Computing power network dynamic trust evaluation and access control method based on GRU-BiLSTM fusion model

Abstract

The invention discloses a computing power network dynamic trust evaluation and access control method based on a GRU-BiLSTM fusion model, and the method comprises the steps: capturing the behavior dependence characteristics of a user and a node in a short period through a GRU, and employing the BiLSTM to model a bidirectional time sequence evolution rule of a trust relation, thereby achieving the classification prediction of the credibility level of the user and the regression estimation of the credibility degree of the node, and achieving the access control of the network. On the basis, a four-level access strategy is designed in combination with a system real-time load and a trust volatility, adaptive optimization of access control is realized by dynamically adjusting an access threshold, joint prediction of a bidirectional trust relationship between a user and a node is realized, and a more accurate behavior basis is provided for an access decision; and secondly, a trust volatility index is introduced, and a dynamic closed-loop mechanism of trust evaluation-policy execution-volatility updating is constructed, so that the access control policy can be adaptive to evolution of trust situation, and intelligent and refined authority management is realized.

Description

Technical Field

[0001] This invention relates to the field of time series prediction technology, and in particular to a method for dynamic trust assessment and access control of computing power networks based on a GRU-BiLSTM fusion model. Background Technology

[0002] With the deep integration of 5G, IoT (IoT), and edge computing technologies, computing networks, as the core carrier of cloud-edge-device resource collaboration, have become a key infrastructure supporting the "Eastern Data, Western Computing" project and the Digital China strategy. Through the unified orchestration of distributed computing, storage, and network resources, they achieve elastic scheduling and efficient reuse of computing power. However, their openness and heterogeneity also pose serious security challenges.

[0003] On the one hand, the dynamic access of numerous edge nodes and users leads to highly complex and uncertain entity behavior patterns, where malicious entities may impersonate trusted users to steal resources or leak sensitive data. On the other hand, in the frequent interactions between multiple users and multiple nodes, trust relationships exhibit significant temporal evolution characteristics, making it difficult for traditional static trust assessment methods to effectively capture the gradual decay or abrupt changes in trust levels. Simultaneously, fixed access control policies cannot adapt to the dynamic demands brought about by trust fluctuations and system load changes, easily leading to resource waste and security degradation, thus limiting the practical application of computing networks in critical scenarios such as cross-domain collaboration and high-sensitivity task scheduling.

[0004] Therefore, constructing an intelligent trust management mechanism with temporal modeling capabilities, accurate trust prediction, and dynamic access control adjustment has become a key issue in ensuring the security and efficient operation of computing power networks. Although existing research has achieved some results in trust assessment and access control, such as trust modeling methods based on scoring mechanisms, game theory models, or shallow machine learning, none have fully captured the evolution of trust in long-term behavior. While deep learning trust models have improved temporal modeling capabilities in recent years—Hua et al. used an LSTM model to predict node trust—they neglect bidirectional temporal dependencies. Traditional access control methods often rely on static rules or role divisions; for example, Chunge et al.'s trust threshold mechanism lacks awareness of the overall trust situation of the system, and Stodt et al.'s IoT dynamic zero-trust architecture does not optimize policy granularity for the 'cross-domain scheduling + high load fluctuation' characteristics of computing power networks. This makes it difficult for existing solutions to meet the core needs of computing power networks for 'dynamic trust awareness + fine-grained access control'. Summary of the Invention The purpose of this invention is to provide a dynamic trust assessment and access control method for computing power networks based on the GRU-BiLSTM fusion model. This method addresses the increasing complexity of user behavior and the heterogeneity of nodes, as well as the high-frequency dynamic interaction behavior of multiple users and nodes in computing power networks. This makes it difficult for traditional static trust mechanisms to capture long-term trust evolution patterns, and fixed access control strategies cannot adapt to real-time changes in trust levels and system load, leading to problems such as resource waste and data leakage.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: 1. A method for dynamic trust assessment and access control of computing power networks based on the GRU-BiLSTM fusion model, the method comprising the following steps; Step 1: Time series data preprocessing. The computing power network acquires user behavior data and node operation data, performs feature filtering on the user behavior data and node operation data, processes the filtered features using the Z-score normalization algorithm, and uses a sliding window to truncate continuous records to obtain a time series feature matrix. Step 2: Construct and train the GRU-BiLSTM fusion model. The GRU-BiLSTM fusion model includes a GRU layer, a BiLSTM layer and a Dropout layer. The time-series feature matrix is ​​input into the GRU layer, which captures short-term behavioral features. Then, the BiLSTM layer extracts long-term bidirectional correlation features. Finally, the Dropout layer suppresses overfitting, and the fully connected layer outputs the trust evaluation result. Step 3: Collect the real-time system load L of the computing power network, and dynamically calculate the trust access threshold by combining the trust volatility of users and nodes. Step four: Based on the trust access threshold and the trust assessment result, a multi-level access strategy is executed. After the multi-level access strategy is executed, the computing power network records the interaction results and feeds them back to the trust volatility calculation, forming a closed-loop adjustment mechanism.

[0006] Furthermore, the filtering features of the user behavior data include five indicators: task completion rate, payment rate, number of defaults, historical compliance frequency, and data upload integrity. The filtering features of the node operation data include four indicators: load coefficient, response time, task completion rate, and error rate. The sliding window is set to a window size of 12, and after truncation, a time-series feature matrix of 12×5 on the user side and 12×4 on the node side is formed. The Z-score standardization converts the original feature value x into a standardized value z by calculating the feature mean μ and standard deviation σ.

[0007] Furthermore, the trust assessment results include user trust level and node trust score. The user trust level is obtained by inputting the 12×5 temporal feature matrix on the user side into the GRU-BiLSTM fusion model, and the node trust score is obtained by inputting the 12×4 temporal feature matrix on the node side into the GRU-BiLSTM fusion model. The user trust level is divided into three categories: low-medium, medium, and high trust level. The node trust score outputs a continuous trust score in the range of 0-1.

[0008] Furthermore, the trust access threshold includes a rejection threshold T_reject, a restricted access threshold T_limit, and a preferred access threshold T_preferred. Simultaneously, when the real-time system load L increases, the trust access threshold decreases; when the system volatility increases, the trust access threshold increases.

[0009] Furthermore, in step four, the multi-level access strategy is as follows: when the user's trust level is low or the node's trust score is lower than T_reject, access is denied; when the user's trust level is medium and the node's trust score is between T_reject and T_limit, access is restricted; when the user's trust level is medium and the node's trust score is higher than T_limit, or when the user's trust level is high and the node's trust score is lower than T_excellent, access to the regular resource pool is allowed; and when the user's trust level is high, the node's trust score is not lower than T_excellent and the load L ≤ 0.6, priority access is given to allocate high-quality resources.

[0010] Furthermore, the GRU layer updates the gate. With Reset Door Selective memorization of short-term behavioral characteristics can be achieved through the following formula; ; in, for Input features at all times, for Hide your status at all times. / / , is the input layer weight matrix. / / The hidden layer weight matrix is... / / , where is the bias peak, and σ is the sigmoid activation function. is the Hadamard product, and tanh is the hyperbolic tangent activation function.

[0011] Furthermore, the BiLSTM layer uses forward LSTM to hide the state and capture the impact of historical behavior on the current trust, and uses backward LSTM to hide the state and capture the correction of future behavior on the current trust. The forward and backward states are then combined to form a complete temporal feature, which is achieved through the following formula. Forward LSTM hidden state: ; Backward LSTM hidden state: ; Bidirectional feature fusion: ; in, For sequence elements, This is the previous hidden state. This is the backward hidden state for the next time step. This represents the current backward hidden state. The current state is hidden from view. Furthermore, the range of the real-time system load of the computing power network is as follows: The trust volatility includes user trust volatility. And node trust volatility The user trust volatility The node trust volatility is the standard deviation of the trust values ​​from the last 10 user entity interactions. The standard deviation of trust values ​​from the last 10 node entity interactions is used to calculate the trust access threshold. The specific formula for this calculation is as follows: ; in, The threshold for denying access, To limit access thresholds, the threshold decreases as L increases. / It increases as it rises.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: First, the GRU-BiSTM fusion model enables joint prediction of the two-way trust relationship between users and nodes, providing more accurate behavioral basis for access decisions.

[0013] Second, the trust volatility index is introduced to construct a dynamic closed-loop mechanism of "trust assessment - strategy execution - volatility update", which enables access control policies to adapt to the evolution of trust status and achieve intelligent and refined access control. Attached Figure Description

[0014] The present invention will be further described below with reference to the accompanying drawings: Figure 1 This invention provides a trust-aware dynamic access control framework. Figure 2 This is the GRU-BiLSTM fusion architecture of the present invention; Figure 3This is a flowchart illustrating the dynamic trust assessment and access control method for computing power networks based on the GRU-BiLSTM fusion model of the present invention. Figure 4 This is a comparison of the training curves of the GRU-BiLSTM fusion model of the present invention with those of four other time series models; Figure 5 This invention relates to the GRU-BiLSTM fusion model and its connection to four other time series models. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0016] The technical solution of the present invention will be described in detail below with reference to specific embodiments. The following specific embodiments may be combined or substituted with each other according to the actual situation, and the same or similar concepts or processes may not be described again in some embodiments.

[0017] Example 1, as Figures 1 to 3 As shown, this invention provides a dynamic trust assessment and access control method for computing power networks based on the GRU-BiLSTM fusion model. This invention employs a distributed architecture of 1 server + 8 edge nodes to simulate a computing power network. Server: Intel Core i7-12700K processor (12 cores, 20 threads, base frequency 3.6GHz, boost frequency 5.0GHz), 64GB DDR4 memory (3200MHz, dual-channel), 2TB NVMe SSD storage (read speed 3500MB / s); Edge nodes: Intel Core i5-1135G7 processor (4 cores, 8 threads, base frequency 2.4GHz, boost frequency 4.2GHz), 16GB LPDDR4 memory (3733MHz), 512GB SSD storage (read speed 2400MB / s).

[0018] Network environment: Connected via Gigabit Ethernet (CAT6 cable), with an average communication latency of <5ms and a packet loss rate of <0.1%, ensuring data transmission efficiency meets the requirements of real-time trust assessment.

[0019] Operating System: Windows 10 Professional (version 21H2, Build 19044.3086); Development Tools: PyCharm 2023.1.3 (Professional Edition), Python 3.9.13 (64-bit); Deep Learning Framework: TensorFlow 2.10.0 (accelerated by CUDA 11.7 and cuDNN 8.4.1); Data Processing Libraries: Pandas 1.5.2 (data cleaning, feature engineering), NumPy 1.23.5 (matrix operations, standardization), Scikit-learn 1.2.2 (data partitioning, indicator calculation); Visualization Tools: Matplotlib 3.6.2 (curve plotting, chart generation), Seaborn 0.12.1 (statistical chart enhancement).

[0020] S1. Time-series data preprocessing: A fully simulated dataset was constructed using industry-standard driven and scenario feature mapping. The dataset for the computing power network was built using the Python Faker library and a custom statistical generator, generating 200,000 interaction records from 1,000 user behavior data points and 1,000 edge node operation data points. Feature filtering was performed on these 200,000 interaction records. On the user side, five behavioral features conforming to the ETSIGSNFV-IFA027 computing power network entity behavior evaluation standard were extracted: task completion rate, payment rate, number of defaults, historical compliance frequency, and data upload integrity. On the node side, four features were selected based on Huawei Cloud Stack edge nodes: load coefficient (weight 0.3), response time (weight 0.25), task completion rate (weight 0.3), and error rate (weight 0.15). The operational characteristics of the monitoring system, and the specific features of these technical fields, anchor the trust assessment to the behavioral dimensions of real-world scenarios. Simultaneously, Z-score standardization is used to process the original features, transforming features of different dimensions within the technical fields, such as the proportion of task completion rates and the duration of response times, into standardized data that the algorithm can directly process. For example, the KL divergence between the standardized user "task completion rate" and real-world scenario data is only 0.08, ensuring the compatibility between the technical data and the algorithm input. Then, a sliding window technique (window size 12, capturing 12 hours of historical data) is used to capture continuous records, ultimately forming a 12x5 time-series feature matrix on the user side and a 12x4 matrix on the node side, with approximately 240 samples in each. The above data is used to train the GRU-BiLSTM fusion model and evaluate the trust assessment results.

[0021] S2, Construct a GRU-BiLSTM fusion model, the model adopts; (1) Input layer: receives the standardized temporal feature matrix; (2) GRU layer: Focuses on short-term behavioral mutations, filters noise through update gates and reset gates, and retains key instantaneous features; (3) BiLSTM layer: Models long-term bidirectional dependencies, forward propagation captures historical cumulative effects, and backward propagation captures future trend predictions; (4) Dropout layer: Randomly deactivates some neurons to suppress overfitting; (5) Output layer: User trust evaluation uses 3 neurons sofmax output, and node trust evaluation uses 1 neuron sigmoid output (0-1 continuous value); This invention adopts a hierarchical structure of "input layer - GRU layer - BiLSTM layer - Dropout layer - fully connected output layer": the GRU layer has 64 neurons, which are updated through the update gate; With reset door; The gating mechanism accurately captures short-term behavioral mutations such as sudden user defaults and instantaneous node failures. The GRU layer effectively addresses the issue of trust states in computing power networks being susceptible to transient behavioral changes. The BiLSTM layer, with 128 neurons, captures the cumulative effect of historical behavior through forward LSTM. Backward LSTM correction for future trends Then, through the bidirectional hidden state fusion formula... The model fully embodies the temporal characteristics of trust status, which has a long-term temporal evolution. In the output layer, the softmax activation function outputs three discrete trust levels of users: "low", "medium", and "high". The sigmoid activation function outputs a continuous trust score within the range of node invention ([0,1] invention). These algorithm outputs are directly transformed into the basis for "user risk classification" and "node service stability quantification" in terms of technology, providing operable trust assessment results for subsequent access control and realizing the precise response of algorithm features to technical requirements.

[0022] S3, Calculation of the dynamic trust access threshold: The real-time system load L of the computing power network is obtained by sampling the CPU utilization (weight 0.6) and memory usage (weight 0.4) every 5 minutes from the central server, with a weighted average value. This reflects the scarcity of computing network resources; trust volatility (user trust volatility) With node trust volatility The stability of entity behavior is reflected by calculating the "standard deviation of trust values ​​over the last 10 interactions." Both factors contribute to the threshold calculation formula: Rejection Threshold. Limiting threshold Priority threshold (Fixed value) so that the threshold can adapt to the dynamic changes in computing power network resources (e.g., when L increases from 0.3 to 0.7, From 0.33 to 0.37 The threshold was reduced from 0.555 to 0.495, easing access restrictions for high-trust entities to improve resource utilization, and adjustments can be made based on entity behavior risk (e.g., ...). + When it rises from 0.1 to 0.4, From 0.32 to 0.38 The threshold was raised from 0.585 to 0.63, tightening permissions for low-trust entities to enhance security controls.

[0023] S4, based on the aforementioned dynamic threshold and the trust assessment results output by the algorithm, this invention sets a four-level access strategy: "deny, restrict, allow, and prioritize": Deny access corresponds to "user trust level is low, or node trust score < " "), prohibits entities from participating in any resource interactions; restricts access to entities with a medium user trust level and a low node trust score." , Access is only permitted to non-sensitive computing resources (such as ordinary data storage nodes); access is permitted to those with a user trust level of medium and a node trust score ≥ Or, the user level is high and the node score is ≤ "Allows access to the regular resource pool (such as standard compute nodes); prioritizes access to nodes with a high user trust level and a node trust score ≥ " Furthermore, the system load L≤0.6” prioritizes the allocation of high-quality resources (such as low-latency edge nodes and high-performance GPU nodes). This strategy combines the algorithm’s trust assessment results with the resource classification and load status of the technical field, directly solving the problem of “security and efficiency coordination”.

[0024] Furthermore, the closed-loop feedback mechanism enables a continuous optimization interaction between technical features and algorithmic features. After the strategy is executed, the system records technical data such as the success rate of high-trust user access (target ≥ 0.83), the rejection rate of low-trust user access (target ≥ 0.95), the core resource access interception rate (target ≤ 0.6), and the risk-permission correlation (target ≤ -0.8). This data is fed back into the calculation of trust volatility, dynamically updating the system. and This allows for the adjustment of the access threshold in the next round, achieving closed-loop management of "trust assessment - strategy execution - volatility update - threshold adjustment," thereby continuously improving the accuracy of trust assessment and the adaptability of access control.

[0025] Compared with existing technologies, this invention uses a GRU-BiLSTM fusion model to jointly predict the two-way trust relationship between users and nodes, providing more accurate behavioral basis for access decisions; it introduces a trust volatility index to construct a dynamic closed-loop mechanism of "trust assessment – ​​policy execution – volatility update", enabling access control policies to adapt to the evolution of trust status and achieve intelligent and refined permission management.

[0026] Specifically, the model captures the short-term behavioral dependency characteristics of users and nodes using GRU, and models the bidirectional temporal evolution of trust relationships using BiLSTM, thereby achieving classification prediction of user trust levels and regression estimation of node trust levels. Based on this, and combining real-time system load and trust volatility, a four-level access strategy (deny access, restrict access, allow access, and prioritize access) is designed, and adaptive optimization of access control is achieved by dynamically adjusting access thresholds.

[0027] Example 2: Training the GRU-BiLSTM fusion model using the following strategy: The training of the GRU-BiLSTM invention fusion model revolves around two core tasks: invention "user trust level classification" and invention "node trust score regression." Through customized loss functions, optimization strategies, and multi-dimensional verification, it achieves accurate assessment of the trust status of the computing power network. To adapt to the characteristics of the dual tasks, the model designs a sparse cross-entropy loss function for the user trust classification task to enhance the identification accuracy of risky samples such as inventions with "low trust." The formula is as follows: (where N is the sample size, It is the true label of the invention of the i-th sample and k-th category, which is "low invention / medium invention / high invention". (For predicting probabilities); a mean squared error loss function is designed for the node trust regression task to penalize large prediction errors and improve the stability of trust scores. The formula is: ;(in This represents the node's true trust value. (For predicted values). In terms of optimization strategy and parameter configuration, the Adam optimizer was selected to leverage its adaptive learning rate feature to accelerate training convergence. The initial learning rate was set to 0.001 and decayed by 10% every 5 training rounds (to avoid oscillations in the later stages of training). Simultaneously, weight decay (coefficient) was introduced. To suppress overfitting, the training iterations were set to 30 rounds, with an early stopping strategy (training was terminated and the optimal model was saved when the validation set loss stopped decreasing after 3 consecutive iterations). The batch size was set to 64 to balance training efficiency and memory usage. To verify the model performance, the dataset was divided into training, validation, and test sets according to a ratio of 7:1:2. The results on the test set showed that the optimal configuration was determined to be 64 neurons in the GRU, 128 neurons in the BiLSTM, and a Dropout rate of 0.2.

[0028] like Figure 4 and Figure 5 As shown, the GRU-BiLSTM model of the present invention compares five types of time series models, including GRU, which focuses on short-term time series independence, LSTM, which focuses on unidirectional long-term time series, BiLSTM, which can bidirectionally capture long-term time series correlation, and unidirectional cascaded GRU-LSTM.

[0029] Input: The user model consists of 12x5 time series features, and the node model consists of 12x4 time series features; Parameters: All deep learning models are trained for 30 epochs, batch size is 64, and learning rate is 0.001; Repeated experiments: Each model was run independently 10 times, and the mean ± standard deviation of the index was taken as the result. The dynamic access control mechanism was used to verify the experiment, as shown in the training curves of the five models in the figure. The BiLSTM model's loss fluctuated significantly in the middle of training (10-15 epochs), and the final loss... The higher loss rate of the fusion model indicates that bidirectional long-term modeling alone cannot efficiently integrate short-term behavioral mutation information, resulting in limited fitting efficiency and accuracy. The GRU model shows a rapid decrease in training loss, but the validation loss increases slightly after 10 epochs, suggesting that GRU's overemphasis on short-term features leads to insufficient fitting of long-term patterns in the validation set, posing a certain risk of overfitting. The GRU-LSTM model shows a better decreasing trend in both training and validation losses than the single model, but the difference is still significant (lower training loss, higher validation loss), indicating that the unidirectional cascade fusion of GRU and LSTM does not fully leverage the advantages of bidirectional temporal data, and there is redundancy in the interaction between short- and long-term features. Both training and validation losses of the LSTM model show a decreasing trend, but the validation loss fluctuates drastically (suddenly spikes after about 10 epochs). This reflects that LSTM is not robust enough to short-term burst noise in unidirectional long-term modeling and is easily affected by abnormal samples. The GRU-BiLSTM model of this invention shows that the training and validation losses decrease rapidly and synchronously, and basically converge after 25 epochs with minimal difference. This demonstrates that GRU can quickly capture short-term features (rapid loss decrease in the early stage) while BiLSTM avoids overfitting through bidirectional long-term modeling, achieving an optimal balance between efficient training and low overfitting risk, as shown in the table below. Table 1 Fitting results of different models As shown in Table 1, the accuracy of the GRU-BiLSTM of this invention is improved by 5.54% and the F1 score is improved by 6.08% compared with BiLSTM. The main reason is that the GRU layer effectively captures short-term behaviors such as sudden user defaults, while the BiLSTM layer simultaneously mines the long-term positive impact of historical compliance behavior on current trust and the reverse constraint of future behavior trends on trust, thus achieving the synergistic utilization of short-term and long-term features.

[0030] Comparison of the effects of dynamic access control mechanisms: Table 2 Access control results of different models As shown in Table 3, regarding the pass rate for high-trust users, both GRU-BiLSTM and GRU-LSTM achieved the highest value of 0.835, but the former's standard deviation was much lower than the latter, outperforming LSTM, BiLSTM, and GRU. This difference stems from the hierarchical collaborative architecture of GRU-BiLSTM: its GRU layer can accurately capture users' short-term sudden compliance behaviors, while the BiLSTM layer can model the two-way long-term compliance cumulative effect, avoiding the fluctuations in misjudgment of high-trust users by a single model. Regarding the rejection rate for low-trust users, all models maintained a high security level above 0.95, with GRU-LSTM slightly higher. However, while maintaining a high rejection rate, GRU-BiLSTM did not exhibit the excessive control of core resources seen in other models, initially demonstrating the design goal of security without sacrificing efficiency.

[0031] The core resource control of GRU-BiLSTM is significantly less stringent than other models, and it has the smallest standard deviation. This result directly verifies the effectiveness of the dynamic threshold adjustment algorithm: after GRU-BiLSTM outputs the trust assessment result, the mechanism lowers the threshold according to the increase in system load and raises the threshold according to the increase in trust volatility—for example, when a high-trust user is matched with a high-trust node and L0.6, priority access is automatically triggered to reduce core resource interception; while other models only set a fixed threshold based on trust level, resulting in core resources being in a state of high interception rate for a long time, which contradicts the requirement of the computing power network to dynamically adapt to load and trust volatility.

[0032] In the risk-permission correlation index, GRU-BiLSTM shows a strong negative correlation with both GRU and GRU-LSTM, significantly outperforming LSTM and BiLSTM. Considering the model design motivation, the model incorporating the GRU structure can quickly capture short-term risk mutations such as instantaneous user defaults and temporary node failures, thus mapping high-risk signals to low-permission policies in real time. While BiLSTM can characterize long-term risk evolution, its response to short-term risks is lagging, leading to a decrease in the matching degree between permission allocation and risk status. The balance between correlation strength and result stability of GRU-BiLSTM demonstrates that it can provide real-time and accurate decision-making basis for access control mechanisms.

[0033] While maintaining an industry-leading pass rate for high-trust users and a rejection rate for low-trust users, the GRU-BiLSTM model achieves the goal of minimizing the strictness of core resource control and accurately refining the correlation between risk and permissions through multi-dimensional collaborative regulation of trust, load, and volatility. This solves the research objective of adapting traditional static strategies and provides data support for the subsequent application of scenarios such as edge node access in "East Data West Computing" and cross-domain medical data scheduling.

[0034] The comparison conclusions between the GRU-BiLSTM model of this invention and five types of time series models are as follows: 1. The GRU-BiLSTM model outperforms the comparison models in both user trust classification (accuracy 97.44%, F1=97.43%) and node trust regression tasks. The fusion architecture effectively balances the requirements of short-term mutation capture (GRU) and bidirectional long-term modeling (BiLSTM). 2. Dynamic access control mechanism achieves synergistic optimization of security and resource efficiency: high-trust user pass rate 0.835, low-trust user rejection rate 0.956, risk-permission negative correlation -0.803, core resource control strictness 0.594; 3. Single sample preprocessing time is approximately 0.5ms, model training time is approximately 1.5h (based on 64GB memory and CUDA 11.7), and policy execution time is approximately 10ms / time, meeting the real-time requirements of a 1000-node computing network (latency <50ms).

[0035] By comparison, the present invention, employing the GRU-BiLSTM fusion model, has the following advantages over existing technologies in predicting trust levels and dynamically adjusting access permissions: 1. GRU's gating mechanism reduces the loss of short-term behavioral features: On user sudden default samples, GRU-BiLSTM achieved an F1 score of 97.43%, a 5.52% improvement over BiLSTM (91.91%), validating its ability to capture short-term mutations; BiLSTM's bidirectional processing captures long-term trust evolution; 2. Threshold Adjustment Formula for Responding to Entity Behavior Fluctuations: When the trust volatility increases from 0.05 to 0.2, T changes from... The rejection rate for low-trust users increased by 4.2% from 0.35 to 0.45, effectively suppressing the risk of high volatility; 3. Engineering efficiency is adapted to large-scale deployment: at a scale of 2000 users / nodes, the policy execution latency is ≤15ms, meeting the real-time access control requirements of computing power networks.

[0036] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for users of ordinary technical skills, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic trust assessment and access control of computing power networks based on a GRU-BiLSTM fusion model, characterized in that, The method includes the following steps; Step 1: Time series data preprocessing. The computing power network acquires user behavior data and node operation data, performs feature filtering on the user behavior data and node operation data, processes the filtered features using the Z-score normalization algorithm, and uses a sliding window to truncate continuous records to obtain a time series feature matrix. Step 2: Construct and train the GRU-BiLSTM fusion model. The GRU-BiLSTM fusion model includes a GRU layer, a BiLSTM layer and a Dropout layer. The time-series feature matrix is ​​input into the GRU layer, which captures short-term behavioral features. Then, the BiLSTM layer extracts long-term bidirectional correlation features. Finally, the Dropout layer suppresses overfitting, and the fully connected layer outputs the trust evaluation result. Step 3: Collect the real-time system load L of the computing power network, and dynamically calculate the trust access threshold by combining the trust volatility of users and nodes. Step four: Based on the trust access threshold and the trust assessment result, a multi-level access strategy is executed. After the multi-level access strategy is executed, the computing power network records the interaction results and feeds them back to the trust volatility calculation, forming a closed-loop adjustment mechanism.

2. The method for dynamic trust assessment and access control of computing power networks based on the GRU-BiLSTM fusion model according to claim 1, characterized in that, The filtering features of the user behavior data include five indicators: task completion rate, payment rate, number of defaults, historical compliance frequency, and data upload integrity. The filtering features of the node operation data include four indicators: load coefficient, response time, task completion rate, and error rate. The sliding window is set to a window size of 12, and after truncation, a time-series feature matrix of 12×5 on the user side and 12×4 on the node side is formed. The Z-score standardization converts the original feature value x into a standardized value z by calculating the feature mean μ and standard deviation σ.

3. The method for dynamic trust assessment and access control of computing power networks based on the GRU-BiLSTM fusion model according to claim 1, characterized in that, The trust assessment results include user trust level and node trust score. The user trust level is obtained by inputting the 12×5 temporal feature matrix on the user side into the GRU-BiLSTM fusion model, and the node trust score is obtained by inputting the 12×4 temporal feature matrix on the node side into the GRU-BiLSTM fusion model. The user trust level is divided into three categories: low-medium, medium-high, and medium-high. The node trust score outputs a continuous trust score in the range of 0-1.

4. The method for dynamic trust assessment and access control of computing power networks based on the GRU-BiLSTM fusion model according to claim 1, characterized in that, The trusted access threshold includes a rejection threshold T_reject, a restricted access threshold T_limit, and a preferred access threshold T_preferred. Furthermore, when the real-time system load L increases, the trusted access threshold decreases; when the system volatility increases, the trusted access threshold increases.

5. The method for dynamic trust assessment and access control of computing power networks based on the GRU-BiLSTM fusion model according to claim 1, characterized in that, In step four, the multi-level access strategy is as follows: when the user's trust level is low or the node's trust score is lower than T_reject, access is denied; when the user's trust level is medium and the node's trust score is between T_reject and T_limit, access is restricted; when the user's trust level is medium and the node's trust score is higher than T_limit, or when the user's trust level is high and the node's trust score is lower than T_excellent, access to the regular resource pool is allowed; and when the user's trust level is high, the node's trust score is not lower than T_excellent and the load L ≤ 0.6, priority access is given to allocate high-quality resources.

6. The method for dynamic trust assessment and access control of computing power networks based on the GRU-BiLSTM fusion model according to claim 1, characterized in that, The GRU layer uses an update gate. With Reset Door Selective memorization of short-term behavioral characteristics can be achieved through the following formula; ； in, for Input features at all times, for Hide your status at all times. / / , is the input layer weight matrix. / / The hidden layer weight matrix is... / / , where is the bias peak, and σ is the sigmoid activation function. is the Hadamard product, and tanh is the hyperbolic tangent activation function.

7. The method for dynamic trust assessment and access control of computing power networks based on the GRU-BiLSTM fusion model according to claim 1, characterized in that, The BiLSTM layer uses a forward LSTM to hide the state and capture the impact of historical behavior on the current trust, and uses a backward LSTM to hide the state and capture the correction of the current trust by future behavior. The forward and backward states are then concatenated to form a complete temporal feature, achieved through the following formula: Forward LSTM hidden state: ; Backward LSTM hidden state: ; Bidirectional feature fusion: ; in, For sequence elements, This is the previous hidden state. This is the backward hidden state for the next time step. This represents the current backward hidden state. The current state is hidden from view.

8. The method for dynamic trust assessment and access control of computing power networks based on the GRU-BiLSTM fusion model according to claim 1, characterized in that, The range of the system load is as follows: The trust volatility includes user trust volatility. And node trust volatility The user trust volatility The node trust volatility is the standard deviation of the trust values ​​from the last 10 user entity interactions. The standard deviation of trust values ​​from the last 10 node entity interactions is used to calculate the trust access threshold. The specific formula for this calculation is as follows: ； in, The threshold for denying access, To limit access thresholds, To prioritize access to the threshold, the threshold decreases as L increases. / It increases as it rises.

Images