, the pass conditions are verified sequentially. The pass conditions are: the gateway's uplink traffic on the current day is higher than the preset uplink threshold on the current day; the gateway's average uplink traffic over the past few days is higher than the preset average uplink threshold; the gateway is not on the list of gateways with sudden characteristics; the gateway's home broadband access ratio is higher than the preset home broadband access ratio threshold; the PCDN prediction flag is retained for gateways that meet all the pass conditions; the PCDN prediction flag is cleared for gateways that do not meet the pass conditions.

9. The method according to claim 8, characterized in that, The removal of non-PCDN explanation tags related to live streaming includes: Non-PCDN explanation tags for live streaming that meet the misidentification filtering criteria based on the proportion of home broadband access and the application stream recognition rate will be removed.

10. The method according to claim 9, characterized in that, The non-PCDN interpretation tag and PCDN prediction flag of the downstream device with the largest uplink traffic within the integrated gateway are used to obtain the gateway's PCDN service identification results, including: Based on the uplink traffic, select the downstream device with the largest uplink traffic within the gateway; Calculate the proportion of uplink traffic of the gateway to the total uplink traffic of the connected devices (cameras) within the gateway. Gateways whose uplink traffic accounts for a preset threshold of their total uplink traffic are classified as high-traffic non-PCDN users. Gateways whose downstream device terminal manufacturers belong to a preset set of camera manufacturer determinations are also classified as high-traffic non-PCDN users. The gateway with the evidence-based non-PCDN interpretation tag of the downstream device with the largest uplink traffic is read. Gateways carrying the non-PCDN interpretation tag are classified as high-traffic non-PCDN users. Gateways that are not high-traffic non-PCDN users but carry a PCDN prediction flag are classified as PCDN users. Gateways that are not high-traffic non-PCDN users and do not carry a PCDN prediction flag are classified as normal users, thus obtaining the gateway PCDN service identification result.