A cloud storage resource dynamic load balancing method and system based on deep learning

Abstract

This invention discloses a deep learning-based dynamic load balancing method and system for cloud storage resources, relating to the field of cloud storage technology. This deep learning-based dynamic load balancing method for cloud storage resources collects access sequences and node load data from the cloud storage cluster, inputs them into a prediction model, and obtains hotspot prediction results, confidence scores, and adversarial scores. It constructs a multi-tenant association graph with tenants as nodes and temporal synchronization as edge weights. Group behavior features are extracted through a graph neural network, fused with tenant historical behavior baselines and temporal compliance features, and then output causal inference strength via a structural causal model. This invention establishes a dual verification mechanism combining an adversarial discrimination subnetwork and a structural causal model, combined with the group behavior features extracted from the multi-tenant association graph, forming a closed loop from identification to blocking to feedback updates. This provides a multi-source verification and continuously evolving scheduling architecture for cloud storage load balancing.

Description

Technical Field

[0001] This invention relates to the field of cloud storage technology, specifically to a method and system for dynamic load balancing of cloud storage resources based on deep learning. Background Technology

[0002] As a crucial component of cloud computing infrastructure, cloud storage systems virtualize numerous storage devices into a unified storage resource pool, providing elastic and scalable data storage services to multiple tenants. With the rapid development of applications such as short videos, social media, and online office work, cloud storage clusters face access requests characterized by high concurrency, large fluctuations, and strong bursts. How to rationally allocate storage media and computing resources within the cluster while ensuring service quality has become a key issue in the design and operation of cloud storage systems.

[0003] Load balancing technology, as a crucial resource scheduling tool, distributes access requests rationally across different storage nodes, preventing response latency and service degradation caused by overload of a single node. Traditional load balancing methods primarily employ threshold-based reactive scheduling strategies. These strategies monitor metrics such as CPU utilization, memory utilization, and disk I / O latency, triggering task migration or resource adjustments when these metrics exceed set thresholds. While simple and easy to implement, these methods consistently lag behind load changes, making them ill-suited for handling sudden load spikes.

[0004] In recent years, deep learning technology has been gradually introduced into the field of load balancing due to its advantages in temporal prediction and pattern recognition. Researchers have attempted to build predictive models based on Long Short-Term Memory (LSTM) networks, temporal convolutional networks, or Transformer structures. By analyzing historical access data, they predict future load trends and use the prediction results as a preliminary basis for resource scheduling. These methods transform load balancing from passive response to proactive prediction, achieving better scheduling results in stable load scenarios. Meanwhile, reinforcement learning methods have also been applied to optimize resource scheduling decisions, learning optimal scheduling strategies through interaction with the environment. Furthermore, regarding the issue of hot data in cloud storage, existing research has proposed methods for hotspot identification combining multi-dimensional features such as access frequency and spatiotemporal distribution, and performing operations such as data preheating or replica expansion after identification to improve access efficiency. In multi-tenant scenarios, the access behaviors between tenants often exhibit certain correlations, and some studies attempt to optimize resource allocation by mining collaborative access patterns among tenants.

[0005] The limitations of existing technologies include at least the following problems: In deep learning-based cloud storage load balancing methods, the predictive model assumes that all the load it handles comes from naturally formed data access patterns. It lacks a mechanism for identifying and defending against maliciously constructed adversarial access behaviors in a multi-tenant environment. Malicious tenants take advantage of the high sensitivity of deep learning models to spatiotemporal features to deliberately construct adversarial access sequences that are highly similar to the actual hotspot distribution. This causes the model to misjudge these as hotspot events that are about to erupt, thereby triggering unnecessary data preheating and resource reservation operations. These ineffective preheating operations consume a large amount of high-performance storage media and computing resources, causing other normal tenants on the same physical node to fall into a state of resource starvation and service response time to deteriorate sharply. This defect exposes the design blind spot of prediction-based execution in the current intelligent scheduling architecture. The deep learning load balancing mechanism, which was originally intended to improve efficiency, has instead become an attack entry point that can be used by malicious tenants to seize shared resources. The requests of normal tenants are continuously waiting due to the ineffective preheating of resources, and the overall service performance of the cluster will gradually degrade as the attack continues. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for dynamic load balancing of cloud storage resources based on deep learning. This solves the problems of existing deep learning-based cloud storage load balancing methods being susceptible to malicious attack access deception, leading to invalid resource preheating and crowding out, resource starvation for normal tenants, and performance degradation of cluster services.

[0007] To achieve the above objectives, this invention provides the following technical solution: a dynamic load balancing method for cloud storage resources based on deep learning, comprising the following steps: collecting access sequence data and spatiotemporal characteristic data of node load in a multi-tenant environment cloud storage cluster; inputting this data into a multi-tenant load identification and prediction model; obtaining hotspot prediction results, confidence scores of hotspot prediction results, and adversarial disturbance discrimination scores. This model includes a hotspot prediction subnetwork and an adversarial disturbance discrimination subnetwork. The adversarial disturbance discrimination subnetwork optimizes parameters using real hotspot features as positive samples and adversarial access sequence features constructed by a generative adversarial network as negative samples; constructing a multi-tenant association graph with tenants as nodes and the temporal synchronization degree of access behavior between tenants as edge weights; and collecting... Graph neural networks are used to extract tenant group behavior features. These features are then fused with the tenant historical behavior baseline and access sequence time-series compliance features and input into the structural causal model. The output is the causal inference strength between the access sequence and hot events. The confidence score of the hot event prediction result, the adversarial disturbance discrimination score, and the causal inference strength are input into the gating fusion unit. The access sequence type is determined based on the comparison between the weighted fusion result and the preset execution threshold. Data preheating and resource reservation operations are performed on valid prediction sequences, and corresponding operations are blocked for malicious adversarial access sequences. Such access sequences are fed back as negative samples to the multi-tenant load identification and prediction model and the structural causal model to update the adversarial disturbance discrimination sub-network parameters and the tenant historical behavior baseline, respectively.

[0008] Furthermore, the specific steps for collecting access sequence data and node load spatiotemporal characteristic data of cloud storage clusters in a multi-tenant environment are as follows: Collect tenant access sequence data, including request timestamps, request object identifiers, and request operation types; collect storage node load spatiotemporal characteristic data, including CPU utilization, memory utilization, disk I / O wait time, and network bandwidth utilization; align the two types of data by timestamp to construct a multi-dimensional feature matrix.

[0009] Furthermore, the specific steps for optimizing the parameters of the adversarial perturbation discriminant subnetwork using real hotspot features as positive samples and adversarial access sequence features constructed by the generative adversarial network as negative samples are as follows: Construct a triplet loss function using real hotspot features as anchor samples, historical normal access sequence features of the same tenant as positive samples, and adversarial access sequence features constructed by the generative adversarial network as negative samples; input the three types of samples into the adversarial perturbation discriminant subnetwork and extract the corresponding feature vectors; update the subnetwork parameters by reducing the distance between the anchor sample and the positive sample feature vector and increasing the distance between the anchor sample and the negative sample feature vector.

[0010] Furthermore, the specific steps for constructing a multi-tenant association graph with tenants as nodes and the temporal synchronization degree of access behavior between tenants as edge weights are as follows: slice the access sequence of each tenant according to the time window; calculate the temporal synchronization degree coefficient between tenants based on the peak value of the cross-correlation function of the access timestamp sequence of tenants within the same time window; use the temporal synchronization degree coefficient as the initial weight of the corresponding edge of the association graph; perform embedding learning on the association graph using a graph attention network to assign differentiated attention weights to different neighbor nodes; and extract the behavioral features of the tenant group through the output layer of the graph attention network.

[0011] Furthermore, the specific steps for fusing this feature with the tenant's historical behavior baseline and access sequence time-series compliance features before inputting it into the structural causal model are as follows: Input the tenant's historical behavior baseline into the structural causal model, and output long-term access preference features through a fully connected layer; input the access sequence time-series compliance features into the structural causal model, and output short-term pattern matching scores through a fully connected layer; input the group behavior features into the structural causal model, and output tenant cluster anomaly deviation through a fully connected layer; concatenate the above three types of features to form a fused feature vector; input the fused feature vector into the output layer of the structural causal model to obtain the causal inference strength.

[0012] Furthermore, the specific steps for determining the access sequence type based on the comparison between the weighted fusion result and the preset execution threshold are as follows: Calculate the numerical difference between the confidence score of the hotspot prediction result and the causal inference strength; dynamically adjust the fusion weights of the two based on the magnitude of the difference, increasing the weight of the causal inference strength when the difference exceeds the set range, and increasing the weight of the confidence score of the hotspot prediction result when the difference is less than the set range; add the adversarial disturbance discrimination score to the weighted fusion result as a penalty term; when the adversarial disturbance discrimination score exceeds the preset adversarial threshold, directly determine the corresponding access sequence as a malicious adversarial access sequence; distinguish between valid prediction sequences and malicious adversarial access sequences based on the relationship between the weighted fusion result and the preset execution threshold.

[0013] Furthermore, the specific steps for performing data preheating and resource reservation operations on the effective prediction sequence are as follows: determine the identifier of the data block to be preheated and the reserved capacity of the high-performance storage medium based on the hot spot prediction results; migrate the data block to be preheated from the low-frequency storage medium to the high-frequency storage medium; divide the high-performance storage medium into independent cache intervals and bind them to the corresponding tenant access sequences; when the confidence score of the hot spot prediction result is lower than the set migration threshold, the migration operation from the low-frequency storage medium to the high-frequency storage medium is triggered only when the data block is accessed for the first time.

[0014] Furthermore, the specific steps for feeding this type of access sequence as a negative sample into the multi-tenant load identification and prediction model are as follows: Based on the degree of deviation between the adversarial perturbation discrimination score and the causal inference strength of the malicious adversarial access sequence, the priority of the negative samples is set; the negative samples are stored in the replay buffer in descending order of priority; during the incremental training phase of the model, negative samples are extracted from the buffer according to priority; the extracted samples are input into the adversarial perturbation discrimination sub-network, the triplet loss function is calculated, and the network parameters are updated.

[0015] Furthermore, the specific steps for feeding this type of access sequence as a negative sample back to the structural causal model are as follows: The structural causal model records the frequency and time distribution of the judgment of tenant malicious adversarial access sequences; determines whether the frequency of tenant malicious judgment exceeds the set threshold within a unit time window; marks the historical behavior baseline of tenants with excessive frequency as a high-risk baseline; when calculating the causal inference strength, reduces the fusion weight of the tenant group behavior characteristics corresponding to the high-risk baseline.

[0016] A deep learning-based dynamic load balancing system for cloud storage resources includes: a data acquisition module for collecting access sequence data and spatiotemporal characteristic data of node load in a multi-tenant environment cloud storage cluster; a load identification and prediction module with a built-in multi-tenant load identification and prediction model for receiving data output from the data acquisition module and outputting hotspot prediction results, confidence scores, and anti-disturbance discrimination scores; and a behavior feature extraction module for constructing a multi-tenant association graph with tenants as nodes and the temporal synchronization degree of access behavior between tenants as edge weights, using a graph neural network to extract tenant group behavior features, and combining these features with the tenant's historical behavior baseline and access sequence temporal data. After feature fusion, the data is input into the structural causal model, which outputs the causal inference strength of the access sequence and hotspot events. The gating and scheduling module receives the confidence score of the hotspot prediction result, the adversarial disturbance discrimination score, and the causal inference strength. Based on the comparison between the weighted fusion result and the preset execution threshold, it determines the type of access sequence, performs data preheating and resource reservation operations on valid prediction sequences, and blocks the corresponding operations for malicious adversarial access sequences. The feedback and update module feeds the access sequence data corresponding to the malicious adversarial access sequence as negative samples to the multi-tenant load identification and prediction model and the structural causal model, respectively updating the parameters of the adversarial disturbance discrimination sub-network and the tenant historical behavior baseline.

[0017] The present invention has the following beneficial effects:

[0018] (1) The cloud storage resource dynamic load balancing method based on deep learning outputs the adversarial disturbance discrimination score through the adversarial disturbance discrimination sub-network, and presents the adversarial disturbance hidden in the access sequence in the form of explicit score. This enables the system to identify potential attack behaviors before the scheduling decision is executed. Combined with the judgment mechanism of the gating fusion unit, the corresponding data preheating and resource reservation operations are directly blocked for malicious adversarial access sequences. This avoids the encroachment of ineffective preheating on high-performance storage media and computing resources from the source, and ensures the available resources of normal tenants.

[0019] (2) The deep learning-based dynamic load balancing method for cloud storage resources constructs an association graph by utilizing the access time synchronization degree between tenants, and associates tenants whose access behavior has a synchronous relationship within the same time window. It assigns differentiated attention weights to different neighbor nodes through a graph attention network, extracts group behavior features that reflect the overall access status of the cluster, and inputs these features into a structural causal model after integrating them with the tenant's historical behavior baseline and access sequence time compliance features. It comprehensively judges the true situation of the access sequence from three dimensions: the degree of consistency of group behavior, the degree of continuity of historical behavior, and the degree of rationality of the time sequence pattern, so that the system has the ability of multi-source cross-validation when facing maliciously constructed adversarial access sequences.

[0020] (3) The deep learning-based cloud storage resource dynamic load balancing method adds a gating fusion unit between the prediction model and the scheduling execution unit. On the one hand, it dynamically adjusts the fusion weight according to the difference between the confidence score of the hot spot prediction result and the causal inference strength, so that the system can seek a dynamic balance between prediction and verification. On the other hand, it adds the adversarial disturbance discrimination score to the fusion result in the form of a penalty term. When the adversarial disturbance discrimination score exceeds the preset adversarial threshold, it is directly judged as a malicious adversarial access sequence and subsequent operations are blocked, thereby effectively avoiding the interference of single signal misjudgment on scheduling decision.

[0021] (4) The cloud storage resource dynamic load balancing system based on deep learning works in collaboration with load identification and prediction, behavior feature extraction, gating scheduling and feedback update modules to continuously collect access data that are judged as malicious adversarial access sequences during operation. It assigns priority to negative samples based on the degree of deviation between the adversarial disturbance discrimination score and the causal inference strength, and samples from the replay buffer for incremental training according to the priority. This enables the adversarial disturbance discrimination sub-network to continuously learn the characteristics of new deceptive access sequences. At the same time, the structural causal model records the historical frequency and time distribution of each tenant being judged as a malicious adversarial access sequence. The historical behavior baseline of tenants with excessive frequency is marked as a high-risk baseline. In the subsequent causal inference strength calculation, the fusion weight of the behavior characteristics of this type of tenant group is reduced. Through the closed-loop cooperation between the modules, the system's defense capability can be continuously adjusted with the changes in attack behavior.

[0022] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0023] Figure 1 This is a flowchart of a dynamic load balancing method for cloud storage resources based on deep learning, according to the present invention.

[0024] Figure 2 This is a flowchart illustrating the specific steps of optimizing parameters in the adversarial perturbation discrimination subnetwork of the present invention, which uses real hotspot features as positive samples and adversarial access sequence features constructed by a generative adversarial network as negative samples in a deep learning-based dynamic load balancing method for cloud storage resources.

[0025] Figure 3 This is a block diagram of a cloud storage resource dynamic load balancing system based on deep learning, according to the present invention. Detailed Implementation

[0026] Please see Figure 1 This invention provides a technical solution: a dynamic load balancing method for cloud storage resources based on deep learning, comprising the following steps: collecting access sequence data and spatiotemporal characteristic data of node load in a multi-tenant environment cloud storage cluster; inputting the data into a multi-tenant load identification and prediction model; obtaining hotspot prediction results, confidence scores of hotspot prediction results, and adversarial disturbance discrimination scores. The model includes a hotspot prediction subnetwork and an adversarial disturbance discrimination subnetwork. The adversarial disturbance discrimination subnetwork optimizes parameters using real hotspot features as positive samples and adversarial access sequence features constructed by a generative adversarial network as negative samples; constructing a multi-tenant association graph with tenants as nodes and temporal synchronization of access behavior between tenants as edge weights, using graph neural networks... The network extracts tenant group behavior characteristics, integrates these characteristics with tenant historical behavior baselines and access sequence time-series compliance characteristics, and inputs them into the structural causal model. The output is the causal inference strength between access sequences and hot events. The confidence score of hot event prediction results, adversarial disturbance discrimination score, and causal inference strength are input into the gating fusion unit. Based on the comparison between the weighted fusion result and the preset execution threshold, the access sequence type is determined. Data preheating and resource reservation operations are performed on valid prediction sequences, and corresponding operations are blocked for malicious adversarial access sequences. Such access sequences are fed back as negative samples to the multi-tenant load identification prediction model and the structural causal model to update the adversarial disturbance discrimination sub-network parameters and tenant historical behavior baselines, respectively.

[0027] The input data for the multi-tenant load identification and prediction model needs to undergo normalization preprocessing. The specific normalization formula is as follows:

[0028] ;

[0029] In the formula, x represents the original input data (access sequence data or spatiotemporal characteristic data of node load). The minimum value of this type of data. The maximum value of this type of data is given, and the normalized data ranges from [0,1].

[0030] The hotspot prediction subnetwork adopts a 3-layer temporal convolutional structure with kernel sizes of 3, 5 and 3 respectively, a stride of 1 for each layer, and the ReLU activation function. The output dimension is 1 (corresponding to the hotspot prediction result, where 1 indicates a predicted hotspot and 0 indicates a predicted non-hotspot).

[0031] The adversarial perturbation discriminant subnetwork adopts a 2-layer fully connected structure, with 128 neurons in the hidden layer and 1 neuron in the output layer (corresponding to the adversarial perturbation discrimination score; the closer the score is to 1, the higher the probability that the current accessed sequence is an adversarial sequence, and the closer it is to 0, the higher the probability that it is a real sequence).

[0032] The structural causal model consists of four fully connected layers with 256, 128, 64, and 1 neurons respectively. The output layer uses the Sigmoid activation function, with an output range of [0,1] (corresponding to the causal inference strength; the closer the strength is to 1, the stronger the causal correlation between the access sequence and the hot events; the closer it is to 0, the weaker the correlation).

[0033] Specifically, the steps for collecting access sequence data and spatiotemporal characteristic data of node load in a multi-tenant cloud storage cluster are as follows:

[0034] Collect tenant access sequence data, including request timestamps, request object identifiers, and request operation types, specifically:

[0035] The cloud storage cluster access log collector is used to collect access request data for each tenant in real time, with a collection frequency of 1 time / second.

[0036] The requested timestamp should be in UTC time format, accurate to milliseconds.

[0037] For example, if tenant i's access request timestamp is "2024-05-20 10:00:00.123", it is denoted as... ,in, (N is the total number of tenants, N≥2). ( (The total number of access requests for tenant i).

[0038] The request object identifier uses a unique encoding format of "tenant ID-file path-file name";

[0039] For example, the file accessed by tenant 1 is identified as "Tenant1: / data / docs / report.pdf", denoted as ;

[0040] The request operation types include read (denoted as 0), write (denoted as 1), delete (denoted as 2), and modify (denoted as 3);

[0041] For example, if tenant i's k-th request is a read operation, denoted as ;

[0042] The final access sequence data for tenant i is as follows:

[0043] ;

[0044] The "0" padding value for the request object identifier corresponds to the default encoding when there is no access request and does not correspond to any real file identifier.

[0045] Collect spatiotemporal load characteristic data of storage nodes, including CPU utilization, memory utilization, disk I / O latency, and network bandwidth utilization, specifically:

[0046] A node monitoring agent is used to collect real-time load data for each storage node in the cloud storage cluster. The collection frequency is the same as that of the access sequence data (1 time / second).

[0047] Let the total number of storage nodes be K, and the node number be K. The CPU utilization rate of node n collected at time t is denoted as . (Unit: %, value range [0, 100]);

[0048] For example, node 1 has a CPU utilization rate of 35.2% at time t=10:00:00, that is... ;

[0049] Memory usage is denoted as (Unit: %, value range [0, 100]), disk I / O latency is denoted as... (Unit: ms, value range [0, 1000]), network bandwidth utilization is denoted as... (Unit: Mbps, value range [0, 1000]);

[0050] The final spatiotemporal feature vector of the load at node n at time t is:

[0051] ;

[0052] Align the two types of data by timestamp and construct a multidimensional feature matrix, specifically as follows:

[0053] Based on the timestamp, the access sequence data of all tenants collected at the same time t are aligned with the spatiotemporal characteristics of the load of all storage nodes. For tenants without access requests, their access sequence data at the corresponding time is filled with (0,0,0) (timestamp is filled with 0, object identifier is filled with 0, and operation type is filled with 0).

[0054] For nodes for which no load data was collected, the load feature vector at the corresponding time point is filled with [0,0,0,0].

[0055] Let the time series length be T (i.e., the total number of time points collected), then the constructed multidimensional feature matrix is: ;

[0056] Each row of the matrix corresponds to all the data at a time t, and each column corresponds to a specific feature (time stamp of tenant access sequence, object identifier, operation type, or various indicators of node load).

[0057] For example, when N=2, K=3, and T=100, the dimension of the feature matrix is ​​100×(2×3+3×4)=100×18.

[0058] In this implementation plan, when collecting tenant access sequence data, the format and encoding rules of request timestamp, request object identifier, and request operation type are clearly defined, and uniform filling processing is performed on times without access requests, so that the access behaviors of different tenants have a basis for alignment in the same time dimension. When collecting storage node load spatiotemporal characteristic data, CPU utilization, memory utilization, disk I / O wait time, and network bandwidth utilization are obtained simultaneously, covering key indicators affecting node processing capabilities. After aligning the two types of data by timestamp, a multi-dimensional feature matrix is ​​constructed, and filling processing is performed on nodes for which load data has not been collected, ensuring the continuity and integrity of the time series.

[0059] Specifically, such as Figure 2 As shown, the specific steps for optimizing the parameters of the adversarial perturbation discriminant subnetwork, using real hotspot features as positive samples and adversarial access sequence features constructed by the generative adversarial network as negative samples, are as follows:

[0060] Using real hotspot features as anchor samples, historical normal access sequence features of the same tenant as positive samples, and adversarial access sequence features constructed by the generative adversarial network as negative samples, a triplet loss function is constructed, which is as follows:

[0061] Anchor point sample This is a feature vector representing the historical hotspot access sequence of the cloud storage cluster, with dimension d (d=3, corresponding to the feature encoding values ​​of timestamp, object identifier, and operation type, and the feature encoding is normalized).

[0062] For example, (Timestamp feature encoding 0.8, object identifier feature encoding 0.6, operation type feature encoding 0.2);

[0063] Positive sample p is the feature vector of the historical normal access sequence of the same tenant, with the same dimension as the anchor sample. For example, the features of the historical normal access sequence of tenant i. ;

[0064] The negative sample n is the feature vector of the adversarial access sequence constructed by the generative adversarial network (GAN), with the same dimension as the anchor sample;

[0065] For example, (Deliberately mimicking the characteristics of trending topics, with a high degree of similarity to real trending topics);

[0066] The formula for the triplet loss function is:

[0067] ;

[0068] In the formula, Describing the L2 norm, This is a marginal parameter with a value of 0.2, used to ensure that the distance between a positive sample and an anchor sample is less than the distance between a negative sample and an anchor sample by at least 0.2.

[0069] Both anchor samples and positive samples are real and valid sequence features, while negative samples are deliberately constructed adversarial sequence features, which are used to train the adversarial perturbation discriminant sub-network to identify adversarial sequences.

[0070] The three types of samples are input into the adversarial perturbation discriminant subnetwork and the corresponding feature vectors are extracted, specifically as follows:

[0071] Anchor sample a, positive sample p, and negative sample n are input into the adversarial discriminant sub-network, respectively. After passing through the first fully connected layer (input dimension d=3, output dimension 128), the ReLU activation function, and the second fully connected layer (input dimension 128, output dimension 64), feature vectors of the three classes of samples are extracted, denoted as follows: ;

[0072] For example, the feature vector extracted from anchor sample a is (Total 64 elements), the feature vector extracted from the positive sample p is The feature vector extracted from negative sample n is ;

[0073] The sub-network parameters are updated by reducing the distance between the anchor sample and the feature vector of the positive sample, and increasing the distance between the anchor sample and the feature vector of the negative sample. Specifically:

[0074] The stochastic gradient descent (SGD) optimization algorithm was adopted, with a learning rate of 0.001, momentum of 0.9, and 100 iterations. During each iteration, the triplet loss function was used. Calculate the gradient, and update the weights and biases of the fully connected layers of the adversarial discrimination subnetwork through gradient backpropagation;

[0075] Let the weight matrix of the fully connected layer be... ;

[0076] The bias vector is ;

[0077] The weight update formula is:

[0078] ;

[0079] The bias update formula is:

[0080] ;

[0081] In the formula, The learning rate;

[0082] Iterate until the loss function value converges (convergence condition is that the change in the loss function value is less than 1e-5 for 10 consecutive rounds), and complete the sub-network parameter update.

[0083] In this implementation scheme, a triplet loss function is constructed, using real hotspot features as anchor samples, historical normal access sequence features of the same tenant as positive samples, and adversarial access sequence features constructed by the generative adversarial network as negative samples. This allows the adversarial perturbation discriminant subnetwork to simultaneously face real effective sequences and constructed adversarial sequences during training. After inputting the three types of samples into the subnetwork, feature vectors are extracted through a fully connected layer. The gradient descent algorithm is used to iteratively update the network parameters, gradually narrowing the distance between the anchor samples and the positive sample feature vectors, and widening the distance between the anchor samples and the negative sample feature vectors. This enables the adversarial perturbation discriminant subnetwork to learn the distribution differences between adversarial access sequences and normal access sequences in the feature space.

[0084] Specifically, the steps for constructing a multi-tenant association graph with tenants as nodes and the synchronization degree of access behavior between tenants as edge weights are as follows:

[0085] The access sequence of each tenant is sliced ​​according to a time window, specifically as follows:

[0086] Set the time window length to The entire access time series is divided into multiple consecutive and non-overlapping time windows, denoted as seconds (adjustable according to the actual cluster size). ,in, , Total access time (in seconds). Indicates rounding down;

[0087] For example, if the total access time is 3600 seconds, it is divided into 60 time windows, with each window corresponding to 60 seconds of access data;

[0088] For each time window ( Extract the access sequence fragment for each tenant i within the window. ;

[0089] If tenant i has no access request within this window, then Empty;

[0090] The start time of the time window is calculated from the moment the data is first collected and accessed, ensuring the continuity of the time series.

[0091] Based on the peak value of the cross-correlation function of the tenant access timestamp sequence within the same time window, the synchronization coefficient of access time sequence among tenants is calculated, specifically as follows:

[0092] For the same time window Any two tenants i and j within ( Extract the access timestamp sequence of both. (m and n are the number of times tenants i and j access the window, respectively).

[0093] Calculate the cross-correlation function of two timestamp series:

[0094] ;

[0095] In the formula, ;

[0096] Tenants i and j are in the window respectively Average access timestamp within, For time delay (value range is) );

[0097] Take the peak value of the cross-correlation function After normalization, they are used as tenants i and j in the window. The timing synchronization coefficient within is:

[0098] ;

[0099] In the formula, The minimum value of the peak value of the cross-correlation function for all tenants. The maximum value, ;

[0100] For example, the peak value of the cross-correlation function between tenant 1 and tenant 2 in window W1 is 0.8. =0.2, =0.9, then the synchronization coefficient ;

[0101] The temporal synchronization coefficient is used as the initial weight of the corresponding edge in the association graph, specifically as follows:

[0102] Constructing a multi-tenant relationship graph , where the node set ( (corresponding tenant i), edge set (There is an edge between any two tenants), weight set ;

[0103] side initial weights This is the average of the synchronization coefficients across all time windows, i.e.:

[0104] ;

[0105] For example, if the average synchronization coefficient of tenant 1 and tenant 2 across 60 windows is 0.78, then the initial weight of edge (v1, v2) is 0.78.

[0106] If two tenants do not access any windows simultaneously (i.e., all...) If the initial weight is 0, then the initial weight is set to 0.

[0107] A graph attention network is used to perform embedding learning on the association graph, assigning differentiated attention weights to different neighbor nodes, specifically as follows:

[0108] The Graph Attention Network (GAT) consists of one graph attention layer and one output layer. The input is an initial feature vector for each node (3 dimensions, representing the tenant's access frequency, average access interval, and operation type percentage; access frequency is the number of accesses per unit time; average access interval is the average time difference between two adjacent accesses; and operation type percentage is the percentage of requests for each type of operation (read, write, delete, modify)). This is denoted as... ;

[0109] For nodes Each neighbor node Calculate the attention coefficient:

[0110] ;

[0111] In the formula, For the characteristic transformation matrix ( (for the transformed feature dimensions) For attention weight vectors, For nodes The set of neighboring nodes (i.e., all other tenant nodes). This indicates feature concatenation, with LeakyReLU as the activation function (negative slope set to 0.2).

[0112] For example, if node v1's neighboring nodes are v2 and v3, the calculation is as follows: That is, the attention weight of v2 is 0.6, and the attention weight of v3 is 0.4;

[0113] The higher the attention coefficient, the greater the impact of the behavior of the corresponding neighboring nodes on the current tenant node.

[0114] The behavioral features of the tenant group are extracted through the output layer of the graph attention network, specifically as follows:

[0115] Each node The features of neighboring nodes are summed by attention weights to obtain the node. Embedding features:

[0116] ;

[0117] In the formula, Use the Sigmoid activation function;

[0118] The embedded features are input to the output layer (fully connected layer, input dimension 16, output dimension 8) to obtain the group behavior feature vector of tenant i. ;

[0119] For example, the group behavior feature vector of tenant 1 is This vector reflects the correlation characteristics of tenant i's access behavior in a multi-tenant group.

[0120] In this implementation, the continuous access sequence is divided into discrete windows by slicing time windows. Within each window, the temporal synchronization coefficient is calculated based on the peak value of the cross-correlation function of the tenant access timestamp sequence. The synchronization degree of access behavior between tenants is quantified as edge weights. A multi-tenant association graph is constructed with tenants as nodes and the temporal synchronization degree of access behavior between tenants as edge weights. A graph attention network is used to perform embedding learning on the association graph, and differentiated attention weights are assigned to different neighbor nodes. This enables the model to identify neighbor nodes that have a greater impact on the current tenant's behavior and assign higher weights to tenants with higher association. The group behavior features of tenants are extracted through the output layer. The access behavior of a single tenant is analyzed from the perspective of the group interaction of multiple tenants. This provides a group behavior dimension for verification for subsequent structural causal models to determine the causal relationship between access sequences and hot events.

[0121] Specifically, the steps for fusing this feature with the tenant's historical behavior baseline and access sequence time-series compliance features and inputting it into the structural causal model are as follows:

[0122] The baseline of tenant historical behavior is input into the structural causal model, and after mapping through a fully connected layer, long-term access preference features are output, specifically:

[0123] Tenant i's historical behavior baseline is access sequence statistics over the past 30 days, including historical access frequency. (Unit: times / hour), historical peak access times (For example, [9:00-11:00, 14:00-16:00]), percentage of historical access object types (These correspond to the percentages of read, write, delete, and modify operations, respectively.) );

[0124] Encode historical behavioral baselines into feature vectors:

[0125] ;

[0126] in, The start and end times of peak periods are encoded by normalizing a 24-hour day to [0,1].

[0127] For example, 9:00 corresponds to 0.375 (9 / 24), 11:00 corresponds to 0.458 (11 / 24), 14:00 corresponds to 0.583 (14 / 24), and 16:00 corresponds to 0.667 (16 / 24), with a value range of [0,1].

[0128] The feature vector is input into the first fully connected layer of the structural causal model (input dimension 7, output dimension 16), and the mapping formula is as follows:

[0129] ;

[0130] In the formula, This is the weight matrix. For bias vectors, This refers to the long-term access preference characteristics of tenant i;

[0131] For example, the historical behavior baseline feature vector of tenant i:

[0132] ;

[0133] After mapping, we get:

[0134] (A total of 16 elements);

[0135] Long-term access preference features are used to reflect tenants’ long-term access habits and provide long-term behavioral references for causal inference.

[0136] The access sequence temporal compliance features are input into the structural causal model, and after mapping through a fully connected layer, a short-time pattern matching score is output, which is as follows:

[0137] Access sequence temporal compliance features are used to measure the degree of matching between the temporal distribution of the current access sequence and the normal temporal pattern. The calculation indicators include the variance of the current access interval. ( (Average access interval of tenant i) and matching degree during the current peak access period (1 if the current access time is within the historical peak period, 0 otherwise), Operation type sequence similarity (Cosine similarity between the current operation type sequence and the historical operation type sequence, with values ​​ranging from [0,1]).

[0138] Encode time-series compliance features into vectors The second fully connected layer of the input structural causal model (input dimension 3, output dimension 8) has the following mapping formula:

[0139] ;

[0140] In the formula, This is the weight matrix. For bias vectors, This refers to short-term pattern matching scoring;

[0141] For example, the time-series compliance feature vector of tenant i After mapping, we get (A total of 8 elements);

[0142] The group behavior characteristics are input into the structural causal model, and the output, after mapping through a fully connected layer, is the tenant cluster anomaly deviation degree, which is as follows:

[0143] The group behavior feature vector of tenant i is (Output from the graph attention network), calculate the average of this feature and the behavioral features of all tenant groups. Euclidean distance The distance is normalized and used as the base value for the anomaly deviation. ( (The maximum value of the base value of abnormal deviation for all tenants);

[0144] Will and spliced ​​as (Dimension 9), the third fully connected layer of the input structure causal model (input dimension 9, output dimension 8), the mapping formula is:

[0145] ;

[0146] In the formula, This is the weight matrix. For bias vectors, This refers to the abnormal deviation of the tenant cluster;

[0147] For example, tenant i's , After splicing, the mapping is obtained ;

[0148] The above three types of features are concatenated to form a fused feature vector, which is as follows:

[0149] Long-term access preference characteristics Short-term pattern matching score Tenant cluster anomaly deviation By concatenating them in order, we obtain the fused feature vector. ;

[0150] For example, It is a 16-dimensional vector. It is an 8-dimensional vector. The 8-dimensional vectors are concatenated to obtain a 32-dimensional fused feature vector. ;

[0151] The fused feature vector is input into the output layer of the structural causal model to obtain the causal inference strength, specifically:

[0152] fuse feature vectors The fourth fully connected layer of the input structural causal model (input dimension 32, output dimension 1) uses the Sigmoid activation function and outputs the causal inference strength:

[0153] ;

[0154] In the formula, This is the weight matrix. For bias vectors, ;

[0155] The closer to 1, the stronger the causal relationship between the current access sequence and the hot events; the closer to 0, the weaker the causal relationship.

[0156] For example, after fusing the feature vector input, we get This indicates a strong causal relationship between the current access sequence and hot events.

[0157] In this implementation, the historical behavior baseline of tenants is mapped to long-term access preference features through a fully connected layer, the temporal compliance features of access sequences are mapped to short-term pattern matching scores through a fully connected layer, and the group behavior features are mapped to tenant cluster anomaly deviation degree through a fully connected layer. The three types of features are concatenated and input into the output layer of the structural causal model to obtain the causal inference strength. Through a hierarchical mapping method, features are extracted from three dimensions: long-term historical habits, short-term temporal patterns, and group interaction relationships. These features are then fused to form a comprehensive judgment basis, so that the causal inference strength can comprehensively reflect the strength of the causal relationship between access sequences and hot events. This provides the gating fusion unit with an independent verification signal that is different from the confidence score of the hot event prediction result, and solves the problem that the output of a single prediction model lacks multi-source verification.

[0158] Specifically, the steps for determining the access sequence type based on the comparison between the weighted fusion result and the preset execution threshold are as follows:

[0159] The numerical difference between the confidence score of the hotspot prediction result and the strength of causal inference is calculated as follows:

[0160] Let the confidence score of the hotspot prediction result be . (Output from the multi-tenant load identification and prediction model; the closer to 1, the more reliable the hotspot prediction result). The causal inference strength is... The formula for calculating the numerical difference is:

[0161] ;

[0162] For example, ,but ;

[0163] like ,but;

[0164] The fusion weights of the two factors are dynamically adjusted based on the magnitude of the difference. When the difference exceeds a set range, the weight of the causal inference strength is increased; conversely, the weight of the confidence score of the hotspot prediction result is increased. Specifically:

[0165] Set upper limit for difference Lower limit of difference The fusion weights include the weights of the confidence scores of the hotspot prediction results. Weights of causal inference strength ,and ;

[0166] The adjustment rules are as follows:

[0167] when hour, (Increase the weighting of causal inference strength);

[0168] when hour, (Increase the weighting of confidence scores in hotspot prediction results);

[0169] when hour, (Equal weights);

[0170] For example, ,but ;

[0171] ,but ;

[0172] The adversarial perturbation discrimination score is added to the weighted fusion result as a penalty term, specifically as follows:

[0173] Let the discriminant score against disturbances be denoted as . (Output from the multi-tenant load identification and prediction model; the closer to 1, the higher the probability that the current access sequence is an adversarial sequence), the penalty term coefficient is... (This can be adjusted according to actual defense needs), the formula for calculating the weighted fusion result is:

[0174] ;

[0175] For example, ,but ;

[0176] like ,but ;

[0177] When the adversarial discrimination score exceeds the preset adversarial threshold, the corresponding access sequence is directly determined to be a malicious adversarial access sequence, specifically as follows:

[0178] Preset adversarial threshold (Confirmed through extensive experimental verification, subject to adjustment);

[0179] when At that time, regardless of the weighted fusion result The size directly determines whether the current access sequence is a malicious and adversarial access sequence;

[0180] For example, ,even though It is still judged as a malicious adversarial access sequence;

[0181] when Then, proceed to the next step of judgment based on the weighted fusion results;

[0182] Based on the relationship between the weighted fusion result and the preset execution threshold, a distinction is made between valid prediction sequences and malicious adversarial access sequences, specifically as follows:

[0183] Preset execution threshold (Through experiments, it was determined that a balance between resource utilization and defensive effectiveness can be achieved.)

[0184] when and When the current access sequence is determined to be a valid prediction sequence, that is, the access sequence corresponds to a real hot event, data preheating and resource reservation operations need to be performed;

[0185] when and If the current access sequence is determined to be a malicious and adversarial access sequence, the corresponding operation must be blocked.

[0186] For example, It was determined to be a valid prediction sequence;

[0187] It was determined to be a malicious, adversarial access sequence.

[0188] In this implementation scheme, the system calculates the numerical difference between the confidence score of the hotspot prediction result and the strength of causal inference, and dynamically adjusts the fusion weight of the two based on the magnitude of the difference. This allows the system to maintain balance when the prediction result is consistent with the verification result, and to tilt towards the more reliable signal when there is a deviation. The adversarial disturbance discrimination score is added to the weighted fusion result as a penalty term. When the adversarial disturbance discrimination score exceeds the preset adversarial threshold, it is directly judged as a malicious adversarial access sequence and subsequent operations are blocked. This multi-level judgment mechanism, which does not depend on the numerical value of the weighted fusion result, ultimately distinguishes between valid prediction sequences and malicious adversarial access sequences based on the comparison result of the weighted fusion result and the preset execution threshold. This multi-level judgment mechanism organically combines prediction signals, causal verification signals, and adversarial discrimination signals, enabling scheduling decisions to have the ability of multi-source verification and rapid response, and avoiding the impact of misjudgment of a single signal on system scheduling.

[0189] Specifically, the steps for performing data preheating and resource reservation operations on effective predicted sequences are as follows:

[0190] Based on the hotspot prediction results, the identifiers of data blocks to be preheated and the reserved capacity of high-performance storage media are determined, specifically as follows:

[0191] The hotspot prediction results include a set of predicted hotspot data block identifiers. (m is the number of hot data blocks);

[0192] For example, ;

[0193] Calculate the size of each hot data block (Unit: GB)

[0194] For example, ;

[0195] Reserved capacity for high-performance storage media (such as SSDs):

[0196] ;

[0197] In the formula, To reserve a redundancy factor to cope with fluctuations in data block size;

[0198] For example, ;

[0199] The process of migrating the data block to be preheated from the low-frequency storage medium to the high-frequency storage medium is as follows:

[0200] Low-frequency storage media are mechanical hard disk drives (HDDs), and high-frequency storage media are solid-state drives (SSDs).

[0201] The migration uses a parallel migration method, with the number of migration threads set to [value missing]. (Maximum 8 threads to avoid consuming too many network resources);

[0202] Monitor migration progress in real time during the migration process:

[0203] ;

[0204] In the formula, The size of data block k that has been migrated;

[0205] When the migration progress of all data blocks At that time, data preheating is completed;

[0206] For example, the migration progress of data block 1 gradually increases from 0% to 100%. After the migration is completed, the data block is migrated from HDD to SSD.

[0207] Within the high-performance storage medium, independent cache regions are partitioned and bound to corresponding tenant access sequences, specifically as follows:

[0208] In the SSD, an independent cache region is allocated for each tenant i. Cache range size:

[0209] ;

[0210] In the formula, The total size of the hotspot data blocks corresponding to tenant i;

[0211] For example, the total size of hot data blocks for tenant 1 is 2GB, for tenant 2 it is 5GB, and the total reserved capacity is 8.4GB;

[0212] Then the size of tenant 1's cache range ;

[0213] Tenant 2's cache range size ;

[0214] cache range Access sequence with tenant i Binding means that only the access request of tenant i can access the data block in this cache range, so as to avoid cache conflicts between different tenants;

[0215] When the confidence score of the hotspot prediction result is lower than the set migration threshold, the migration operation from the low-frequency storage medium to the high-frequency storage medium is triggered only when the data block is accessed for the first time. Specifically:

[0216] Set migration threshold ;

[0217] when When the reliability of hotspot prediction results is low, data preheating migration is not performed in advance. Instead, real-time migration is triggered when the data block is accessed by the tenant for the first time.

[0218] After the migration is complete, the data blocks will be stored in the corresponding tenant's cache area;

[0219] For example, When tenant i accesses data block obj1 for the first time, it triggers a real-time migration of obj1 from HDD to SSD. After the migration is completed, subsequent accesses will read directly from the SSD cache.

[0220] The first access refers to the tenant's first access request after the data block is determined to be a valid prediction sequence.

[0221] In this implementation plan, the identifiers of data blocks to be preheated and the reserved capacity of high-performance storage media are determined based on the hotspot prediction results. By statistically analyzing the size of hotspot data blocks and setting a reserved redundancy coefficient, the reserved capacity can cover the fluctuation range of data blocks. The data blocks to be preheated are migrated from low-frequency storage media to high-frequency storage media. A parallel migration method is used to shorten the data preparation time. Within the high-performance storage media, an independent cache interval is allocated for each tenant and bound to the corresponding tenant access sequence to ensure that the cache spaces of different tenants are isolated from each other, avoiding cache conflicts and resource contention. When the confidence score of the hotspot prediction result is lower than the set migration threshold, a lazy migration strategy is adopted, and the migration operation is triggered only when the data block is accessed for the first time, avoiding the resource overhead caused by performing invalid migration when the prediction reliability is insufficient.

[0222] Specifically, the steps for feeding this type of access sequence as a negative sample into the multi-tenant load identification and prediction model are as follows:

[0223] Based on the deviation between the adversarial perturbation discrimination score and the causal inference strength of the malicious adversarial access sequence, a negative sample priority is set, specifically as follows:

[0224] Negative sample priority The formula used to measure the contribution of a negative sample to model optimization is as follows:

[0225] ;

[0226] In the formula, To counteract disturbances, a discrimination score is set (the higher the score, the stronger the countermeasure). The strength of causal inference (the smaller the value, the weaker the correlation with trending events). ;

[0227] The higher the priority, the greater the probability that the sample will be sampled during model training;

[0228] For example, ,but (High priority);

[0229] ,but (Medium priority);

[0230] ,but (Low priority);

[0231] Negative samples are stored in the replay buffer in descending order of priority, specifically as follows:

[0232] Set the maximum capacity of the playback buffer to (This can be adjusted according to model training requirements). The buffer uses a priority queue to store negative samples, and each negative sample contains malicious adversarial access sequence data. Disturbance discrimination score Strength of causal inference and priority ;

[0233] When the buffer is not full, new negative samples are directly inserted into the queue in descending order of priority;

[0234] When the buffer is full, delete the lowest priority sample and then insert a new sample;

[0235] For example, if the buffer is full, the new sample has a priority of 1.6, and the lowest priority in the buffer is 0.7, then the sample with priority 0.7 is deleted, and the new sample is inserted.

[0236] During the incremental training phase of the model, negative samples are drawn from the buffer according to priority, specifically as follows:

[0237] The incremental training cycle for the multi-tenant load identification and prediction model is set to 1 hour. During each training session, negative samples are sampled from the replay buffer according to priority, with the number of samples... (Maximum of 100 samples per sampling session);

[0238] The sampling uses a weighted random sampling method, and the probability of a sample being sampled is:

[0239] ;

[0240] In the formula, The set of negative samples in the buffer;

[0241] For example, if there are 3 samples in the buffer with priorities of 1.6, 1.1, and 0.7, and the total priority is 3.4, then the sampling probability of the sample with priority 1.6 is 1.6 / 3.4≈0.47, the sampling probability of the sample with priority 1.1 is 1.1 / 3.4≈0.32, and the sampling probability of the sample with priority 0.7 is 0.7 / 3.4≈0.21.

[0242] The extracted samples are input into the adversarial perturbation discriminant subnetwork, the triplet loss function is calculated, and the network parameters are updated, specifically as follows:

[0243] The sampled negative samples are used as negative samples n in the triplet loss function, the anchor sample a is still the real hotspot feature, and the positive sample p is the historical normal access sequence feature of the same tenant.

[0244] The loss function value for each training iteration is calculated using the triplet loss function calculation method for parameter updates in the adversarial perturbation discriminant subnetwork.

[0245] The Adam optimization algorithm (learning rate 0.0001, decay coefficient 0.9, where the decay coefficient refers to the momentum decay coefficient) is used to update the weights and biases of the fully connected layers of the adversarial discriminant subnetwork through gradient backpropagation.

[0246] Each incremental training iteration consists of 50 rounds until the loss function value converges (convergence condition is that the loss change is less than 1e-6 for 5 consecutive rounds), thus completing the network parameter update.

[0247] For example, after sampling 100 negative samples and training for 50 rounds, the loss function value converges from 0.3 to 0.05, completing the parameter update;

[0248] Incremental training only updates the adversarial perturbation discriminant subnetwork and does not affect the parameters of the hotspot prediction subnetwork, thus ensuring the stability of model predictions.

[0249] In this implementation scheme, negative samples are prioritized based on the deviation between the adversarial perturbation discrimination score and the causal inference strength of the malicious adversarial access sequence. Samples with high deception level and weak correlation with hot events are given higher priority. Negative samples are stored in the replay buffer in descending order of priority. When the buffer capacity is limited, high-priority samples are retained and low-priority samples are discarded to ensure that valuable historical attack samples are continuously retained. During the incremental training phase, negative samples are extracted from the buffer according to priority. High-priority samples have a higher probability of being sampled, enabling the adversarial perturbation discrimination subnetwork to learn typical attack patterns first under limited training resources. The extracted samples are input into the adversarial perturbation discrimination subnetwork to calculate the triplet loss function and update the network parameters. Incremental training only updates the adversarial perturbation discrimination subnetwork without affecting the hotspot prediction subnetwork, so that the model's defense capability continues to evolve with the emergence of new attack patterns.

[0250] Specifically, the steps for feeding this type of access sequence as a negative sample back into the structural causal model are as follows:

[0251] The structural causal model records the frequency and temporal distribution of determinations of malicious adversarial access sequences by tenants, specifically as follows:

[0252] The structural causal model has a built-in logging module that records in real time the number of times each tenant i is identified as having a malicious adversarial access sequence. and the timestamp of each determination ;

[0253] For example, if tenant 1 is identified as having a malicious adversarial access sequence 3 times within 1 hour, with timestamps of 10:05:12, 10:20:35, and 10:45:21, then... ;

[0254] Simultaneously record the frequency of judgments within each unit time window (set to 10 minutes). (s is the time window number);

[0255] For example, the frequency of tenant 1's judgment within the 10:00-10:10 window is 1;

[0256] The value in the window between 10:10 and 10:20 is 0;

[0257] The value in the window is 1 from 10:20 to 10:30;

[0258] The value in the window is 0 between 10:30 and 10:40.

[0259] The value in the window is 1 between 10:40 and 10:50.

[0260] The determination of whether the frequency of malicious tenant identifications exceeds a set threshold within a unit time window is as follows:

[0261] Set the frequency threshold for a unit time window (10 minutes) (Can be adjusted according to cluster security requirements);

[0262] For each tenant i, its statistics are collected in each time window. Frequency of judgment within ;

[0263] like If so, it is determined that the tenant has engaged in malicious access behavior within the window;

[0264] For example, if the frequency of tenant 2's access during the 10:00-10:10 window is 3 > 2, it is determined to be a malicious access cluster.

[0265] like If so, it is considered a normal fluctuation;

[0266] Tenant historical behavior baselines that exceed the frequency limit are marked as high-risk baselines, specifically as follows:

[0267] If the frequency of judgments for tenant i exceeds the threshold in three or more consecutive time windows. If a tenant's historical behavior baseline is set if five or more windows exceed the limit within one hour, the tenant's historical behavior baseline will be reset. Baselines marked as high-risk are denoted as ;

[0268] For example, if tenant 3's judgment frequency is 3>2 in three consecutive windows of 10:00-10:10, 10:10-10:20, and 10:20-10:30, then its historical behavior baseline is marked as a high-risk baseline;

[0269] The high-risk baseline will serve as an important reference for subsequent causal inference strength calculations. Once marked, it will remain in effect until the tenant does not exceed the limit in any unit time window for 24 consecutive hours, at which point the high-risk mark will be automatically removed.

[0270] The unit time window is consistent with the set 10-minute window;

[0271] When calculating the strength of causal inference, the fusion weight of the tenant group behavior characteristics corresponding to the high-risk baseline is reduced, specifically as follows:

[0272] For tenant i whose historical behavioral baseline is marked as a high-risk baseline, their group behavioral characteristics are reduced when calculating the fused feature vector in the structural causal model. The fusion weights;

[0273] Let the normal weights be... (Corresponding to the weight of long-term access preference features) Short-time pattern matching scoring weight The weighting of group behavior characteristics of high-risk tenants was adjusted to... At the same time, increase the weight of long-term access preference features to Short-time pattern matching score weighting remains ;

[0274] The adjusted fusion feature vector is calculated as follows:

[0275] ;

[0276] Then the strength of causal inference is calculated. To reduce the impact of the behavioral characteristics of high-risk tenant groups on causal inference and prevent their malicious behavior from interfering with hotspot judgment;

[0277] For example, tenant 3 is a high-risk tenant. ;

[0278] but .

[0279] In this implementation, the structural causal model records the frequency and time distribution of tenants' malicious adversarial access sequences. By comparing the frequency of determination within a unit time window with a set threshold, it identifies tenants with concentrated malicious access behavior. The historical behavior baselines of tenants with excessive frequency are marked as high-risk baselines, enabling the system to identify tenants with persistent malicious behavior characteristics. In subsequent causal inference strength calculations, the fusion weight of the tenant group behavior characteristics corresponding to the high-risk baseline is reduced, while the weight of long-term access preference characteristics is increased. This reduces interference from the behavior characteristics of high-risk tenant groups when the structural causal model judges the causal relationship between access sequences and hot events. This mechanism allows the structural causal model to adaptively adjust the reliability of verification signals based on the actual behavior of tenants, avoiding the impact of a single tenant's malicious behavior on the overall judgment.

[0280] Please see Figure 3This invention provides a technical solution: a deep learning-based dynamic load balancing system for cloud storage resources, comprising: a data acquisition module for acquiring access sequence data and spatiotemporal characteristic data of node load in a multi-tenant environment cloud storage cluster; a load identification and prediction module, which has a built-in multi-tenant load identification and prediction model, for receiving data output from the data acquisition module and outputting hotspot prediction results, hotspot prediction result confidence scores, and anti-disturbance discrimination scores; and a behavior feature extraction module for constructing a multi-tenant association graph with tenants as nodes and the temporal synchronization degree of access behavior between tenants as edge weights, using a graph neural network to extract tenant group behavior features, and comparing these features with the tenant's historical behavior baseline. The system integrates access sequence time-series compliance features and inputs them into the structural causal model, outputting the causal inference strength between the access sequence and hot events. A gating and scheduling module receives the confidence score of hotspot prediction results, the adversarial discrimination score, and the causal inference strength. Based on the comparison between the weighted fusion result and the preset execution threshold, it determines the access sequence type, performs data preheating and resource reservation operations on valid prediction sequences, and blocks corresponding operations for malicious adversarial access sequences. A feedback and update module feeds the access sequence data corresponding to malicious adversarial access sequences as negative samples back to the multi-tenant load identification and prediction model and the structural causal model, respectively updating the adversarial discrimination sub-network parameters and the tenant historical behavior baseline.

[0281] The modules interact with each other via a high-speed data bus with a data transmission rate of no less than 10Gbps to ensure real-time performance.

[0282] The data acquisition module includes an access log collector and a node monitoring agent. The access log collector is deployed on the gateway node of the cloud storage cluster and is responsible for collecting access request data from all tenants. The node monitoring agent is deployed on each storage node and is responsible for collecting load data from the corresponding node. Both adopt a lightweight design, occupying no more than 5% of the node's CPU resources and no more than 100MB of memory resources.

[0283] The load identification and prediction module is deployed on edge computing nodes and uses GPU-accelerated computing (GPU model is NVIDIA A100). The inference latency is no more than 50ms, which meets the real-time load scheduling requirements of cloud storage clusters.

[0284] The graph neural network for the behavior feature extraction module is implemented using the PyTorch framework. After training, it is fixed into a model file and loaded into memory during deployment. It supports online inference with an inference latency of no more than 100ms.

[0285] The gating and scheduling module has a built-in configuration interface for parameters such as preset execution thresholds and adversarial thresholds, which can dynamically adjust parameter values ​​according to the scale of the cloud storage cluster (number of tenants, number of storage nodes);

[0286] For example, when the number of tenants exceeds 1000, the threshold will be enforced. Adjusted to 0.65, execution threshold Adjusted to 0.55;

[0287] The feedback update module adopts an incremental update mechanism. Each update only modifies some parameters of the model without retraining the entire model. The update time is no more than 10 seconds, thus avoiding affecting the normal operation of the cloud storage cluster.

[0288] In addition, each module has a fault self-check function with a self-check frequency of once per minute. When a module experiences faults such as abnormal data transmission or inference failure, an alarm mechanism is automatically triggered (alarm methods include background pop-ups and email notifications), and the system switches to a backup module to ensure continuous and stable operation. The backup module and the main module adopt a hot backup mode, and the switching delay does not exceed 100ms.

[0289] The deployment location of each module can be flexibly adjusted according to the scale of the cloud storage cluster. Small-scale clusters can deploy multiple modules on the same server, while large-scale clusters adopt distributed deployment.

[0290] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0291] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A deep learning-based cloud storage resource dynamic load balancing method, characterized in that, Includes the following steps: Access sequence data and spatiotemporal characteristic data of node load of cloud storage cluster in multi-tenant environment are collected and input into multi-tenant load identification and prediction model to obtain hot spot prediction results, hot spot prediction result confidence scores and adversarial disturbance discrimination scores. The model includes a hot spot prediction sub-network and an adversarial disturbance discrimination sub-network. The adversarial disturbance discrimination sub-network optimizes parameters with real hot spot features as positive samples and adversarial access sequence features constructed by generative adversarial network as negative samples. Construct a multi-tenant association graph with tenants as nodes and the temporal synchronization degree of access behavior between tenants as edge weights. Use graph neural networks to extract tenant group behavior features. After fusing these features with the tenant historical behavior baseline and access sequence temporal compliance features, input them into the structural causal model and output the causal inference strength between access sequence and hot events. The confidence score of hotspot prediction, the adversarial discrimination score, and the causal inference strength are input into the gating fusion unit. The access sequence type is determined based on the comparison between the weighted fusion result and the preset execution threshold. Data preheating and resource reservation operations are performed on valid prediction sequences, and corresponding operations are blocked for malicious adversarial access sequences. Such access sequences are fed back as negative samples to the multi-tenant load identification prediction model and the structural causal model to update the parameters of the adversarial discrimination sub-network and the baseline of tenant historical behavior, respectively. 2.The deep learning based cloud storage resource dynamic load balancing method of claim 1, wherein, The specific steps for collecting access sequence data and spatiotemporal characteristic data of node load in a multi-tenant cloud storage cluster are as follows: Collect tenant access sequence data, including request timestamp, request object identifier, and request operation type; Collect spatiotemporal load characteristics data of storage nodes, including CPU utilization, memory utilization, disk I / O latency, and network bandwidth utilization; Align the two types of data by timestamp to construct a multidimensional feature matrix. 3.The deep learning based cloud storage resource dynamic load balancing method of claim 1, wherein, The specific steps for optimizing the parameters of the adversarial perturbation discriminant subnetwork, using real hotspot features as positive samples and adversarial access sequence features constructed by the generative adversarial network as negative samples, are as follows: A triplet loss function is constructed using real hotspot features as anchor samples, historical normal access sequence features of the same tenant as positive samples, and adversarial access sequence features constructed by generative adversarial networks as negative samples. The three types of samples are input into the adversarial perturbation discriminant subnetwork and the corresponding feature vectors are extracted; The subnetwork parameters are updated by reducing the distance between the anchor sample and the positive sample feature vector and increasing the distance between the anchor sample and the negative sample feature vector.

4. The deep learning-based dynamic load balancing method for cloud storage resources according to claim 1, characterized in that, The specific steps for constructing a multi-tenant association graph with tenants as nodes and the temporal synchronization degree of access behavior between tenants as edge weights are as follows: Slice the access sequence of each tenant according to the time window; Based on the peak value of the cross-correlation function of the tenant access timestamp sequence within the same time window, the access timing synchronization coefficient between tenants is calculated. Use the temporal synchronization coefficient as the initial weight of the corresponding edge in the association graph; A graph attention network is used to perform embedding learning on the association graph, assigning differentiated attention weights to different neighbor nodes; The behavioral characteristics of tenant groups are extracted using the output layer of a graph attention network.

5. The cloud storage resource dynamic load balancing method based on deep learning according to claim 1, characterized in that, The specific steps for fusing this feature with the tenant's historical behavior baseline and access sequence time-series compliance features and inputting it into the structural causal model are as follows: Input the tenant's historical behavior baseline into the structural causal model, and output long-term access preference features through a fully connected layer; The access sequence temporal compliance features are input into the structural causal model, and the short-time pattern matching score is output after mapping through a fully connected layer. The group behavior characteristics are input into the structural causal model, and the tenant cluster anomaly deviation is output after mapping through a fully connected layer. The above three types of features are combined to form a fused feature vector; By inputting the fused feature vectors into the output layer of the structural causal model, the causal inference strength is obtained.

6. The cloud storage resource dynamic load balancing method based on deep learning according to claim 1, characterized in that, The specific steps for determining the access sequence type based on the comparison between the weighted fusion result and the preset execution threshold are as follows: Calculate the numerical difference between the confidence score of the hotspot prediction result and the strength of causal inference; The fusion weights of the two are dynamically adjusted based on the magnitude of the difference. When the difference exceeds the set range, the weight of the causal inference strength is increased, and vice versa, the weight of the confidence score of the hot spot prediction result is increased. The adversarial perturbation discrimination score is added to the weighted fusion result as a penalty term; When the adversarial discrimination score exceeds the preset adversarial threshold, the corresponding access sequence is directly determined to be a malicious adversarial access sequence; Based on the relationship between the weighted fusion result and the preset execution threshold, effective prediction sequences and malicious adversarial access sequences are distinguished.

7. The deep learning-based dynamic load balancing method for cloud storage resources according to claim 1, characterized in that, The specific steps for performing data warm-up and resource reservation operations on the effective predicted sequence are as follows: Based on the hotspot prediction results, determine the identifiers of the data blocks to be preheated and the reserved capacity of the high-performance storage media; Migrate the data blocks to be preheated from low-frequency storage media to high-frequency storage media; Divide the high-performance storage medium into independent cache regions and bind them to the corresponding tenant access sequences; When the confidence score of the hotspot prediction result is lower than the set migration threshold, the migration operation from the low-frequency storage medium to the high-frequency storage medium is triggered only when the data block is accessed for the first time.

8. The cloud storage resource dynamic load balancing method based on deep learning according to claim 1, characterized in that, The specific steps for feeding this type of access sequence as a negative sample into the multi-tenant load identification and prediction model are as follows: Based on the degree of deviation between the adversarial perturbation discrimination score and the causal inference strength of the malicious adversarial access sequence, the negative sample priority is set; Negative samples are stored in the replay buffer in descending order of priority. During the incremental training phase of the model, negative samples are drawn from the buffer according to priority. The extracted samples are input into the adversarial perturbation discriminant subnetwork, the triplet loss function is calculated, and the network parameters are updated.

9. The cloud storage resource dynamic load balancing method based on deep learning according to claim 1, characterized in that, The specific steps for feeding this type of access sequence as a negative sample back into the structural causal model are as follows: The structural causal model records the frequency and temporal distribution of determinations of malicious adversarial access sequences by tenants; Determine whether the frequency of malicious judgments against tenants exceeds a set threshold within a unit time window; Mark the historical behavior baselines of tenants that exceed the frequency limit as high-risk baselines; When calculating the strength of causal inference, the fusion weight of the tenant group behavior characteristics corresponding to the high-risk baseline is reduced.

10. A deep learning-based dynamic load balancing system for cloud storage resources, employing the deep learning-based dynamic load balancing method for cloud storage resources as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect access sequence data and spatiotemporal characteristic data of node load in a multi-tenant environment cloud storage cluster. The load identification and prediction module has a built-in multi-tenant load identification and prediction model. It is used to receive data output from the data acquisition module and output hot spot prediction results, hot spot prediction result confidence scores and anti-disturbance discrimination scores. The behavioral feature extraction module is used to construct a multi-tenant association graph with tenants as nodes and the temporal synchronization degree of access behavior between tenants as edge weights. It uses a graph neural network to extract the behavioral features of tenant groups, and then integrates these features with the historical behavior baseline of tenants and the temporal compliance features of access sequences into the structural causal model. The output is the causal inference strength between access sequences and hot events. The gating and scheduling module is used to receive the confidence score of hotspot prediction results, the anti-disturbance discrimination score, and the causal inference strength. Based on the comparison between the weighted fusion result and the preset execution threshold, it determines the access sequence type, performs data preheating and resource reservation operations on valid prediction sequences, and blocks the corresponding operations on malicious and anti-disturbance access sequences. The feedback update module is used to feed the access sequence data corresponding to the malicious adversarial access sequence as negative samples to the multi-tenant load identification and prediction model and the structural causal model, respectively updating the parameters of the adversarial disturbance discrimination sub-network and the baseline of tenant historical behavior.

Images