A method, electronic device and medium for identifying website collections

By combining online training with a lightweight Transformer, the problem of high costs in collecting real traffic and computation in multi-tab website recognition is solved, achieving efficient and low-cost multi-tab website recognition.

CN123074949APending Publication Date: 2026-11-10CHONGQING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611585794.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing multi-tab website fingerprinting technologies suffer from problems such as high cost of collecting real multi-tab training traffic, insufficient model generalization ability, high computational cost, and inaccurate data synthesis, especially inefficient when dealing with combinatorial explosion and overlapping window issues.

Method used

An online training method is adopted to dynamically synthesize multi-tab training samples based on single-tab traffic trajectories under multi-granularity state set constraints. Temporal convolutional coding and padding mask processing are used, combined with a lightweight Transformer for global aggregation, and the website existence probability is output to achieve single-stage inference.

Benefits of technology

It reduces the need for real traffic collection, improves the model's generalization ability, reduces computational costs and storage overhead, and achieves efficient multi-tab website recognition.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074949A_ABST
    Figure CN123074949A_ABST
Patent Text Reader

Abstract

This invention discloses a website set identification method, including online training and single-stage inference. The online training includes: dynamically synthesizing multi-tab training samples online based on the traffic trajectories of a single tab homepage and subpages, under the constraint of a multi-granularity set of used states, generating fixed-length combined trajectories and website-level multi-hot tags; temporally convolutionally encoding the fixed-length combined trajectories and generating a padding mask; dividing the encoded features into overlapping local windows, and performing mask mean pooling in combination with the padding mask; after window projection and the introduction of a learnable classification token, global aggregation by Transformer, outputting the probability of candidate website existence, and updating model parameters based on website-level multi-hot tags. The training samples are directly randomly combined online from single-tab homepage / subpage records in each iteration, avoiding large-scale collection of real concurrent traffic compared to existing technologies, and continuously expanding the training combination distribution without requiring real multi-tab training traffic and subsequent fine-tuning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of information technology and provides a method, electronic device and medium for identifying website collections. Background Technology

[0002] Multi-tab website fingerprinting technology aims to recover and identify the set of websites currently visited from the mixed encrypted network traffic generated when a user simultaneously visits multiple websites. In the early stages of this technology's development, the mainstream approach was to first segment the mixed traffic or locate the traffic boundaries corresponding to different tabs, and then identify each website segment individually. With the development of deep learning technology, current research has gradually adopted end-to-end technical solutions such as convolutional neural networks, sliding windows, attention mechanisms, and multi-tab classification for multi-tab website fingerprinting. Some studies have also synthesized multi-tab training data using single-tab traffic samples to alleviate the difficulty of obtaining real samples; however, these solutions mostly use pre-generated fixed synthetic datasets or still require a small amount of real multi-tab traffic to participate in model fine-tuning.

[0003] Existing multi-tab website fingerprinting technology still has several technical shortcomings.

[0004] (1) The cost of collecting training traffic for real multi-tab pages is high, and the combination of concurrent websites is explosive. Fixed collection or fixed offline synthesis data can only cover limited co-occurrence patterns. The model is prone to learning specific combination relationships, and its generalization ability to unseen website combinations and subpage access is insufficient.

[0005] (2) When using only the homepage / subpage trajectory of a single tab to dynamically construct training samples, it is necessary to simultaneously constrain different granular states such as "whether the specific trajectory has been used, whether the main class of a website homepage has appeared, and whether the subclass of a subpage has appeared"; if such constraints are lacking, duplicate trajectories or training samples that do not conform to the predetermined combination rules may be generated.

[0006] (3) Online splicing introduces artificial segment boundaries, and fixed-length input generates zero-padding regions. If global self-attention is applied directly to the entire long sequence, the computational cost is high; if the windowing process does not distinguish between valid flow and zero-padding positions, invalid positions will change the local descriptor. Therefore, it is necessary to combine overlapping windows with padding masks: the former reduces the sensitivity to fixed split points and artificial splicing boundaries, while the latter eliminates the contribution of zero-padding positions when windows converge. Summary of the Invention

[0007] In view of the above, the purpose of this invention is to provide a method for identifying website collections.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying website collections, comprising online training and single-stage inference; The online training includes: Based on the traffic trajectory of a single tab homepage and subpage, multi-tab training samples are dynamically synthesized online under the constraint of multi-granularity used state set, generating fixed-length combined trajectory and website-level multi-hot tags; Temporal convolutional encoding of fixed-length combined trajectories and generation of padding masks; The encoded features are divided into overlapping local windows, and mask mean pooling is performed in combination with the padding mask. After window projection and the introduction of learnable classification tags (Tokens), the Transformer is globally aggregated to output the probability of candidate websites. The model parameters are then updated based on website-level multi-hot tags. The single-stage reasoning includes: The preprocessed real multi-tab encrypted traffic is directly input into the trained model. No sample synthesis is performed and the number of active tabs does not need to be known. The time-series encoding, window aggregation, global aggregation and multi-tab output links of the trained model are used for a single forward calculation. The set of visited websites is obtained based on the probability of the output websites.

[0009] Furthermore, the multi-granularity used state set includes a trajectory identity set, a homepage website main class set, and a subpage page subclass set; When synthesizing training samples online, the number of target segments k is sampled, and the trajectory of the main page or subpage is selected with a preset probability; The homepage trajectory verifies the trajectory identity and confirms that the website's main class is not in use. The subpage trajectory verifies the trajectory identity and confirms that the page's subclass is not in use. Once the verification is successful, the corresponding status set is updated.

[0010] Furthermore, the maximum number of attempts for the candidate fragment is set to max(20,12k). When the candidate segment set is empty, a subpage trajectory is randomly selected as the backtracking sample.

[0011] Furthermore, the generation of fixed-length combined trajectories includes shuffling candidate segments and writing them into a fixed-length sequence, randomly truncating sub-segments from ultra-long segments, and padding the remaining positions with zeros; The website-level multi-hot tags are obtained by ORing the main website class of the selected segment, representing the selected website set rather than time-based location tags.

[0012] Furthermore, the temporal convolutional coding employs a multi-level one-dimensional convolutional block extraction local flow pattern; The padding mask is a binary mask that distinguishes between the actual traffic flow and the zero-padding region. It forms position validity information based on the valid data positions in the fixed-length combined trajectory and transforms the position validity information to a time scale corresponding to the encoded features.

[0013] Furthermore, the overlapping local window reduces the sensitivity to splicing boundaries and fixed split points; Mask mean pooling uses a padding mask to shield zero-filling positions and sets a stabilizing term to avoid zero in the denominator.

[0014] Furthermore, the window descriptor is projected and concatenated to form a learnable [CLS] token, which is then fed into the Transformer encoder. The [CLS] output is taken as the global representation, and after passing through a linear classification head and Sigmoid activation, the independent existence probability of each website is obtained.

[0015] Furthermore, during the training phase, Asymmetric loss (ASL) is used to supervise model updates; During the inference phase, the probability of existence is determined based on a configurable threshold, and the set of identified websites is output.

[0016] An electronic device for identifying website collections, implementing any of the methods described above, includes: The online sample synthesis module is used to generate fixed-length combined trajectories and website-level multi-hot tags from single-tab trajectories under the constraints of multi-granularity used state sets; The feature encoding and mask generation module is used for temporal convolutional encoding and generating aligned padding masks; The window aggregation module is used to divide overlapping local windows and perform mask mean pooling; The global aggregation and multi-tab output module is used for window projection, token splicing, global aggregation with Transformer, and outputting the probability of candidate websites. The model training module is used to update parameters based on multiple hot tags at the website level; The inference execution module is used to take real multi-tab encrypted traffic as input, and outputs the set of visited websites in a single forward calculation without knowing the number of active tabs.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described website collection identification method.

[0018] The beneficial effects of this invention are: 1. In the technical concept of this invention, the training samples are directly randomly combined online from the single tab home page / subpage records in each iteration. Compared with the existing technology, this avoids large-scale collection of real concurrent traffic and continuously expands the training combination distribution without the need for real multi-tab training traffic and subsequent fine-tuning.

[0019] 2. In the technical concept of this invention, the combination rules simultaneously cover the homepage, subpage, and trajectory identity; by restricting specific trajectories, homepage main class, and subpage subclass by the trajectory identity set, homepage website main class set, and subpage page subclass set, it avoids duplicates from the same source and allows the same website to appear together in the form of homepage and different subpages, which is closer to actual browsing behavior.

[0020] 3. PaddingMask eliminates the contribution of zero-padding positions to the window descriptor through MaskedMeanPooling; overlapping windows reduce the sensitivity caused by manually combined boundaries and fixed window split points; PaddingMask and overlapping windows solve different problems respectively.

[0021] 4. The lightweight Transformer aggregates evidence from various windows, and the classification head directly outputs the independent existence probability of each candidate website. There is no need to estimate the number of active tabs or use auxiliary decision modules. Single-stage multi-tab reasoning does not depend on the number of tabs.

[0022] 5. The model structure is lightweight, reducing storage and deployment overhead while maintaining multi-tab recognition capabilities. Attached Figure Description

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a block diagram of the main steps of method one of the present invention.

[0024] Figure 2 This is a block diagram of a sub-step of the method of the present invention.

[0025] Figure 3 This represents the main data relationship between online training and single-stage inference in the technical concept of this invention. Detailed Implementation

[0026] The present invention will be further described below with reference to specific embodiments. The accompanying drawings are for illustrative purposes only, representing schematic diagrams rather than actual physical objects, and should not be construed as limiting the scope of this patent. To better illustrate the embodiments of the present invention, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0027] refer to Figure 1-2 As shown, a method for identifying website collections includes the following steps: S100, online training.

[0028] S200, Single-stage reasoning.

[0029] S100 includes: S110: Based on the traffic trajectory of a single tab homepage and subpage, multi-tab training samples are dynamically synthesized online under the constraint of multi-granularity used state sets, generating fixed-length combined trajectories and website-level multi-hot tags.

[0030] In specific implementation, the multi-granularity used state set includes a trajectory identity set Uid, a homepage website main class set Umain, and a subpage page subclass set Usub; when synthesizing training samples online, the number of target segments sampled is k, with a preset probability. Choose either the homepage or subpage track; the homepage track verifies the track identity and confirms that the website main class is not in use, while the subpage track verifies the track identity and confirms that the page subclass is not in use. After successful verification, update the corresponding status set.

[0031] Furthermore, the maximum number of attempts for a candidate segment is set to max(20, 12k); when the candidate segment set is empty, a subpage trajectory is randomly selected as a backtracking sample. The generation of fixed-length combined trajectories includes shuffling the candidate segments and writing them into a fixed-length sequence, randomly truncating sub-segments from excessively long segments, and padding the remaining positions with zeros; the website-level multi-hot tags are obtained by ORing the main website class of the selected segments, representing the selected website set rather than time-sequential position tags.

[0032] The specific processing and principles in step S110 are as follows.

[0033] Establish a homepage trajectory set H and a subpage trajectory set S.

[0034] Each homepage trajectory has a trajectory identity and a website main class, while subpage trajectories also have a page subclass and its corresponding website main class. Each time a training sample is constructed, an all-zero fixed-length sequence x, all-zero website-level multi-hot labels y, an empty fragment list Q, and three used state sets are initialized: the trajectory identity set Uid, the homepage website main class set Umain, and the subpage page subclass set Usub.

[0035] The target fragment number k is sampled from a predefined distribution PK. In a specific training configuration, k is uniformly sampled from {5, 6}.

[0036] In the candidate selection process, probability is used Choose the homepage track; otherwise, choose the subpage track.

[0037] Homepage candidates must meet the following conditions: their trajectory identity does not appear in the trajectory identity set Uid, and their website main class does not appear in the homepage website main class set Umain. Subpage candidates must meet the following conditions: their trajectory identity does not appear in the trajectory identity set Uid, and their page subclass does not appear in the subpage page subclass set Usub.

[0038] After a candidate passes the constraint, it is clipped to the active segment (the last non-zero signal position), added to Q, and the corresponding website main class is set to 1 in y. At the same time, the corresponding state set is updated.

[0039] In one specific implementation, a maximum number of attempts is set to max(20,12k). If Q is still empty in the end, a subpage trajectory is randomly selected as a backtracking sample.

[0040] The collected fragments are randomly shuffled and then written into a fixed-length input of length L.

[0041] If the current segment length exceeds the remaining space, a continuous sub-window of equal length to the remaining space is randomly selected from the segment. If there are still empty spaces after all segments are written, they are padded with zero values.

[0042] The website-level tag y is formed based on its website class when the fragment is accepted, thus representing the set of websites selected during the sample-building phase, rather than position-by-position tags for each time position after truncation.

[0043] The process of generating online random combinations and website-level multi-hot tags can be represented as follows:

[0044] Among them, s i For the i-th selected single-tab fragment, c i For its main website class, ⊕ indicates sequential concatenation, Truncate(·) indicates obtaining a fixed length L by truncation or zero padding, e_{c i} represents the main class c of the website. i The one-hot vector, ∨ represents the element-wise OR operation to obtain multi-hot website tags.

[0045] like Figure 3 As shown, this scheme organizes training data construction, model training, and deployment inference into interconnected data flows. During training, the single-tab homepage / subpage trajectory is first randomly combined online by state constraints such as the trajectory identity set Uid, the homepage website main class set Umain, and the subpage page subclass set Usub, to obtain a fixed-length combined trajectory x and a website-level multi-hot tag y.

[0046] The combined trajectory is sequentially processed by temporal convolutional encoding and PaddingMask, corresponding overlapping windows, MaskedMeanPooling, window projection, and global aggregation with a Transformer with [CLS]Token, and finally outputs the existence probability of each candidate website.

[0047] Website-level labels y are used together with model output for ASL training supervision.

[0048] During inference, real multi-tab traffic directly enters the recognition model's path, without performing online random combination or requiring input of the actual number of tabs. The solid arrows in the diagram represent forward processing relationships, while the dashed lines represent supervision relationships during training.

[0049] S120. Perform temporal convolutional encoding on the fixed-length combined trajectory and generate a padding mask.

[0050] In practical implementation, the temporal convolutional coding employs multi-level one-dimensional convolutional blocks to extract local flow patterns. The padding mask is a binary mask that distinguishes between the actual flow and the zero-padding region. Position validity information is formed based on the effective data positions in the fixed-length combined trajectory, and this position validity information is transformed to a time scale corresponding to the coding features.

[0051] The specific processing and principles involved in this step are as follows.

[0052] The fixed-length flow sequence first enters the temporal convolutional encoder.

[0053] The temporal convolutional encoder uses multi-level one-dimensional convolutional blocks to progressively compress the time length and extract local flow patterns that are independent of absolute position.

[0054] In one specific embodiment of the structural configuration, the encoder includes four ConvStages, each consisting of two convolutional blocks, and outputs a local feature map with a time length of 1250 and 256 channels. Simultaneously, a binary PaddingMask is generated to distinguish between real traffic regions and zero-padding regions, and downsampled over the encoded time scale to correspond to the temporal position of the local feature map.

[0055] S130. Divide the encoded features into overlapping local windows and perform mask mean pooling in combination with the padding mask.

[0056] In practical implementation, the overlapping local window reduces the sensitivity to splicing boundaries and fixed split points; the mask mean pooling uses a filling mask to shield zero-filling positions and sets a stable term to avoid the denominator being zero.

[0057] The specific processing and principles involved in this step are as follows.

[0058] The encoded feature map is still quite long. Instead of directly applying self-attention to all time locations, it is divided into overlapping local windows.

[0059] In one specific embodiment of the structural configuration, the feature window size is 256×48, the corresponding Mask window length is 48, and a total of 51 windows are formed.

[0060] Overlapping windows allow the same artificially stitched boundary to be in different relative positions in different windows, thereby reducing the model's sensitivity to fixed cut points and stitching artifacts.

[0061] MaskedMeanPooling for the nth window can be represented as:

[0062] Among them, F n For the nth local feature window, M′ n The PaddingMask corresponds to this window, where W is the window length, ⊙ represents element-wise multiplication (the Mask is broadcast in the channel dimension), and ε is a stabilizing term to avoid a denominator of zero. This formula directly uses the Mask in the summation and normalization, ensuring that zero-padding positions are not considered as evidence of normal flow in the window descriptor h. n The formation of.

[0063] When a fixed-length trajectory has a lot of zeros padded at the end, overlapping windowing may result in the entire window consisting of zero-padded positions and containing no real traffic at all. The associated code further determines whether each window contains at least one valid position; for windows with no valid positions, they are masked during subsequent global aggregation by the Transformer and do not participate in attention calculations.

[0064] S140. After window projection and the introduction of learnable classification tags (Tokens), the Transformer performs global aggregation, outputting the probability of candidate websites existing, and updating the model parameters based on website-level multi-hot tags.

[0065] In practice, the window descriptor is projected and concatenated to form a learnable [CLS] token, which is then fed into the Transformer encoder. The [CLS] output is taken as the global representation, and the probability of each website's independent existence is obtained by passing it through a linear classification head and Sigmoid activation.

[0066] The specific processing and principles involved in this step are as follows.

[0067] The 51 window descriptors are first projected through a linear layer, GELU, and Dropout.

[0068] In one specific architecture, each 256-dimensional window descriptor is projected to 160 dimensions, resulting in a 51×160 window sequence. A learnable [CLS] token is then added, forming a 52×160 sequence, which is then input into a lightweight Transformer encoder. The Transformer integrates dependencies between distant windows through self-attention, and the [CLS] output serves as a global representation of the entire traffic trajectory.

[0069] Finally, the linear classification head outputs the multi-label logits of each candidate website, and the independent existence probability of each website is obtained through the Sigmoid function.

[0070] During training, Asymmetric Loss (ASL) is used to handle multi-label class imbalance, where γ + =1.0, γ - =2.0. During inference, the classification head obtains the independent existence probability of each candidate website through the Sigmoid function and forms a website set according to the multi-label judgment rule; the specific judgment threshold is a configurable parameter. During the training phase, the Asymmetric Loss (ASL) is used to supervise model updates; during the inference phase, the existence probability is determined according to the configurable judgment threshold, and the identified website set is output.

[0071] S200 includes: The preprocessed real multi-tab encrypted traffic is directly input into the trained model. No sample synthesis is performed and the number of active tabs does not need to be known. The time-series encoding, window aggregation, global aggregation and multi-tab output links of the trained model are used for a single forward calculation. The set of visited websites is obtained based on the probability of the output websites.

[0072] In this step, the real multi-tab encrypted traffic is preprocessed according to the model input requirements and directly enters the classification model. Online combination is not performed, and the actual number of tabs is not input. The model directly obtains the complete website set through a single forward computation, without the need for an auxiliary tab count estimation module or prior knowledge of the number of active tabs.

[0073] An electronic device for identifying website collections, implementing the above-described method, includes: The online sample synthesis module is used to generate fixed-length combined trajectories and website-level multi-hot tags from single-tab trajectories under the constraints of multi-granularity used state sets; The feature encoding and mask generation module is used for temporal convolutional encoding and generating aligned padding masks; The window aggregation module is used to divide overlapping local windows and perform mask mean pooling; The global aggregation and multi-tab output module is used for window projection, token splicing, global aggregation with Transformer, and outputting the probability of candidate websites. The model training module is used to update parameters based on multiple hot tags at the website level; The inference execution module is used to take real multi-tab encrypted traffic as input, and outputs the set of visited websites in a single forward calculation without knowing the number of active tabs.

[0074] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the website collection identification method.

[0075] Through the above implementation method, compared with the existing multi-tab website identification schemes, the main technical advantages of this solution are as follows: 1. No need for real multi-tab training traffic and subsequent fine-tuning. Training samples are directly obtained from online random combinations of single-tab homepage / subpage records in each iteration, avoiding large-scale collection of real concurrent traffic and continuously expanding the training combination distribution.

[0076] 2. The combination rules simultaneously cover the homepage, subpages, and browsing history identities. By using Uid, Umain, and Usub to restrict specific browsing history, the main website class of the homepage, and the subpage subclass, it avoids duplicate entries from the same origin while allowing the same website to appear as both a homepage and different subpages, more closely resembling actual browsing behavior.

[0077] 3. PaddingMask and overlapping windows address different issues. PaddingMask uses MaskedMeanPooling to eliminate the contribution of zero-padding positions to the window descriptor; overlapping windows reduce the sensitivity to manually combined boundaries and fixed window split points.

[0078] 4. Single-stage multi-label inference does not depend on the number of tabs. The lightweight Transformer aggregates evidence from various windows, and the classification head directly outputs the independent existence probability of each candidate website, without the need to estimate the number of active tabs or use auxiliary decision-making modules.

[0079] 5. Lightweight model structure. In one implemented architecture, the model contains 441,874 trainable parameters and is approximately 1.69 MB in size, reducing storage and deployment overhead while maintaining multi-tab recognition capabilities.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying website collections, characterized in that, This includes online training and single-stage inference; The online training includes: Based on the traffic trajectory of a single tab homepage and subpage, multi-tab training samples are dynamically synthesized online under the constraint of multi-granularity used state set, generating fixed-length combined trajectory and website-level multi-hot tags; Temporal convolutional encoding of fixed-length combined trajectories and generation of padding masks; The encoded features are divided into overlapping local windows, and mask mean pooling is performed in conjunction with the padding mask. After window projection and the introduction of learnable classification tags (Tokens), the Transformer is globally aggregated to output the probability of candidate websites. The model parameters are then updated based on website-level multi-hot tags. The single-stage reasoning includes: The preprocessed real multi-tab encrypted traffic is directly input into the trained model. No sample synthesis is performed and the number of active tabs does not need to be known. The time-series encoding, window aggregation, global aggregation and multi-tab output links of the trained model are used for a single forward calculation. The set of visited websites is obtained based on the probability of the output websites.

2. The website collection identification method according to claim 1, characterized in that, The multi-granularity used state set includes a trajectory identity set, a homepage website main class set, and a subpage page subclass set; When synthesizing training samples online, the number of target segments k is sampled, and the trajectory of the main page or subpage is selected with a preset probability; The homepage trajectory verifies the trajectory identity and confirms that the website's main class is not in use. The subpage trajectory verifies the trajectory identity and confirms that the page's subclass is not in use. Once the verification is successful, the corresponding status set is updated.

3. The website collection identification method according to claim 2, characterized in that, Set the maximum number of attempts for the candidate fragment to max(20, 12k); When the candidate segment set is empty, a subpage trajectory is randomly selected as the backtracking sample.

4. The website collection identification method according to claim 1, characterized in that, The process of generating a fixed-length combined trajectory includes shuffling candidate segments and writing them into a fixed-length sequence, randomly truncating sub-segments from an excessively long segment, and padding the remaining positions with zeros. The website-level multi-hot tags are obtained by ORing the main website class of the selected segment, representing the selected website set rather than time-based location tags.

5. The website collection identification method according to claim 1, characterized in that, The temporal convolutional coding employs a multi-level one-dimensional convolutional block extraction local flow pattern. The padding mask is a binary mask that distinguishes between the actual traffic flow and the zero-padding region. It forms position validity information based on the valid data positions in the fixed-length combined trajectory and transforms the position validity information to a time scale corresponding to the encoded features.

6. The website collection identification method according to claim 1, characterized in that, The overlapping local window reduces sensitivity to splicing boundaries and fixed cutting points; Mask mean pooling uses a padding mask to shield zero-filling positions and sets a stabilizing term to avoid zero in the denominator.

7. The website collection identification method according to claim 1, characterized in that, The window descriptor is projected and concatenated to learnable [CLS] tokens, which are then fed into the Transformer encoder. The [CLS] output is taken as the global representation, and the probability of each website's independent existence is obtained by passing it through a linear classification head and Sigmoid activation.

8. The website collection identification method according to claim 1, characterized in that, During the training phase, Asymmetric loss (ASL) is used to supervise model updates. During the inference phase, the probability of existence is determined based on a configurable threshold, and the set of identified websites is output.

9. An electronic device for identifying a collection of websites, characterized in that, Implementing the method according to any one of claims 1-8, comprising: The online sample synthesis module is used to generate fixed-length combined trajectories and website-level multi-hot tags from single-tab trajectories under the constraints of multi-granularity used state sets; The feature encoding and mask generation module is used for temporal convolutional encoding and generating aligned padding masks; The window aggregation module is used to divide overlapping local windows and perform mask mean pooling; The global aggregation and multi-tab output module is used for window projection, token splicing, global aggregation with Transformer, and outputting the probability of candidate websites. The model training module is used to update parameters based on multiple hot tags at the website level; The inference execution module is used to take real multi-tab encrypted traffic as input, and outputs the set of visited websites in a single forward calculation without knowing the number of active tabs.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the website collection identification method according to any one of claims 1-8.