Space-time perception-based network hot review confrontation effectiveness adaptive evaluation method and system

Abstract

The application provides a network hot review confrontation performance self-adaptive evaluation method and system based on space-time perception and a terminal device, and is suitable for the technical field of network security. The method comprises the following steps: inputting multi-source network data into a multi-scale space-time data fusion model to generate multi-scale space-time features; generating a fusion feature vector through a feature depth fusion processing model based on the multi-scale space-time features; training a three-party strategy network model to be trained based on the fusion feature vector, determining a three-party game equilibrium strategy and an evaluation score corresponding to the three-party game equilibrium strategy; adjusting evaluation indexes through an adaptive evaluation model based on environmental state features to obtain new evaluation index weights; and determining an adaptive evaluation score based on the new evaluation index weights and the evaluation score.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of network security, and particularly relates to a network hot comment confrontation effectiveness adaptive evaluation method and system based on space-time perception. BACKGROUND

[0002] With the wide application of mobile Internet and social media, network comments have become an important carrier of public opinion generation and emotion diffusion, and the propagation process is easily affected by multiple confrontation interference such as malicious manipulation, false information injection and emotion polarization. Under this background, the public opinion monitoring system needs to have high-precision perception and evaluation capability for the dynamic propagation process to ensure the healthy development of the network public opinion environment.

[0003] However, the evaluation index system used by the existing public opinion monitoring technology is mostly fixed design, and in the real scene of dynamic change of network environment, continuous evolution of confrontation strategy and change of propagation mode, it is difficult to match the evaluation needs in different scenes, and the adaptability to complex dynamic confrontation scene is obviously insufficient. SUMMARY

[0004] Therefore, the embodiment of the application provides a network hot comment confrontation effectiveness adaptive evaluation method and system based on space-time perception and a terminal device, which can solve the problem of insufficient adaptability to complex dynamic confrontation scenes.

[0005] The first aspect of the embodiment of the application provides a network hot comment confrontation effectiveness adaptive evaluation method based on space-time perception, comprising:

[0006] inputting multi-source network data into a multi-scale space-time data fusion model to generate multi-scale space-time features;

[0007] generating a fusion feature vector through a feature depth fusion processing model based on the multi-scale space-time features;

[0008] training a three-party strategy network model to be trained based on the fusion feature vector, determining a three-party game equilibrium strategy and an evaluation score corresponding to the three-party game equilibrium strategy;

[0009] adjusting the evaluation index through an adaptive evaluation model based on the environment state features to obtain a new evaluation index weight;

[0010] determining an adaptive evaluation score based on the new evaluation index weight and the evaluation score.

[0011] The second aspect of the embodiment of the application provides a network hot comment confrontation effectiveness adaptive evaluation system based on space-time perception, comprising:

[0012] a space-time perception collection module configured to input multi-source network data and generate multi-scale space-time features;

[0013] a feature deep fusion processing module configured to generate a fusion feature vector based on the multi-scale spatio-temporal features;

[0014] a multi-party game decision module configured to train a three-party strategy network model based on the fusion feature vector, determine a three-party game equilibrium strategy and an evaluation score corresponding to the three-party game equilibrium strategy;

[0015] an adaptive evaluation score determination module configured to adjust the evaluation indexes by an adaptive evaluation model based on the environment state features to obtain new evaluation index weights, and determine an adaptive evaluation score based on the new evaluation index weights and the evaluation score.

[0016] A third aspect of the embodiments of the present application provides a terminal device, which comprises a memory and a processor, the memory stores a computer program capable of running on the processor, and the processor implements the steps of the network hot review countermeasure performance adaptive evaluation method based on spatio-temporal perception according to any one of the above first aspect when executing the computer program.

[0017] A fourth aspect of the embodiments of the present application provides a computer readable storage medium, which comprises a computer program stored thereon, and the computer program is executed by a processor to implement the steps of the network hot review countermeasure performance adaptive evaluation method based on spatio-temporal perception according to any one of the above first aspect.

[0018] A fifth aspect of the embodiments of the present application provides a computer program product, which, when running on a terminal device, causes the terminal device to execute the network hot review countermeasure performance adaptive evaluation method based on spatio-temporal perception according to any one of the above first aspect.

[0019] Compared with the prior art, the network hot review countermeasure performance adaptive evaluation method based on spatio-temporal perception has the beneficial effects that: the multi-scale spatio-temporal data fusion is used to generate multi-scale spatio-temporal features, the deep fusion is used to obtain a fusion feature vector, the three-party strategy network is constructed to determine a game equilibrium strategy, and finally the adaptive evaluation model is used to obtain an evaluation score, so as to realize real-time, accurate and dynamic evaluation of the network hot review countermeasure performance, effectively improve the dynamic adaptability, multi-layer game modeling capability and evaluation precision, and provide technical support for network public opinion management. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0021] Figure 1 FIG. 1 is a first implementation flow diagram of a network hot review confrontation performance adaptive evaluation method based on space-time perception provided by an embodiment of the present application;

[0022] Figure 2 FIG. 2 is a second implementation flow diagram of a network hot review confrontation performance adaptive evaluation method based on space-time perception provided by an embodiment of the present application;

[0023] Figure 3 FIG. 3 is a third implementation flow diagram of a network hot review confrontation performance adaptive evaluation method based on space-time perception provided by an embodiment of the present application;

[0024] Figure 4 FIG. 4 is a fourth implementation flow diagram of a network hot review confrontation performance adaptive evaluation method based on space-time perception provided by an embodiment of the present application;

[0025] Figure 5 FIG. 5 is a fifth implementation flow diagram of a network hot review confrontation performance adaptive evaluation method based on space-time perception provided by an embodiment of the present application;

[0026] Figure 6 FIG. 6 is a sixth implementation flow diagram of a network hot review confrontation performance adaptive evaluation method based on space-time perception provided by an embodiment of the present application;

[0027] Figure 7 FIG. 7 is a seventh implementation flow diagram of a network hot review confrontation performance adaptive evaluation method based on space-time perception provided by an embodiment of the present application;

[0028] Figure 8 FIG. 8 is a structure diagram of a network hot review confrontation performance adaptive evaluation system based on space-time perception provided by an embodiment of the present application;

[0029] Figure 9 FIG. 9 is a schematic diagram of a terminal device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0030] In the following description, specific details are set forth, such as particular system configurations, techniques, etc., in order to provide a thorough understanding of the present application. However, persons skilled in the art will understand that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

[0031] Actor-Critic Algorithm is a reinforcement learning method combining policy gradient and temporal difference learning, including two parts, Actor and Critic. Actor refers to the policy function, that is, learning a policy to get the highest return, used to generate actions and interact with the environment. Critic refers to the value function, which estimates the value function of the current policy, that is, evaluates the performance of the actor, and guides the action of the actor in the next stage.

[0032] In this application, the Actor-Critic architecture is the key support for implementing the three-party game strategy network model training and subsequent adaptive evaluation model dynamic adjustment of evaluation index weight. Among them, the Actor network is responsible for outputting the strategy actions of each party (attacker, defender, platform party) according to the input fusion feature vector and other information, to explore and determine the possible strategy in the three-party game process. The Critic network evaluates the value of the strategy output by the Actor network, calculates the Q value (state-action value) under the corresponding strategy, and provides guidance for the strategy optimization of the Actor network through this evaluation feedback, so that the three-party strategy network model can continuously learn and converge to Nash equilibrium, and obtain the three-party game equilibrium strategy and the corresponding evaluation score.

[0033] In order to illustrate the technical solutions described in the present application, the following will be described through specific embodiments.

[0034] Figure 1 The first implementation flowchart of the network hot evaluation confrontation effectiveness adaptive evaluation method based on space-time perception provided by the embodiment of the present application is shown. The details are as follows:

[0035] Step S100: inputting multi-source network data into a multi-scale space-time data fusion model to generate multi-scale space-time features;

[0036] Step S200: generating a fusion feature vector through a feature depth fusion processing model based on the multi-scale space-time features;

[0037] Step S300: training the three-party strategy network model based on the fusion feature vector, determining the three-party game equilibrium strategy and the evaluation score corresponding to the three-party game equilibrium strategy;

[0038] Step S400: adjusting the evaluation index through an adaptive evaluation model based on the environment state features to obtain new evaluation index weights;

[0039] Step S500: determining the adaptive evaluation score based on the new evaluation index weights and the evaluation score.

[0040] The network hot comment confrontation performance adaptive evaluation method based on space-time perception generates multi-scale space-time features through multi-scale space-time data fusion, obtains a fusion feature vector through deep fusion, determines a three-party game equilibrium strategy by constructing a three-party strategy network model, and finally obtains an evaluation score from an adaptive evaluation model, thereby realizing real-time, accurate and dynamic evaluation of the network hot comment confrontation performance, effectively improving the dynamic adaptability, multi-layer game modeling capability and evaluation accuracy of the evaluation, and providing technical support for network public opinion management.

[0041] Figure 2 A second implementation process schematic diagram of the network hot comment confrontation performance adaptive evaluation method based on space-time perception provided by the embodiments of the application is shown. In one embodiment, step S100 includes:

[0042] S101, sampling multi-source network data according to a time window to obtain sampling data;

[0043] S102, performing parallel processing on the sampling data through a real-time stream data processing algorithm to obtain temporary storage data;

[0044] S103, constructing an influence network based on the temporary storage data to generate node attribute features;

[0045] S104, performing distributed storage on the temporary storage data to generate multi-scale time markers;

[0046] S105, performing space-time feature fusion based on the node attribute features and the multi-scale time markers to generate multi-scale space-time features;

[0047] The time window includes an hour-level time window, a minute-level time window and a second-level time window, and the sampling data includes hour topic sampling data corresponding to the hour-level time window, minute-level forwarding sampling data corresponding to the minute-level time window, and second-level comment sampling data corresponding to the second-level time window.

[0048] Exemplarily, second-level comment sampling is performed through a second-level time window of 1 second to capture user immediate reactions and emotional fluctuations; minute-level forwarding sampling is performed through a minute-level time window of 60 seconds to track information propagation paths and influence ranges; and hour-level topic sampling is performed through an hour-level time window of 3600 seconds to monitor long-term trends and topic development.

[0049] In step S101, exemplarily, the second-level time window is 1 second, which is used to capture comment interaction dynamics; the minute-level time window is 1 minute, which is used to track forwarding wave times; and the hour-level time window is 1 hour, which is used to monitor public opinion evolution. Specifically, the multi-source network data includes comment information, forwarding chain information and user interaction information.

[0050] In step S102, specifically, the hourly topic sampling data, the minute-level forwarding sampling data, and the second-level comment sampling data are input into real-time stream data processing algorithms for parallel processing to obtain temporary storage data. The real-time stream data processing algorithms are based on a Storm stream computing framework to build a three-layer processing topology: a Spout layer (parallelism degree 16) is responsible for data access, a Bolt layer (including a filtering Bolt with a parallelism degree of 8, an analysis Bolt with a parallelism degree of 12, and an aggregation Bolt with a parallelism degree of 4) is responsible for data cleaning and preliminary feature extraction, and a result temporary storage layer uses a ring buffer (capacity 10,000 records, and replaced according to a FIFO strategy when full) to realize temporary storage, that is, to obtain the temporary storage data.

[0051] In step S102, exemplarily, three parallel message channels are established by using a Kafka message queue to shunt the hourly topic sampling data, the minute-level forwarding sampling data, and the second-level comment sampling data, wherein the second-level message channel is configured with 16 partitions, the minute-level message channel is configured with 8 partitions, and the hourly message channel is configured with 4 partitions; the shunted sampling data is high-speed stored by a real-time buffering mechanism to obtain the temporary storage data.

[0052] Exemplarily, step S103 includes:

[0053] S1031, user basic features of a target user and an interaction relationship between the target user and other users are extracted from the temporary storage data. Specifically, the user basic features of the target user include a number of followers, a comment interaction frequency, and a forwarding diffusion range. The interaction relationship between the target user and other users includes a comment, an @ (mention), a forwarding, a quote, and a follow relationship. The number of followers is the total number of follow relationships established by all other users with the target user.

[0054] S1031, the user basic features are input into an improved PageRank algorithm model for iterative calculation to obtain a first user influence value of each target user.

[0055] Specifically, a user association directed graph is constructed through the interaction relationship between the target user and other users, wherein a node in the graph represents a target user, and a directed edge represents the interaction relationship between the target user and other users. Specifically, S1031a. a user v comments / forwards the content of a user u, and there is a directed edge from v to u. S1031b. the out-degree OutDegree(v) of each user v in the graph is counted, that is, the total number of directed edges from the target user to other users. S1031c. the initial influence value Influence_initial(u) of all users is initialized, and the initial value is allocated according to a preset weight based on the user basic features; the user influence value is iteratively calculated, and the calculation formula of the k+1 iteration is:

[0056] Influence k+1 (u)=(1-d)+d×Σ(Influence k (v) / OutDegree(v))

[0057] wherein d=0.85 is a damping coefficient, used to indicate the probability of the user continuing to browse the current user content; Σ represents the sum of the calculation results of all associated users v pointing to user u (i.e. there is a directed edge from v to u); Influence k (v) is the influence value of user v at the kth iteration; Influence k+1 (v) is the influence value of user v at the (k+1)th iteration. S1031d. Repeat step S1031c. until the difference between the influence values of all users in the adjacent two iterations is less than a preset threshold (such as 10 -6 ), at this time, the converged user influence value Influence(u) is obtained, which is the first user influence value.

[0058] S1032, determine the time interval between the time when each target user behavior occurs and the current processing time, calculate the time decay factor based on the time interval, and dynamically correct the first user influence value of the corresponding target user with the time decay factor to obtain the second user influence value of each target user.

[0059] Specifically, the timestamp T_action of the occurrence of the user behavior and the timestamp T_current of the current processing time are determined, and the time interval Δt=T_current-T_action is calculated; the time interval Δt is substituted into the time decay factor formula TimeDecay=exp(-λ×Δt), wherein λ=0.01 is the decay coefficient, and the time decay factor TimeDecay corresponding to the user behavior is calculated; the user influence value Influence_initial(u) obtained by improving the PageRank algorithm is corrected with the calculated TimeDecay, and the dynamically corrected user influence value is Influence(u)=Influence_initial(u)×TimeDecay, which is the second user influence value.

[0060] S1033, taking each target user as a node, taking the interaction relationship between the target user and other users as a directed edge, taking the second user influence value of the corresponding target user as the initial weight of the node, and taking the strength of the interaction relationship as the edge weight, a user influence network is constructed;

[0061] S1034, based on the user influence network, multi-dimensional features of each user node are extracted and integrated to generate node attribute features of each user node.

[0062] In step S1034, exemplary, the multi-dimensional features include basic attribute features, behavior attribute features, network association attribute features, influence attribute features. Specifically, the static information (such as user ID, registration duration, active platform identifier) associated with the user from the staging data is taken as the basic attribute features of the node attribute features. Based on the historical behavior records of the user in the staging data, the comment interaction frequency per unit time, the forwarding diffusion range (such as the number of user levels covered by forwarding), and the proportion of interaction user types (such as the proportion of ordinary users / verified user interactions) are calculated to form the behavior attribute features. The user influence network is a directed graph. In the user influence network, the number of incoming edges of the node (the number of associated users pointing to the user), the number of outgoing edges (the number of associated users pointed to by the user), and the edge weight distribution (the average influence transmission coefficient between associated users) are counted to form the network association attribute features. The second user influence value, the influence ranking of the user in the network, and the influence fluctuation amplitude (such as the change rate of the second user influence value in the last 1 hour) are taken as the influence attribute features. The basic attribute features, behavior attribute features, network association attribute features, and influence attribute features are normalized (such as mapped to the [0, 1] interval) to eliminate dimensional differences, and then spliced to form a dimension-unified vector. The vector is the node attribute feature of the corresponding user node in the user influence network.

[0063] In step S104, exemplary, three partition tables comment_table, forward_table, and topic_table are created using HBase partition storage, a Rowkey design strategy of "timestamp_userID_dataType" is adopted, and the staging data is distributed and persistently stored, and a millisecond-level Unix timestamp is recorded for each piece of stored data to generate a multi-scale time marker. Specifically, the staging data is shunted according to data type to a distributed storage architecture for persistent storage, and a multi-scale time marker is generated for each piece of data; wherein the distributed storage architecture adopts HBase to create three partition tables (comment_table table for storing second-level comment data, forward_table table for storing minute-level forwarding data, and topic_table table for storing hour-level topic data), Rowkey design is "timestamp_userID_dataType", and TTL is set to 72 hours to automatically clean up expired data. The multi-scale time marker is a millisecond-level Unix timestamp, and is associated with the time window level (second / minute / hour) to which the corresponding data belongs.

[0064] In step S105, the multi-scale spatio-temporal data fusion algorithm is used to perform spatio-temporal feature fusion on the node attribute features and the multi-scale time labels. Specifically, the multi-scale spatio-temporal data fusion algorithm adopts a hierarchical sampling strategy. The second-level time window is a sliding window, the size of the second-level time window is 1 second, and the step of the second-level time window is 0.5 second. The minute-level time window is a fixed window, and a data point is generated every 60 seconds, and the aggregated features are calculated by the weighted average method. The hour-level time window adopts the exponential smoothing algorithm, the smoothing coefficient a = 0.3, and the fusion algorithm formula is: F_multi = w1xF_sec + w2xF_min + w3xF_hour, wherein the weight w1 = 0.5, w2 = 0.3, and w3 = 0.2, F_sec is the second-level aggregated feature, F_min is the minute-level aggregated feature, and F_hour is the hour-level aggregated feature; F_multi is the first multi-scale spatio-temporal feature obtained by the multi-scale spatio-temporal data fusion algorithm.

[0065] Specifically, in step S1051, the node attribute features generated in step S103 are time-aligned with the multi-scale time labels generated in step S104, so that the node attributes and the time labels in the same time window correspond to each other. Specifically, for the second-level, minute-level, and hour-level time scales, the node attribute features in the corresponding time window are extracted, and are associated with the corresponding time labels to form a node attribute sequence with a time stamp. In step S1052, for the node attribute sequences of different time scales, a hierarchical aggregation strategy is used to extract spatio-temporal features. In the second-level time window, the node attribute features are statistically calculated (such as mean, variance, and extreme value) by a sliding window to obtain the second-level aggregated features, so as to capture the instantaneous behavior fluctuations. In the minute-level time window, the node attribute features are aggregated by a fixed window to obtain the minute-level aggregated features, so as to extract the medium-frequency features such as the forwarding path and the propagation range. In the hour-level time window, the exponential smoothing algorithm (smoothing coefficient a = 0.3) is used to perform trend smoothing on the node attribute features to obtain the hour-level aggregated features, so as to extract the long-term evolution features. In step S1053, the aggregated features of the above three time scales are input into the multi-scale spatio-temporal data fusion algorithm, and are fused by weighting to obtain the first multi-scale spatio-temporal feature. The fusion formula is: the first multi-scale spatio-temporal feature F_multi = w1xF_sec + w2xF_min + w3xF_hour; F_sec is the second-level aggregated feature, F_min is the minute-level aggregated feature, and F_hour is the hour-level aggregated feature. The weight w1 = 0.5, w2 = 0.3, and w3 = 0.2, which are used to balance the importance of the three-level time scale features. In step S1054, the first multi-scale spatio-temporal feature F_multi is input into the spatio-temporal encoder, and the time and space dimension information are fused by using the attention mechanism to generate the final multi-scale spatio-temporal feature.

[0066] The method can realize multi-scale accurate capture of network hot comment propagation dynamics by constructing a three-layer spatiotemporal perception network architecture. Traditional evaluation systems often use a single time window for data collection, which is difficult to balance the rapid response of short-term emergencies and the in-depth analysis of long-term propagation trends. The present application captures instantaneous propagation bursts through 1-second high-frequency sampling, traces propagation diffusion processes through 60-second medium-frequency sampling, and analyzes long-term evolution trends through 3600-second low-frequency sampling, forming a stereoscopic perception network covering different time scales. This three-level spatiotemporal perception mechanism not only can detect the budding state of abnormal propagation patterns in time, but also can accurately identify the evolution trajectory of persistent adversarial behavior, achieving all-round and multi-level accurate capture of propagation fine-grained dynamics, and providing a high-quality data basis for subsequent game decision and risk assessment.

[0067] The method mainly realizes functions such as real-time collection of multi-source data, three-level spatiotemporal perception sampling, data preprocessing and identification, spatiotemporal feature labeling, and distributed storage. The data preprocessing process adopts a multi-level cleaning mechanism, including text formatting processing based on regular expressions, sensitive word filtering based on a vocabulary, and outlier detection based on statistical learning, and performs deep semantic analysis on comment content through natural language processing techniques such as part-of-speech tagging, named entity recognition (NER), and sentiment polarity analysis; At the same time, three parallel message channels of Kafka message queue are established to realize data shunting processing, among which the second channel is configured with 16 partitions to support high-concurrency writing, the minute-level channel is configured with 8 partitions to optimize throughput, and the hour-level channel is configured with 4 partitions to ensure orderly processing; three partition tables are created for mass data persistent storage using HBase partition storage, and the Rowkey design strategy is "timestamp_userID_data type" to optimize query performance, and TTL is set to 72 hours to automatically clean up expired data. Deploy a three-layer cache architecture of Redis, L1 cache stores hotspot user influence data (capacity 1GB), L2 cache stores recent topic trend data (capacity 2GB), L3 cache stores graph data of user influence network (capacity 4GB), and each data record is accurately time-stamped (Unix timestamp accurate to millisecond) and calculates key attribute indicators such as user influence, propagation speed, and sentiment intensity.

[0068] Figure 3 A third implementation flowchart of the network hot comment adversarial effectiveness adaptive evaluation method based on spatiotemporal perception provided by the embodiments of the present application is shown. In one embodiment, step S200 includes:

[0069] S201, input the multi-scale spatio-temporal features into a text feature extractor, perform text semantic encoding based on a pre-trained BERT model to obtain a text semantic vector. Illustratively, the pre-trained BERT model adopts a hierarchical attention model to extract key semantic features, a first layer of attention layers is used to extract word-level features, a second layer of attention layers is used to extract sentence-level features, and a third layer of attention layers is used to extract document-level features.

[0070] S202, input the multi-scale spatio-temporal features into a behavior feature extractor to extract time sequence features of user behaviors.

[0071] S203, input the multi-scale spatio-temporal features into a relationship feature extractor to extract relationship features between users.

[0072] S204, input the text semantic vector and the time sequence features of user behaviors into a double-layer LSTM time sequence encoder to obtain a time sequence feature vector. Illustratively, the double-layer LSTM time sequence encoder is a double-layer bidirectional LSTM network with multiple hidden layers, the hidden layer dimension is 256, the number of layers is 2, and the dropout rate is 0.1, which is used to process comment time sequences, capture short-term memory and long-term dependency through a gating mechanism, and extract forward and backward time sequence information at the same time through a bidirectional structure.

[0073] S205, input the relationship features between users into a graph neural network relationship encoder to obtain a relationship feature vector. Illustratively, the graph neural network relationship encoder adopts a graph convolution network model. Specifically, a propagation graph is constructed based on user-user directed edges and weight information, a user influence network is established using interactive relationships and an adjacency matrix is constructed, a normalized Laplacian matrix is used to process to avoid gradient explosion, and spatial structure features are extracted through a two-layer graph convolution network model.

[0074] S206, input the time sequence feature vector and the relationship feature vector into a multi-head attention Transformer model to obtain preliminary fusion features. Illustratively, the multi-head attention Transformer model adopts a 6-head multi-head Transformer encoder, wherein the dimension of the feature vector d_model=512, the number of attention heads nhead=6, the number of Transformer encoder layers num_layers=6, and the dimension of the hidden layer dim_feedforward=2048.

[0075] S207, input the preliminary fusion features into the deep fusion layer for feature integration to obtain a fusion feature vector. The multi-scale feature extractor is used to capture multi-granularity features from the preliminary fusion features, wherein the multi-scale feature extractor uses a parallel convolutional network model to extract local detail features and global pattern features respectively, so as to ensure that the model can pay attention to semantic information of different granularities at the same time. The multi-granularity features output by the multi-scale feature extractor are input into the deep fusion layer for feature integration, and a high-dimensional fusion feature vector is output. The deep fusion layer includes a residual connection layer and a normalization layer. The multi-scale feature extractor uses a parallel convolutional network model including three different convolution kernel sizes of 3*3, 5*5 and 7*7 to capture feature representations of fine granularity, medium granularity and coarse granularity respectively; the residual connection layer realizes direct transmission of feature information through a skip connection, and the normalization layer uses LayerNorm technology to standardize the feature distribution.

[0076] The key components of the method for deep feature extraction and multi-modal data fusion use an LSTM-Transformer hybrid architecture design. First, the bidirectional long short-term memory network (BiLSTM) is used to encode the time series review data to extract short-term memory features and long-term dependencies. In parallel, the graph convolution network (GCN) model is used to process the user influence network to learn the topological structure features and propagation patterns in the user influence network. Then, the time series features output by the LSTM and the relationship features of the user influence network extracted by the GCN are input into the multi-head attention Transformer encoder to realize deep fusion of spatio-temporal features through the self-attention mechanism. The multi-scale feature extractor further captures feature representations of different granularities to ensure that the model can pay attention to local details and global patterns at the same time. The entire fusion process uses residual connection and layer normalization techniques to improve training stability, and finally outputs a high-dimensional fusion feature vector to provide a rich feature representation basis for subsequent game decision and evaluation analysis.

[0077] Figure 4 A fourth implementation flow diagram of the network hot review confrontation effectiveness adaptive evaluation method based on spatio-temporal perception provided by the embodiments of the present application is shown. In one embodiment, the three-party strategy network model includes a platform strategy network, a defender strategy network and an attacker strategy network, all of which use the Actor-Critic architecture. Before step S300, after the system is first deployed or reset, supervised pre-training is performed using historical network hot review data and corresponding confrontation behavior labels, or reinforcement learning pre-training is performed in a simulator similar to the real environment, so that the three-party strategy network model has basic strategy generation and game analysis capabilities. After formal online operation, the three-party strategy network model is trained and optimized online according to the real-time input fusion feature vector (generated by step S200). Step S300 includes:

[0078] S301, training the three-party strategy network model to be trained based on the fusion feature vector to generate platform strategy parameters corresponding to the platform strategy network, defense strategy parameters corresponding to the defender strategy network, and attack strategy parameters corresponding to the attacker strategy network. Illustratively, the platform strategy network, the defender strategy network, and the attacker strategy network share the environment feature vector and the mutual observable reward function during the training process. The reward function of the platform is:

[0079] ;

[0080] The reward function of the defender is:

[0081] ;

[0082] The reward function of the attacker is:

[0083] ;

[0084] wherein, is a preset weight coefficient.

[0085] Illustratively, the platform strategy parameters include recommendation weight, interaction gate, user satisfaction, and content quality loss; the defense strategy parameters include filtering threshold, trace comparison frequency, detection accuracy, false positive rate, and detection probability; and the attack strategy parameters include malicious comment rate, emotional manipulation intensity, and propagation effect.

[0086] S302, action space constraints are performed on the platform strategy parameters, the defense strategy parameters, and the attack strategy parameters to limit the range of continuous actions (i.e., weight adjustment actions), to obtain a strategy vector. Specifically, when the three-party strategy network model to be trained generates a balanced strategy in parallel, a weight sharing collaborative training method is adopted, the strategy network is updated through gradient back propagation, the training stability is ensured through soft update of the target network, and the strategy converges to equilibrium after 500,000 steps of training.

[0087] S303, inputting the strategy vector into a Nash equilibrium solving iterative optimization algorithm to obtain a candidate equilibrium strategy and a strategy change rate corresponding to the candidate equilibrium measurement. Specifically, the Nash equilibrium solving iterative optimization algorithm adopts an Adam optimizer, and the learning rate is set to 3e -4 , and combines an experience replay buffer to optimize the training process. Specifically, when solving the Nash equilibrium, the iterative optimization algorithm constructs a three-dimensional revenue matrix R(i,j,k), calculates the expected revenue of each party under the equilibrium strategy, and finds a stable strategy equilibrium point. i, j, and k represent the strategy index of the attacker, the defender, and the platform, respectively, and the matrix element R(i,j,k) represents the expected revenue value under the equilibrium strategy. The expected revenue calculation formula is: is a discount factor (value 0.99), For Instant rewards at the moment, the long-term income distribution under the equilibrium strategy is determined by the Monte Carlo sampling algorithm to find the stable strategy equilibrium point.

[0088] S304, if the strategy change rate is less than the preset threshold, it is determined that the to-be-trained three-party strategy network model reaches an equilibrium state, and a candidate equilibrium strategy and a candidate evaluation score corresponding to the candidate equilibrium strategy are output, and the candidate equilibrium strategy is taken as the three-party game equilibrium strategy, and the candidate evaluation score is taken as the evaluation score. Illustratively, when verifying the equilibrium by using the three-dimensional income matrix, it is judged whether the income of the three-party strategy satisfies the preset condition of Nash equilibrium; if yes, it is determined that the equilibrium state is reached; in the equilibrium state, the to-be-trained three-party strategy network model is trained to obtain the three-party strategy network model.

[0089] The method first performs environmental state feature perception, collects current network propagation state, user behavior characteristics, platform operation policy parameters and other information. Subsequently, three Actor-Critic networks respectively represent different participants, and generate their own strategy actions according to the current environmental feature vector state: the attacker strategy network outputs attack parameters such as malicious comment generation rate and emotional manipulation intensity; the defender strategy network generates defense parameter strategies such as content filtering threshold and abnormal detection sensitivity; the platform strategy network determines platform parameter strategies such as recommendation algorithm weight and user interaction gate. By constructing a three-dimensional income matrix, the expected income of each party under different strategy combinations is calculated, and a Nash equilibrium solving algorithm is used to find a stable strategy equilibrium point. In order to finally follow up, the strategy gradient optimization and experience replay mechanism are used to continuously improve the strategy network of each party, so that the system can adapt to the dynamically changing confrontation environment, and provide more accurate strategy prediction and decision support.

[0090] The method can deeply analyze the internal mechanism of network confrontation behavior from the perspective of game theory, predict the evolution trend of each party's strategy, and significantly improve the robustness of malicious manipulation behavior. The present application introduces three-party game theory to build a decision-making model closer to the real network environment, in which the attacker pursues the maximization of propagation effect, the defender tries to maintain information security, and the platform is committed to balancing user experience and content governance. Through the Actor-Critic algorithm deep reinforcement learning framework, the strategy network of each party is trained, and the system can learn and predict the optimal strategy selection of each participant in the dynamic game process, and then obtain a stable game solution through Nash equilibrium solving. This three-party game mechanism makes the system have stronger strategic foresight and confrontation adaptability, and can quickly adjust the defense strategy when facing new attack methods, effectively dealing with the complex and variable network confrontation environment.

[0091] Figure 5A fifth implementation flowchart of the network hot review perception-based adaptive evaluation method is shown. Specifically, the evaluation indicators include four dimensions of propagation performance, confrontation robustness, platform stability, and social influence. Before step S402, the parameters of the Actor policy network and the Critic value network of the three-party strategy network model to be trained are randomly initialized or pre-trained weights are loaded. The three-party strategy network model learns and optimizes online through the evaluation feedback generated by the interaction with the environment during system operation. Step S400 includes:

[0092] S401, based on the three-party game equilibrium strategy and the environment state feature, a 512-dimensional environment feature vector is constructed, wherein the 512-dimensional environment feature vector includes 128-dimensional game state features generated by the three-party strategy network model, 128-dimensional spatiotemporal features, 128-dimensional historical evaluation trends, and 128-dimensional system resource states. Among them, the game state feature is extracted from the current equilibrium strategy of the three-party strategy network model (step S300), including: the statistical quantity of the strategy parameter of each party (attacker, defender, platform party), the current value of the benefit function, etc., wherein the statistical quantity includes mean, variance. The spatiotemporal feature is directly selected or aggregated from the multi-scale spatiotemporal feature generated by the spatiotemporal perception acquisition module. The historical evaluation trend is generated from the pre-stored historical adaptive evaluation score sequence through a time series analysis method, wherein the time series analysis method includes calculating the difference, moving average, and trend extraction. The system resource state is obtained in real time from the monitoring interface of the server and the platform, including CPU / memory usage, network delay, service queue length, and other indicators.

[0093] Specifically, the statistical characteristics of the strategy vector and the benefit convergence indicators are extracted from the game equilibrium strategy to form the game state feature. Key indicators are selected from the multi-scale spatiotemporal feature, such as second-level emotional fluctuation variance, minute-level forwarding network centrality, and hour-level topic heat trend slope, to form the spatiotemporal feature. Feature engineering is performed on the historical adaptive evaluation score sequence, such as calculating the mean, variance, and trend of the recent score, to form the historical evaluation trend. The system monitoring indicators are aggregated to form the system resource state. The 512-dimensional feature data of the above four parts is spliced and normalized to form a unified 512-dimensional environment feature vector. Specifically, the 512-dimensional environment feature vector output by S401 is denoted as .

[0094] S402, input the 512-dimensional environment feature vector into the DDPG weight adjuster of the adaptive evaluation model, and output a weight adjustment action corresponding to the evaluation index through the Actor policy network corresponding to the three-party policy network model. Exemplarily, the Actor policy network adopts a three-layer fully connected structure, the input layer is a 512-dimensional environment feature vector, sequentially passes through a 256-dimensional hidden layer and a 128-dimensional hidden layer, and the output layer outputs a 4-dimensional weight adjustment action. The hidden layer of the Actor policy network adopts a ReLU activation function, and a Dropout of 0.2 is set to prevent overfitting. Before the output layer outputs the 4-dimensional weight adjustment action, Sigmoid normalization processing is performed. Specifically, the 512-dimensional environment feature vector is input into the DDPG weight adjuster of the adaptive evaluation model. The Actor policy network in the DDPG weight adjuster performs forward calculation according to the current environment feature vector, and outputs a 4-dimensional weight adjustment action. The weight adjustment action is a continuous vector, and each dimension corresponds to a preliminary adjustment amount of the four evaluation indexes. Before the weight adjustment action is output, Sigmoid function normalization processing is performed, so that the action value of the weight adjustment action is mapped to the interval [0, 1].

[0095] S403, evaluate the weight configuration Q value of the weight adjustment action through the Critic value network corresponding to the three-party policy network model, wherein the Q value is used to indicate the weight quality. The Critic value network simultaneously receives the 512-dimensional environment feature vector and the 4-dimensional weight adjustment action , and calculates a scalar Q value Q( , ) through the internal value function. The Q value evaluates the long-term expected return that can be brought by executing the weight adjustment action under the current environment feature vector , that is, the weight quality of the weight configuration.

[0096] S404, optimize the Actor policy network based on the Q value, and simultaneously perform constraint optimization processing on the weight adjustment action to obtain a new evaluation index weight. Exemplarily, based on the Q value calculated in S403, the deep deterministic policy gradient (DDPG) algorithm is used to calculate the policy gradient and back propagation to optimize the parameters of the Actor policy network. The optimization target is to enable the Actor network to output a weight adjustment action with a higher Q value in the future. The update formula is represented as: , wherein is a learning rate. is the updated parameter of the Actor policy network, These are the parameters before the Actor policy network is updated. To represent the Q-value of the Critic value network in relation to actions The gradient. This indicates that the output of the Actor policy network is relative to its own parameters. The gradient. Represents the Actor policy network, which is based on the state Output Action . Representing the Critic value network, its evaluation is in state Next action The value of.

[0097] For example, the constraint optimization process includes: calculating the initial adjustment weights based on the basic weight update formula; and adjusting the 4-dimensional weights of the Actor network output for the current period. The process is applied to generate new weights that can be directly used for evaluation. This process ensures that the weights meet the physical constraints of the practical application. The basic weight update formula is:

[0098]

[0099] The `clip` function is used to limit the magnitude of a single adjustment; the initial adjustment weights are optimized using the Lagrange multiplier method to ensure that normalization constraints are met. and range constraints This yields the weights of the new evaluation indicators. Specifically, For the first Initial adjusted weights for each evaluation indicator. For the first The weights of each evaluation indicator in the previous evaluation period; when the system runs for the first time, the preset initial weights are used. For the first The learning rate or step size control factor corresponding to each evaluation indicator. For the current moment 4D weight adjustment action of the Actor network output The first in Each component. For the pruning function, noise will be explored. Alternatively, the adjustment range can be limited to the interval [0.01, 0.1] to prevent excessively large single weight adjustments. Subsequently, constrained optimization algorithms such as the Lagrange multiplier method are used to adjust the initial weight vector. Optimize it to ensure that it meets the following constraints:

[0100]

[0101]

[0102] Finally, output the new evaluation index weight that meets all the constraints .

[0103] Exemplarily, the Critic value network adopts a state-action joint encoding structure, encodes a 512-dimensional environment feature vector through a state encoder (512→256→128), encodes a 4-dimensional weight adjustment action through an action encoder (4→16→32), inputs the encoding results into a value function network (160→128→64→1) after splicing, and outputs the Q value of the current weight configuration: Q(s t ,a t ), which is used to feedback and optimize the Actor policy network. The state encoder is a three-layer fully connected network composed of 512, 256, and 128 neurons, which is used to receive and encode state information. The action encoder is a three-layer fully connected network composed of 4, 16, and 32 neurons, which is used to receive and encode action information. The value evaluator is a four-layer fully connected network composed of 160, 128, 64, and 1 neurons.

[0104] Exemplarily, in the training process, the DDPG weight adjuster stores state-action-reward-next state data in an experience replay buffer with a capacity of 1 million samples; 256 sample batches are randomly sampled from the buffer each time for training, and the network parameters are updated using an exponential decay learning rate with an initial value of 1e -3 , and a decay rate of 0.95; the current network parameters are synchronized with the target network through a soft update coefficient to ensure training stability.

[0105] Specifically, in step S400, steps S401 to S404 are executed every 30 seconds, and a ring queue with a capacity of 100 is maintained to store historical environment feature state vectors and weight adjustment action information for subsequent trend analysis and anomaly detection.

[0106] The method is based on a deep deterministic policy gradient algorithm, and realizes intelligent optimization of evaluation index weight. At the beginning of the process, the system collects the current environmental state characteristics, including network propagation state, confrontation behavior intensity, platform running parameters and other multi-dimensional information, to form an environmental state feature vector. The DDPG weight adjuster adjusts the weight of the Actor network of the DDPG agent according to the environmental feature vector state vector to output the weight adjustment action of four evaluation dimensions, including the weight distribution of propagation efficiency, confrontation robustness, platform stability and social influence. Subsequently, the evaluation index of each dimension is calculated using the adjusted weight, and the evaluation score is obtained by weighted summation. The Critic network evaluates the pros and cons of the current weight configuration, and calculates the Q value function. Based on the evaluation effect feedback, the system calculates the reward signal, and stores the experience data into the replay buffer. Through batch sampling and gradient update, the Actor and Critic network parameters are continuously optimized. The whole process is executed in a 30-second evaluation cycle to ensure that the weight configuration can adapt to the network environment changes in time, and maintain the high precision and robustness of the evaluation model.

[0107] As shown in Figure 6 The sixth implementation process schematic diagram of the network hot evaluation confrontation efficiency adaptive evaluation method based on space-time perception. Step S500 includes:

[0108] S501, based on the game state information corresponding to the three-party game equilibrium strategy and the new evaluation index weight, the evaluation values of the four dimensions of propagation efficiency, confrontation robustness, platform stability and social influence are calculated respectively. Before step S501, based on the three-party game equilibrium strategy, the game state information corresponding to the three-party game equilibrium strategy is extracted. Exemplarily, the game state information includes attack situation indicators, defense efficiency indicators and platform regulation indicators. Specifically, the attack situation indicators include: in the attacker strategy network, the malicious comment quantity growth rate, the number of users affected by malicious comments, the average intensity of emotional manipulation operations, the number of suspected water army accounts put into use, etc. The defense efficiency indicators include: in the defender strategy network, the detection accuracy, the false positive rate (i.e. the false kill rate), the historical adaptive evaluation score sequence, the content filtering threshold, the coverage rate of known attack patterns, etc. The platform regulation indicators include: in the platform strategy network, system resource state, service response time data set, service level agreement (SLA) compliance rate, recommendation algorithm weight distribution, tightness of user interaction gate, diversion or weight reduction coefficient for specific topics, etc.

[0109] Specifically, the propagation efficiency evaluation value = propagation speed × coverage range × influence depth, wherein the propagation speed is obtained by first-order difference processing of the malicious comment quantity growth rate; the coverage range is obtained by logarithmic transformation processing of the number of users affected by malicious comments; and the influence depth is obtained by weighted summation based on the number of forwarding levels obtained from the user influence network.

[0110] Specifically, the anti-robustness evaluation value = detection accuracy x (1-false positive rate) x recovery capability is calculated, the detection accuracy is expressed in the form of F1 score; the recovery capability is evaluated by an exponential decay model, the input of the exponential decay model is: the decline amplitude caused by the attack in the historical adaptive evaluation score sequence, the recovery speed.

[0111] Specifically, the platform stability evaluation value = (1-system load) x response speed x availability is calculated, the system load is obtained by weighted average of CPU / memory usage in system resource status; the response speed is calculated by taking the 95th percentile of the service response time data set; the availability is determined based on the platform service level agreement (SLA) compliance rate.

[0112] Specifically, the social influence evaluation value = user satisfaction x content quality x social harmony is calculated. The user satisfaction is measured by the proportion of positive evaluations in sentiment analysis. The user sentiment distribution containing the sentiment polarity label of each comment is extracted from the multi-scale spatio-temporal features, which has been processed by a pre-trained sentiment classification model; the proportion of positive comments in the total number of comments in the user sentiment distribution is obtained as the user satisfaction by statistical analysis. The content quality (i.e., content quality loss) adopts a composite index of BLEU score and perplexity. The BLEU score is obtained by comparing the comment text in the comment content data set after text preprocessing with a pre-set high-quality reference review corpus, and calculating the n-gram accuracy. The perplexity value is obtained by inputting the comment content data set into a pre-trained language model to calculate the prediction uncertainty of the pre-trained language model for the comment text sequence. The content quality score = α*BLEU score + β*(1 / perplexity) is obtained by linear combination of the BLEU score and the reciprocal of the perplexity through a weighted fusion algorithm, where α and β are pre-set weight coefficients. The user sentiment distribution is extracted from the multi-scale spatio-temporal features, and the sentiment intensity of each comment in the user sentiment distribution is represented in the form of continuous numerical value; the topic polarization degree is obtained by calculating the statistical variance of all comment sentiment intensity values; the social harmony is obtained based on the reciprocal of the topic polarization degree.

[0113] S502, the evaluation values of the four dimensions are weighted and fused with the evaluation scores to obtain real-time comprehensive evaluation scores. Exemplarily, the weighted fusion formula is:

[0114]

[0115] wherein, is the weight of the new evaluation index, are evaluation values corresponding to the propagation efficiency, the adversarial robustness, the platform stability, and the social influence, respectively; is a real-time comprehensive evaluation score. Illustratively, are evaluation values corresponding to the propagation efficiency, the adversarial robustness, the platform stability, and the social influence, respectively; are new evaluation index weights corresponding to the propagation efficiency, the adversarial robustness, the platform stability, and the social influence, respectively.

[0116] S503, the exponential moving average algorithm is used to smooth the real-time comprehensive evaluation score to obtain an adaptive evaluation score. Specifically, the real-time comprehensive evaluation score of the current time t is obtained from step S502 , which is an original evaluation result calculated by weighted fusion based on the new evaluation index weights and the four-dimensional evaluation values . The adaptive evaluation score of the last evaluation period (time t-1) is read from the system cache or storage , which is used to indicate the evaluation result after historical data smoothing. The adaptive evaluation score of the current time t is calculated according to the exponential moving average algorithm formula:

[0117]

[0118] , wherein, is the real-time comprehensive evaluation score, is the adaptive evaluation score of the current time t. is the adaptive evaluation score of the last evaluation period (i.e., time t-1), which is used to indicate the stable evaluation state of the three-party strategy network model at the last time. The smoothing coefficient is used to control the weight proportion of the adaptive evaluation score of the last evaluation period and the adaptive evaluation score of the current time. The adaptive evaluation score calculated at the current time is stored in the system cache as the historical adaptive evaluation score of the next evaluation period (time t+1), realizing continuous updating of the evaluation state. The adaptive evaluation score of the current time t calculated is taken as the final evaluation result at the current time, i.e., the adaptive evaluation score. In the first evaluation period of the system (t=1), since there is no historical data, the initial value may be set as a preset value (such as 0.8), or the real-time comprehensive evaluation score calculated in the first period is directly used for assignment. Starting from the second evaluation period (t≥2), is the adaptive evaluation score calculated and stored in the last period (t-1).

[0119] The method can dynamically optimize the weight configuration of the evaluation model according to the characteristics of different network environments and application scenarios, and ensure that the evaluation results can maintain high accuracy and reliability under various conditions. Traditional evaluation methods usually use a fixed weight evaluation index system, which is difficult to adapt to the rapid changes of network environment and the continuous evolution of countermeasures. Based on the deep deterministic policy gradient algorithm, the adaptive weight adjustment mechanism is constructed, which includes four dimensions of propagation efficiency, counter-robustness, platform stability and social influence. The system can complete intelligent adjustment of the weight within 30 seconds of the evaluation period by continuously monitoring the network state changes, analyzing the historical evaluation effect and learning the optimal weight combination. When the network environment changes significantly, such as the emergence of new attack patterns or the migration of propagation characteristics, the system can automatically identify these changes and adjust the importance weight of each evaluation dimension accordingly, ensuring that the evaluation model always matches the current network environment, thereby maintaining the accuracy and timeliness of the evaluation results.

[0120] As shown in Figure 7 the seventh implementation process schematic diagram of the network hot evaluation countermeasure efficiency adaptive evaluation method based on space-time perception. After obtaining the adaptive evaluation score, it also includes:

[0121] S701, input the adaptive evaluation score into the risk early warning and intervention module, and perform multi-layer risk threshold judgment based on a statistical learning method to obtain a risk level;

[0122] S702, according to the risk level, execute the corresponding grading intervention strategy;

[0123] S703, record the early warning log and the intervention log, evaluate the effectiveness of the intervention measures and feedback optimization.

[0124] In one embodiment, step S701 comprises:

[0125] Compare the adaptive evaluation score with the preset multi-layer risk threshold (safe (>0.8), low risk (0.6-0.8), medium risk (0.4-0.6), high risk (0.2-0.4), and extremely high risk (<0.2)) one by one;

[0126] According to the comparison result, determine the risk level to which the adaptive evaluation score belongs.

[0127] In one embodiment, step S702 comprises:

[0128] If the risk level is low risk (0.6-0.8), execute level 1 intervention, notify the platform to carry out content filtering enhancement and user reminder operation;

[0129] If the risk level is medium risk (0.4-0.6), execute level 2 intervention, notify the platform to implement account restriction and content unloading operation;

[0130] If the risk level is high risk (0.2-0.4), perform level 3 intervention, notify the platform to perform account blocking, emergency response operation;

[0131] If the risk level is extremely high risk (<0.2), perform level 4 intervention, notify the platform to perform system locking, regulatory department notification operation.

[0132] In one embodiment, step S703 comprises:

[0133] Record the trigger time, risk level, adaptive assessment score and other information of the early warning event to the early warning log;

[0134] Record the execution time, intervention level, specific operation and other information of the intervention measure to the intervention log;

[0135] Compare and analyze the propagation situation of network hot comments and the change of confrontation behavior before and after intervention, and evaluate the effectiveness of the intervention measure;

[0136] The evaluation results are fed back to the adaptive assessment model and the three-party strategy network model to optimize the subsequent assessment and intervention strategy.

[0137] Illustratively, the risk early warning and intervention module has millisecond-level risk identification capability and second-level intervention response speed.

[0138] Illustratively, the preset multi-layer risk threshold can be optimized and adjusted according to historical data distribution.

[0139] Illustratively, the execution of the hierarchical intervention strategy is realized through the interface interaction with the platform to ensure the timely communication of the intervention instruction.

[0140] The method adopts a multi-level risk detection system, and divides the risk level into five levels according to the adaptive assessment score: safe, low risk, medium risk, high risk and extremely high risk. When the assessment score is lower than the safe threshold, the system immediately triggers the intervention measure of the corresponding level. The low risk level performs mild intervention, such as content filtering enhancement and user reminder; the medium risk level implements moderate intervention, including account restriction and content unloading; the high risk level starts severe intervention, such as account blocking and emergency response; the extremely high risk level performs the highest level intervention, including system locking and regulatory notification. The whole response process has millisecond-level risk identification capability and second-level intervention response speed. At the same time, the module establishes a perfect log recording system, which records the early warning event, intervention measure and effect tracking data in real time, and evaluates the effectiveness of the intervention measure through comparison and analysis, provides data support for strategy optimization, and forms a closed-loop risk management system.

[0141] For example, Figure 8As shown, the eighth implementation flow diagram of the network thermal evaluation of spatiotemporal perception-based network thermal evaluation of adaptive performance countermeasure is shown. In an embodiment, after obtaining the adaptive evaluation score, the adaptive evaluation score and related system running data are input into the visualization interface module, and based on the Web dashboard architecture, the multi-dimensional data visualization display and interactive operation are realized through the interaction between the front-end display layer and the back-end processing layer.

[0142] S800, the adaptive evaluation score and the related system running data are input into the visualization interface module, and based on the Web dashboard architecture, the multi-dimensional data visualization display and interactive operation are realized through the interaction between the front-end display layer and the back-end processing layer.

[0143] Exemplarily, step S800 includes:

[0144] S801, the front-end display layer utilizes HTML5, JavaScript and ECharts chart library to build user interface, and the back-end processing layer provides API interface and data service through the 5001 port of FlaskWeb service;

[0145] S802, the real-time data communication between the front-end display layer and the back-end processing layer is realized through SocketIO, and the adaptive evaluation score and the related system running data are pushed to the front-end display layer. Specifically, the real-time data communication realized by SocketIO can ensure the timely update of the front-end display layer data, and the delay is not more than the preset time threshold.

[0146] S803, the front-end display layer displays the adaptive evaluation score real-time curve, four-dimensional index radar chart, risk level state indicator, historical risk trend chart, three-party game strategy evolution curve and Nash equilibrium convergence process visualization based on the received data, and provides interactive operation. The interactive operation includes threshold parameter online adjustment, evaluation report and historical data export, system start / stop, chart scaling, system control start / stop operation, etc.

[0147] Exemplarily, the visualization interface module adopts data caching, incremental data updating and optimization of large data volume chart rendering performance strategy to ensure smooth user experience. Specifically, the data caching adopts Redis caching technology to cache the high-frequency access data. Incremental data updating refers to transmitting only the changed part of the data, reducing the data transmission volume.

[0148] The method adopts a front-end and back-end separation design pattern. The front-end is based on HTML5, JavaScript and ECharts chart library to build the user interface, and the back-end provides API interface and data service through Flask Web service. Real-time data communication between the front-end and the back-end is realized through SocketIO to ensure timely updating of interface data. The module provides rich data visualization functions, including real-time curve chart of adaptive evaluation score, radar chart of four-dimensional evaluation index, risk level state indicator, historical risk trend analysis chart, three-party game strategy evolution curve, Nash equilibrium convergence process visualization, etc. The system operation status monitoring panel displays the working status, data processing performance, resource occupation and other key indicators of each module in real time. The interactive function supports online adjustment of threshold parameters, export of evaluation report, historical data query, system control operation, etc. To ensure user experience, the module adopts data caching, incremental updating, asynchronous rendering and other performance optimization strategies to ensure smooth interface response in large data scenarios and provide users with intuitive and efficient system monitoring and management experience.

[0149] A network hot review confrontation effectiveness adaptive evaluation method based on space-time perception is provided in the present application. The specific implementation scheme of the method is as follows: implementation scenario one: confrontation effectiveness evaluation of sudden public opinion events. Taking the network hot review propagation of a sudden public event as an example, it is assumed that a controversial topic about the event appears on a social platform at T=0. The system operation process is as follows:

[0150] In the space-time perception collection stage, the system captures the abnormal surge of the number of comments within T+1s by means of 1-second sampling: from an average of 5 comments per second to a rapid rise to 200 comments per second; through 60-second sampling, it is found that the forwarding propagation presents a network diffusion trend, involving 12 cross-community transmission paths; and 3600-second sampling accurately identifies that the topic evolution direction is deviating towards emotional polarization.

[0151] After entering the feature fusion processing stage, the bidirectional LSTM network successfully captures the time sequence features of the comment sentiment: from neutral (0.1) to negative (-0.8) quickly; the graph convolution network identifies 3 key transmission nodes, and their influence scores are 0.95, 0.87 and 0.73 respectively; then, the Transformer encoder fuses the time sequence and spatial features, and outputs a 128-dimensional feature vector.

[0152] In the multi-party game decision-making stage, the action vector output by the attacker strategy network is [0.8, 0.6, 0.2], corresponding to sentiment amplification, false information implantation, and water army account deployment, respectively; the response vector of the defender strategy network is [0.9, 0.7, 0.8], covering content review enhancement, abnormal account detection, and transmission restriction; and the platform strategy vector is [0.6, 0.4, 0.9], including user reminders, content weight reduction, and topic guidance. After about 150,000 steps of training, the strategies of the three parties converge to a Nash equilibrium state.

[0153] In the adaptive evaluation stage, the initial weight is w = [0.3, 0.3, 0.2, 0.2], and the DDPG weight adjuster adjusts it to w' = [0.4, 0.4, 0.15, 0.05] based on the current environmental feature vector, so as to increase the weight proportion of transmission efficiency and anti-robustness. The adaptive evaluation score E_t = 0.65 is lower than the threshold 0.8, triggering a medium-risk warning.

[0154] In the risk warning and intervention stage, the system automatically triggers a level 2 intervention measure at T+180s, notifying the platform to perform account restriction and content removal operations. Within 30 minutes after the intervention, the transmission speed drops to the normal level (20 per second), and the efficiency evaluation score rises to 0.82, fully verifying the effectiveness of the intervention measures.

[0155] The application provides a network hot evaluation countermeasure efficiency adaptive evaluation method based on space-time perception. The specific implementation scheme of the method is as follows: implementation scenario two: countermeasure efficiency evaluation of long-term network public opinion manipulation. Taking long-term public opinion manipulation in product marketing as an example, the attacker uses a persistent and concealed strategy to guide positive public opinion, and the system runs as follows:

[0156] In the space-time perception collection stage, within the 30-day observation period, the system finds that the comment activity level in the fixed time period (9:00-11:00, 14:00-16:00, 19:00-21:00) increases slightly but regularly through second-level sampling; minute-level sampling identifies that the forwarding path presents obvious centralization characteristics; and hour-level sampling shows that the topic heat presents a periodic fluctuation pattern.

[0157] In the feature fusion processing stage, time series analysis shows that the sentiment value gradually increases from 0.2 to 0.7 within the observation period; spatial analysis finds that there are 15 high-activity accounts forming a core transmission network, and the correlation between these accounts is abnormally high, with an average correlation coefficient of 0.85; and the fused features clearly show that the transmission mode has obvious signs of artificial intervention.

[0158] In the multi-party game decision phase, the attacker adopts a moderate penetration strategy, and the action vector is [0.3, 0.1, 0.6], which corresponds to mild emotional guidance, a small amount of positive content implantation, and a large number of water army accounts for long-term delivery. The defender's strategy network responds weakly in the early stage, with a vector of [0.2, 0.3, 0.1], and gradually increases to [0.7, 0.8, 0.6] as the game evolves. The platform strategy is also adjusted to [0.8, 0.6, 0.7]. After about 250,000 steps of training, the three parties reach a new equilibrium state.

[0159] In the adaptive evaluation phase, the weight experienced 3 significant adjustments within 30 days, and finally stabilized at w' = [0.25, 0.45, 0.2, 0.1], highlighting the importance of adversarial robustness. The adaptive evaluation score showed a downward trend, from 0.9 at the beginning to 0.7, triggering a low-risk warning at the 20th day, and upgrading to a medium-risk warning at the 28th day.

[0160] In the risk warning and intervention phase, the system adopts a gradual intervention strategy: level 1 intervention (content filtering enhancement) is implemented on the 20th day, and level 2 intervention (account restriction) is upgraded on the 28th day. After 15 days of continuous monitoring, 12 of the 15 core propagation accounts were successfully identified and restricted, the topic propagation returned to a natural state, the evaluation score rose to 0.85, and long-term adversarial behavior was effectively curbed.

[0161] The network hot review confrontation efficiency adaptive evaluation method based on spatiotemporal perception described in the application can realize multi-scale accurate capture of network hot review propagation dynamics by constructing a three-layer spatiotemporal perception network architecture of seconds, minutes, and hours. Traditional evaluation systems often use a single time window for data collection, making it difficult to balance the rapid response to short-term emergencies and the in-depth analysis of long-term propagation trends. The present application captures instantaneous propagation bursts through 1-second high-frequency sampling, traces propagation diffusion processes through 60-second medium-frequency sampling, and analyzes long-term evolution trends through 3600-second low-frequency sampling, forming a stereoscopic perception network covering different time scales. This three-level spatiotemporal perception mechanism not only can detect the early stages of abnormal propagation patterns, but also can accurately identify the evolution trajectory of persistent adversarial behavior, achieving all-around, multi-level accurate capture of propagation fine-grained dynamics, and providing a high-quality data foundation for subsequent game decision and risk assessment.

[0162] Based on the innovative design of the attacker-defender-platform three-party game decision network, the system can deeply analyze the internal mechanism of network confrontation behavior from the perspective of game theory, predict the evolution trend of each party's strategy, and significantly improve the robustness to malicious manipulation behavior. The existing technology mainly adopts a binary confrontation model, only considering the game relationship between attack and defense, ignoring the unique role and interest demand of the platform as an important participant. The invention introduces the three-party game theory to build a decision model closer to the real network environment, in which the attacker pursues the maximization of propagation effect, the defender tries to maintain information security, and the platform is committed to balancing user experience and content governance. Through the Actor-Critic algorithm deep reinforcement learning framework to train each party's strategy network, the system can learn and predict the optimal strategy selection of each participant in the dynamic game process, and then obtain a stable game solution through Nash equilibrium solution. This three-party game mechanism makes the system have stronger strategic foresight and confrontation adaptability, which can quickly adjust the defense strategy when facing new attack methods, and effectively cope with the complex and variable network confrontation environment.

[0163] Through the introduction of the adaptive evaluation index weight adjustment mechanism, the system can dynamically optimize the weight configuration of the evaluation model according to the characteristics of different network environments and application scenarios, ensuring that the evaluation results can maintain high accuracy and reliability under various conditions. Traditional evaluation methods usually use a fixed weight evaluation index system, which is difficult to adapt to the rapid changes of network environment and the continuous evolution of confrontation strategies. Based on the deep deterministic policy gradient (DDPG) algorithm, the invention constructs an adaptive weight adjustment mechanism containing four dimensions of propagation efficiency, confrontation robustness, platform stability and social influence. The system can complete intelligent adjustment of weights within 30 seconds of evaluation period by continuously monitoring network state changes, analyzing historical evaluation effects and learning optimal weight combinations. When the network environment changes significantly, such as the emergence of new attack patterns or the migration of propagation characteristics, the system can automatically identify these changes and adjust the importance weight of each evaluation dimension accordingly, ensuring that the evaluation model always matches the current network environment, thereby maintaining the accuracy and timeliness of the evaluation results.

[0164] Compared to traditional methods for evaluating the effectiveness of network adversarial attacks, this system achieves significant improvements across several key performance indicators. Regarding evaluation accuracy, the synergistic effect of technological innovations such as spatiotemporal awareness multi-scale feature fusion, three-party game decision modeling, and adaptive weight optimization greatly enhances the overall evaluation accuracy compared to existing technologies, particularly in complex network environments and hybrid adversarial scenarios. In terms of detecting novel adversarial attacks, thanks to the strategic prediction capabilities of the game theory model and the refined monitoring of the multi-scale spatiotemporal awareness network, the system achieves a high detection rate for unknown attack patterns, significantly improving upon traditional rule-based or single machine learning model-based methods. Regarding system response performance, through optimized distributed architecture design, efficient feature extraction algorithms, and intelligent caching mechanisms, the system can complete the entire evaluation and analysis process and output results within 3 seconds of receiving network data, significantly reducing the average response time compared to existing systems. This truly achieves near real-time network security situational awareness and risk warning, providing more timely, accurate, and reliable technical support for network security protection.

[0165] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0166] Corresponding to the method in the above embodiments, Figure 8 The diagram shows a structural block diagram of the adaptive evaluation system for network heat assessment adversarial effectiveness based on spatiotemporal awareness provided in the embodiments of this application. For ease of explanation, only the parts related to the embodiments of this application are shown. Figure 8 The example of the spatiotemporally aware network hot rating adversarial effectiveness adaptive evaluation system can be the execution subject of the spatiotemporally aware network hot rating adversarial effectiveness adaptive evaluation method provided in the aforementioned embodiment 1.

[0167] Reference Figure 8 The spatiotemporal awareness-based adaptive evaluation system for network heat map adversarial effectiveness includes:

[0168] The spatiotemporal sensing acquisition module is used to input multi-source network data and generate multi-scale spatiotemporal features;

[0169] The feature deep fusion processing module is used to generate fused feature vectors based on multi-scale spatiotemporal features;

[0170] The multi-party game decision-making module is used to train the three-party policy network model to be trained based on the fused feature vector, and to determine the equilibrium strategy of the three-party game and the evaluation score corresponding to the equilibrium strategy.

[0171] The adaptive evaluation score determination module is configured to adjust the evaluation indexes based on the environmental state features through an adaptive evaluation model to obtain new evaluation index weights, and determine an adaptive evaluation score based on the new evaluation index weights and the evaluation score.

[0172] Figure 8 The overall architecture of the network hot review confrontation effectiveness adaptive evaluation system based on the spatio-temporal perception multi-scale game network is shown. The system adopts a hierarchical design, the bottom layer is a data infrastructure layer, including Kafka streaming middleware responsible for multi-source social data access, Redis providing cache service, HBase distributed storage supporting massive data management. The middle layer is the core algorithm processing layer, which is composed of six functional modules: the spatio-temporal perception acquisition module realizes multi-scale data acquisition, the feature depth fusion processing module completes spatio-temporal feature extraction and fusion, the multi-party game decision module analyzes the strategy based on the three-party game theory, the adaptive evaluation module dynamically adjusts the evaluation weight, the risk early warning and intervention module provides security protection, and the visualization interface module supports user interaction. The top layer is the application service layer, which is deployed through cloud server cluster, uses TensorRT to accelerate deep learning model inference, and provides real-time evaluation, risk early warning, strategy analysis and other services for users. The whole architecture realizes the organic unification of data flow, control flow and feedback flow, ensures the efficiency, real-time performance and scalability of the system.

[0173] In one embodiment, the spatio-temporal perception-based network hot review confrontation effectiveness adaptive evaluation system further comprises: a risk early warning and intervention module for multi-layer risk detection based on the adaptive evaluation score, determining the risk level and executing the hierarchical intervention strategy. And a visualization interface module for visualizing and interacting with the adaptive evaluation score, risk warning information and system running state.

[0174] In one embodiment, the spatio-temporal perception acquisition module comprises: a multi-scale sampling unit for sampling second-level comments, minute-level forwarding and hour-level topics through 1-second, 60-second and 3600-second time windows respectively. A data preprocessing unit for text formatting based on regular expressions, sensitive word filtering based on a vocabulary, anomaly detection based on statistical learning, and semantic analysis through part-of-speech tagging, named entity recognition and sentiment polarity analysis. And a distributed storage unit for shunting storage and persistence of processed data through Kafka message queue and HBase partition table, and caching hot data through Redis three-layer cache architecture.

[0175] In an embodiment, the feature deep fusion processing module includes: a time sequence feature extraction unit for processing the comment time sequence using a double-layer bidirectional LSTM network to capture short-term memory and long-term dependence; a spatial structure modeling unit for constructing a propagation graph based on a user influence network and extracting spatial structure features using a graph convolution network; and a multi-modal fusion unit for attention fusion of the LSTM output and the graph convolution embedding through a multi-head Transformer encoder to output a multi-scale spatio-temporal feature tensor as a fusion feature vector.

[0176] In an embodiment, the multi-party game decision module includes: a strategy network training unit for establishing three-party strategy network models with an Actor-Critic architecture for the attacker, the defender, and the platform party respectively, and performing collaborative training based on shared environment information and mutually observable reward functions; and an equilibrium solving unit for solving the Nash equilibrium of the three-party strategies through an iterative optimization algorithm, determining that an equilibrium state is reached when the strategy change rate is less than 0.01, and outputting the three-party optimal strategy and the equilibrium strategy evaluation score.

[0177] In an embodiment, the adaptive evaluation module includes: a weight adjustment unit for adjusting the action using the output weight of the Actor network of the DDPG algorithm, and obtaining the new evaluation index weight after Sigmoid normalization and constraint optimization. A comprehensive evaluation unit for calculating an adaptive evaluation score based on the new evaluation index weight and the evaluation value of each dimension, smoothing short-term fluctuations through an exponential moving average algorithm, and triggering a risk warning based on a dynamic threshold.

[0178] In an embodiment, the risk warning and intervention module includes: a risk detection unit for setting multi-layer risk thresholds based on statistical learning methods, and dividing the risk level into safe, low risk, medium risk, high risk, and extremely high risk. A hierarchical intervention unit for executing corresponding intervention strategies according to different risk levels, including content filtering enhancement, account restriction, account ban, system lock, etc. And a log tracking unit for recording warning logs and intervention logs, and analyzing the effects before and after intervention to evaluate the effectiveness of the measures.

[0179] In an embodiment, the visualization interface module includes: a data display unit for displaying the adaptive evaluation score real-time curve, the four-dimensional index radar chart, the risk level state indicator, and the three-party game strategy evolution curve through a Web dashboard. An interactive operation unit for supporting online adjustment of threshold parameters, export of evaluation reports and historical data, system start-stop control, and chart scaling, etc. And a performance optimization unit for optimizing the performance of large data volume chart rendering using data caching, incremental data updating, etc.

[0180] The application aims to provide a network hot review confrontation performance adaptive evaluation system based on space-time perception. Firstly, a three-level time perception system covering seconds, minutes and hours and a cross-community space propagation path modeling mechanism are proposed to synchronously capture the dynamic characteristics of network hot reviews in instantaneous reaction, diffusion process and long-term trend. Through multi-source data real-time collection and hierarchical sampling strategy, user behavior data in different time windows are accurately recorded, and combined with spatial topology analysis, comprehensive space-time dimension support is provided for confrontation performance evaluation, solving the problem of single-scale and single-dimensional monitoring in traditional technology. Secondly, a "attacker-defender-platform" three-party game model is introduced, and the strategy network of each party is built through the Actor-Critic framework to realize dynamic simulation and strategy evolution prediction of the interactive behavior of the three parties. The strategy networks of the three parties share environmental information and observe the reward function, and converge to Nash equilibrium through collaborative training, effectively predicting the development trend and performance change of the confrontation behavior, making up for the defects of existing technology in multi-party game modeling. Thirdly, aiming at the problem that the evaluation index system is fixed and cannot be adjusted in real time with the change of the environment, a four-dimensional evaluation system including propagation performance, confrontation robustness, platform stability and social influence is established, and DDPG weight adjustment algorithm is introduced to realize dynamic optimization of evaluation weight. The system can automatically adjust the weight of each evaluation dimension according to the real-time network environment characteristics, ensuring the accuracy and adaptability of the evaluation result in different scenarios, overcoming the limitations of single index and fixed weight in traditional evaluation methods. In addition, the risk warning and intervention module and the visualization interface module are integrated to form a complete closed loop from data collection, feature processing, game decision, performance evaluation to risk intervention and visualization display. The risk warning and intervention module can trigger the corresponding level of intervention measures in real time according to the evaluation result to realize rapid response and disposal of malicious confrontation behavior, and the visualization interface module can display the system running state, evaluation result and strategy evolution process through intuitive charts to provide convenient interactive operation and decision support for users.

[0181] Through the innovative design of multi-scale space-time perception, three-party game modeling and adaptive evaluation weight adjustment, the application can comprehensively improve the real-time performance, accuracy and adaptability of network hot review confrontation performance evaluation, effectively cope with various confrontation behaviors in complex network environment, and provide strong technical support for scientific governance of network public opinion. The method can significantly improve the adaptability, real-time performance and accuracy of the network hot review confrontation performance adaptive evaluation system based on space-time perception in complex dynamic confrontation scenarios.

[0182] The process of each module in the network hot review confrontation performance adaptive evaluation system based on space-time perception provided by the embodiments of the application to realize its own function can be specifically referred to the description of the first embodiment of the foregoing Figure 1 application, which will not be described here in detail.

[0183] It will be understood that the term "includes", "comprises", "comprising", "has", "having", "includes" and / or "including" when used in this specification and the following claims specifies the presence of the stated features, integers, steps, operations, elements, and / or components but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0184] It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items, and that the term "at least one of' denotes one, or, optionally, more than one.

[0185] As used in this specification and the appended claims, the term "if' can be construed to mean "when" or "once" or "in response to determining" or "in response to detecting" depending on the context. Similarly, the phrase "if it is determined" or "if [a described condition or event] is detected" can be construed to mean "once it is determined" or "in response to determining" or "once [the described condition or event] is detected" or "in response to detecting [the described condition or event]," depending on the context.

[0186] In addition, the terms "first", "second", "third", etc. as used in the description of the application and the appended claims are only used to distinguish between different descriptions, and cannot be understood as indicating or implying relative importance. It should also be understood that although the terms "first", "second", etc. are used in the text to describe various elements in some embodiments of the application, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, the first table can be named as the second table, and similarly, the second table can be named as the first table, without departing from the scope of various described embodiments. The first table and the second table are both tables, but they are not the same table.

[0187] Reference in the specification to "one embodiment" or "some embodiments" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. The appearances of the phrase "in one embodiment", "in some embodiments", "in other embodiments", "in additional embodiments", and so on, in various places in the specification are not necessarily all referring to the same embodiment, unless otherwise specifically stated. The terms "comprise", "comprising", "has", "having", "includes" and "including" are meant to be open-ended terms that specifically permit the inclusion of one or more of the listed items, but not exclusion of any of the items that are specifically listed.

[0188] Figure 9 is a structural schematic diagram of a terminal device provided by an embodiment of the application. As shown in the figure, the terminal device comprises a processor 1001, a memory 1002 and a communication interface 1003. Figure 9As shown, the terminal device 9 of this embodiment includes: at least one processor 90 ( Figure 9 Only one is shown in the image), and a memory 91 is stored in which a computer program 92 can be run on the processor 90. When the processor 90 executes the computer program 92, it implements the steps in the above-described embodiments of the adaptive evaluation method for network heat assessment adversarial effectiveness based on spatiotemporal awareness, for example... Figure 1 Steps 100 to 500 are shown. Alternatively, when the processor 90 executes the computer program 92, it implements the functions of each module / unit in the above-described device embodiments.

[0189] The terminal device 9 can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor 90 and a memory 91. Those skilled in the art will understand that... Figure 9 This is merely an example of terminal device 9 and does not constitute a limitation on terminal device 9. It may include more or fewer components than shown, or combine certain components, or different components. For example, the terminal device may also include input transmission devices, network access devices, buses, etc.

[0190] The processor 90 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0191] The memory 91 may, in some embodiments, be an internal storage unit of the terminal device 9, such as a hard disk or a memory of the terminal device 9. The memory 91 may also be an external storage device of the terminal device 9, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. equipped on the terminal device 9. Further, the memory 91 may also include both an internal storage unit and an external storage device of the terminal device 9. The memory 91 is used to store an operating system, an application program, a BootLoader, data, and other programs, etc., such as program codes of the computer program, etc. The memory 91 may also be used to temporarily store data that has been transmitted or is to be transmitted.

[0192] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0193] The embodiments of the present application further provide a terminal device, which comprises at least one memory, at least one processor, and a computer program stored in the at least one memory and executable on the at least one processor, and the processor executes the computer program to enable the terminal device to implement the steps in any of the above method embodiments.

[0194] The embodiments of the present application further provide a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps in any of the above method embodiments.

[0195] The embodiments of the present application provide a computer program product, which, when executed on a terminal device, enables the terminal device to implement the steps in any of the above method embodiments.

[0196] The integrated modules / units, if implemented in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. The computer program can implement the steps of each method embodiment when executed by a processor. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms. The computer readable medium can include any entity or device, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. that can carry the computer program code.

[0197] In the above embodiments, the description of each embodiment has its own focus, and the parts not described or recorded in detail in a certain embodiment can be referred to the related description of other embodiments.

[0198] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0199] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, i.e. they can be located in one place or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiment according to actual needs.

[0200] The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. An adaptive evaluation method for the adversarial effectiveness of network hot topic evaluation based on spatiotemporal awareness, characterized in that, include: Multi-source network data is input into a multi-scale spatiotemporal data fusion model to generate multi-scale spatiotemporal features; Based on the aforementioned multi-scale spatiotemporal features, a fused feature vector is generated through a feature deep fusion processing model. The three-party policy network model to be trained is trained based on the fused feature vectors to determine the equilibrium strategy of the three-party game and the evaluation score corresponding to the equilibrium strategy of the three-party game. The three-party strategy network model includes a platform strategy network, a defender strategy network, and an attacker strategy network, all of which adopt the Actor-Critic architecture. Based on environmental characteristics, the evaluation indicators are adjusted through an adaptive evaluation model to obtain new evaluation indicator weights. The evaluation indicators include four dimensions: dissemination effectiveness, resilience against attacks, platform stability, and social impact. The process of adjusting the evaluation indicators based on environmental state characteristics using an adaptive evaluation model to obtain new evaluation indicator weights includes: Based on the three-party game equilibrium strategy and the environmental state characteristics, a 512-dimensional environmental feature vector is constructed, wherein the 512-dimensional environmental feature vector includes 128-dimensional game state characteristics generated by the three-party strategy network model, 128-dimensional spatiotemporal characteristics, 128-dimensional historical evaluation trends, and 128-dimensional system resource status. Based on the weights of the new evaluation indicators and the evaluation scores, an adaptive evaluation score is determined; The step of inputting multi-source network data into a multi-scale spatiotemporal data fusion model to generate multi-scale spatiotemporal features includes: The multi-source network data is sampled according to a time window to obtain sampled data; The sampled data is processed in parallel using a real-time streaming data processing algorithm to obtain temporary data. A user influence network is constructed based on the temporary data, and node attribute features are generated. The temporary data is distributed and stored to generate multi-scale time stamps; Based on the node attribute features and the multi-scale time stamp, spatiotemporal features are fused to generate the multi-scale spatiotemporal features; The time window includes an hourly time window, a minute-level time window, and a second-level time window. The sampled data includes hourly topic sampled data collected corresponding to the hourly time window, minute-level forwarding sampled data collected corresponding to the minute-level time window, and second-level comment sampled data collected corresponding to the second-level time window. The process of constructing a user influence network based on the temporary data and generating node attribute features includes: Extract the target user's basic user characteristics and the interaction relationships between the target user and other users from the temporary data; The user's basic characteristics are input into the improved PageRank algorithm model for iterative calculation to obtain the first user influence value of each target user. Determine the time interval between the occurrence time of each target user's behavior and the current processing time, calculate the time decay factor based on the time interval, and use the time decay factor to dynamically correct the first user influence value of the corresponding target user to obtain the second user influence value of each target user. The user influence network is constructed by using each target user as a node, the interaction relationship between users as directed edges, the second user influence value of the corresponding target user as the initial weight of the node, and the strength of the interaction relationship as the edge weight. Based on the user influence network, multi-dimensional features of each user node are extracted and integrated to generate node attribute features of each user node.

2. The adaptive evaluation method for network heat rating adversarial effectiveness based on spatiotemporal awareness as described in claim 1, characterized in that, The process of generating a fused feature vector based on the multi-scale spatiotemporal features through a feature deep fusion processing model includes: The multi-scale spatiotemporal features are input into the text feature extractor, and text semantic encoding is performed based on the pre-trained BERT model to obtain the text semantic vector; The multi-scale spatiotemporal features are input into the behavior feature extractor to extract the temporal features of user behavior; The multi-scale spatiotemporal features are input into a relationship feature extractor to extract relationship features between users; The text semantic vector and the temporal features of the user behavior are input into a two-layer LSTM temporal encoder to obtain a temporal feature vector; The relationship features between the users are input into a graph neural network relationship encoder to obtain a relationship feature vector; The temporal feature vector and relational feature vector are input into the multi-head attention Transformer model to obtain preliminary fused features; The preliminary fusion features are input into a deep fusion layer for feature integration to obtain the fusion feature vector.

3. The adaptive evaluation method for network heat rating adversarial effectiveness based on spatiotemporal awareness as described in claim 1, characterized in that, The process of training the three-party policy network model based on the fused feature vectors to determine the equilibrium strategy and the corresponding evaluation score for the equilibrium strategy includes: The three-party policy network model to be trained is trained based on the fused feature vector to generate platform policy parameters corresponding to the platform policy network, defense policy parameters corresponding to the defender policy network, and attack policy parameters corresponding to the attacker policy network. The platform strategy parameters, the defense strategy parameters, and the attack strategy parameters are constrained by action space to limit the range of continuous actions, resulting in a strategy vector. The policy vector is input into the Nash equilibrium solution iterative optimization algorithm to obtain the candidate equilibrium policy and the policy change rate corresponding to the candidate equilibrium measurement. If the rate of change of the strategy is less than a preset threshold, it is determined that the three-party strategy network model to be trained has reached an equilibrium state, and a candidate equilibrium strategy and a candidate evaluation score corresponding to the candidate equilibrium strategy are output. The candidate equilibrium strategy is then used as the equilibrium strategy of the three-party game, and the candidate evaluation score is used as the evaluation score.

4. The adaptive evaluation method for network heat rating adversarial effectiveness based on spatiotemporal awareness as described in claim 3, characterized in that, The 512-dimensional environmental feature vector is input into the DDPG weight adjuster of the adaptive evaluation model, and the weight adjustment action corresponding to the evaluation index is output through the Actor policy network corresponding to the three-party policy network model. The weight configuration Q value corresponding to the weight adjustment action is evaluated by the Critic value network corresponding to the three-party policy network model, wherein the Q value is used to indicate the weight quality. The Actor policy network is optimized based on the Q value, and the weight adjustment action is constrained and optimized to obtain the new evaluation index weight.

5. The adaptive evaluation method for network heat rating adversarial effectiveness based on spatiotemporal awareness as described in claim 4, characterized in that, The process of determining an adaptive evaluation score based on the weights of the new evaluation indicators and the evaluation scores includes: Based on the game state information corresponding to the equilibrium strategy of the three-party game and the weight of the new evaluation index, the evaluation values ​​of the four dimensions of propagation effectiveness, adversarial robustness, platform stability and social impact are calculated respectively. The evaluation values ​​of the four dimensions are weighted and fused with the evaluation score to obtain a real-time comprehensive evaluation score; The real-time comprehensive evaluation score is smoothed using an exponential moving average algorithm to obtain the adaptive evaluation score.

6. An adaptive evaluation system for network heat map adversarial effectiveness based on spatiotemporal awareness, characterized in that, include: The spatiotemporal sensing acquisition module is used to input multi-source network data and generate multi-scale spatiotemporal features; A feature deep fusion processing module is used to generate a fused feature vector based on the multi-scale spatiotemporal features; The multi-party game decision module is used to train the three-party strategy network model to be trained based on the fused feature vector, and to determine the equilibrium strategy of the three-party game and the evaluation score corresponding to the equilibrium strategy of the three-party game. The three-party strategy network model includes a platform strategy network, a defender strategy network, and an attacker strategy network, all of which adopt the Actor-Critic architecture. The adaptive evaluation score determination module is used to adjust the evaluation indicators based on environmental state characteristics through an adaptive evaluation model to obtain new evaluation indicator weights. And, for determining an adaptive evaluation score based on the weights of the new evaluation index and the evaluation score; The evaluation indicators include four dimensions: dissemination effectiveness, resilience against attacks, platform stability, and social impact. The process of adjusting the evaluation indicators based on environmental state characteristics using an adaptive evaluation model to obtain new evaluation indicator weights includes: Based on the three-party game equilibrium strategy and the environmental state characteristics, a 512-dimensional environmental feature vector is constructed, wherein the 512-dimensional environmental feature vector includes 128-dimensional game state characteristics generated by the three-party strategy network model, 128-dimensional spatiotemporal characteristics, 128-dimensional historical evaluation trends, and 128-dimensional system resource status. The step of inputting multi-source network data into a multi-scale spatiotemporal data fusion model to generate multi-scale spatiotemporal features includes: The multi-source network data is sampled according to a time window to obtain sampled data; The sampled data is processed in parallel using a real-time streaming data processing algorithm to obtain temporary data. A user influence network is constructed based on the temporary data, and node attribute features are generated. The temporary data is distributed and stored to generate multi-scale time stamps; Based on the node attribute features and the multi-scale time stamp, spatiotemporal features are fused to generate the multi-scale spatiotemporal features; The time window includes an hourly time window, a minute-level time window, and a second-level time window. The sampled data includes hourly topic sampled data collected corresponding to the hourly time window, minute-level forwarding sampled data collected corresponding to the minute-level time window, and second-level comment sampled data collected corresponding to the second-level time window. The process of constructing a user influence network based on the temporary data and generating node attribute features includes: Extract the target user's basic user characteristics and the interaction relationships between the target user and other users from the temporary data; The user's basic characteristics are input into the improved PageRank algorithm model for iterative calculation to obtain the first user influence value of each target user. Determine the time interval between the occurrence time of each target user's behavior and the current processing time, calculate the time decay factor based on the time interval, and use the time decay factor to dynamically correct the first user influence value of the corresponding target user to obtain the second user influence value of each target user. The user influence network is constructed by using each target user as a node, the interaction relationship between users as directed edges, the second user influence value of the corresponding target user as the initial weight of the node, and the strength of the interaction relationship as the edge weight. Based on the user influence network, multi-dimensional features of each user node are extracted and integrated to generate node attribute features of each user node.

7. A terminal device, characterized in that, The terminal device includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 5.

Images