A virtual machine resource opening method and system combined with multi-modal risk assessment

Abstract

The application provides a kind of virtual machine resource opening method and system combined with multi-modal risk assessment in the technical field of virtual machine management, method includes: step S1, create order semantic recognition model, risk assessment model, resource dynamic allocation model, and train and deploy;Step S2, acquisition computing resource data and build resource heat map;Step S3, obtain real-time virtual machine purchase order to verify and parse to obtain order content and user account, obtain real-time multi-modal data based on user account;Step S4, input order content into order semantic recognition model to obtain configuration requirement, input real-time multi-modal data into risk assessment model to obtain risk assessment result;Step S5, based on risk assessment result, configuration requirement, resource heat map, call resource dynamic allocation model to execute resource allocation.The application has the advantages that: greatly improve the efficiency of virtual machine resource opening, resource utilization and security.

Description

Technical Field

[0001] This invention relates to the field of virtual machine management technology, and in particular to a method and system for enabling virtual machine resources that combines multimodal risk assessment. Background Technology

[0002] Virtual machine technology, as a core component of cloud computing infrastructure, constructs isolated virtual computing environments on physical host machines through the Hypervisor virtualization layer. Leveraging its environmental isolation, resource configurability, and deployment flexibility, this technology has fostered a diverse application ecosystem: in development and testing, it supports the parallel construction of multiple environment versions; in education and research, it enables cross-platform operating system experiments; in enterprise IT, it facilitates server resource pooling and integration, improving hardware utilization by 40%-60%; and in the end-user domain, it provides cross-platform application compatibility solutions.

[0003] Commercial virtual machines (VMs) are typically deployed on cloud platforms (servers). VMs are created by purchasing services from corresponding service providers and then utilize the cloud platform's computing resources to perform related computational tasks. The VM service purchase process first requires sending a VM purchase order to the cloud platform. Upon receiving the order, the cloud platform needs to review it. Only after approval will the corresponding VM resources be activated. Traditionally, this review and activation process requires manual intervention, which is time-consuming, labor-intensive, and prone to errors that could affect VM activation. This no longer meets current needs, thus creating a demand for automated VM resource review and activation.

[0004] However, existing automated approval and activation systems have the following problems: 1. They adopt predefined threshold strategies (such as automatic approval for credit scores > 80), but lack natural language processing capabilities and cannot identify semantic conflicts in order texts (e.g., applying for a "financial-grade security environment" but selecting "public network policy"), resulting in a high system misjudgment rate. This forces maintenance personnel to frequently intervene in rule adjustments, thereby affecting the efficiency of virtual machine resource activation; 2. They adopt a linear allocation mechanism using FIFO queues, which is prone to resource fragmentation in scenarios with sudden high concurrency requests (e.g., 36 CPU cores remaining but split into 3 12-core nodes), and at the same time cause regional load imbalance; 3. They only activate resources based on virtual machine purchase orders, without combining multimodal data (historical orders, behavioral data, etc.) for comprehensive consideration and risk assessment. This results in problems such as inappropriate resource allocation, and the orders may be tampered with during transmission, causing the virtual machines to crash during actual operation because they cannot handle the corresponding load.

[0005] Therefore, how to provide a virtual machine resource provisioning method and system that combines multimodal risk assessment to improve the efficiency, resource utilization, and security of virtual machine resource provisioning has become an urgent technical problem to be solved. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and system for virtual machine resource provisioning that combines multimodal risk assessment, so as to improve the efficiency, resource utilization and security of virtual machine resource provisioning.

[0007] In a first aspect, the present invention provides a method for enabling virtual machine resources that incorporates multimodal risk assessment, comprising the following steps:

[0008] Step S1: The cloud platform creates an order semantic recognition model, a risk assessment model, and a resource dynamic allocation model;

[0009] Step S2: The cloud platform acquires a large amount of historical multimodal data, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. After preprocessing and labeling each type of historical multimodal data, a dataset is constructed.

[0010] Step S3: The cloud platform trains the order semantic recognition model, risk assessment model, and resource dynamic allocation model using the dataset, and then deploys the trained order semantic recognition model, risk assessment model, and resource dynamic allocation model.

[0011] Step S4: The cloud platform collects computing resource data in real time and constructs a resource heat map based on the collected computing resource data;

[0012] Step S5: The cloud platform obtains real-time virtual machine purchase orders, verifies and parses the real-time virtual machine purchase orders to obtain order content and user accounts, and obtains real-time multimodal data based on the user accounts;

[0013] Step S6: The cloud platform inputs the order content into the deployed order semantic recognition model to obtain configuration requirements, and inputs the real-time multimodal data into the risk assessment model to obtain risk assessment results;

[0014] Step S7: Based on the risk assessment results, configuration requirements, and resource heatmap, the cloud platform calls the resource dynamic allocation model to allocate computing resources in order to perform virtual machine resource activation operations.

[0015] Step S8: The cloud platform provides feedback on the activation result and records the real-time activation log;

[0016] In step S1, the order semantic recognition model is constructed based on an input representation module, a context encoding module, a multi-task decoding module, and a post-processing enhancement module. The input representation module is used to convert virtual machine purchase orders into order vectors. The context encoding module is constructed based on a BiLSTM network and a Transformer network. The BiLSTM network is used to extract local dependency features from the order vectors, and the Transformer network is used to extract global dependency features from the order vectors. The multi-task decoding module is used to extract demand parameters, demand strength, and implicit demands from the local dependency features and global dependency features. The post-processing enhancement module is used to semantically normalize the demand parameters, demand strength, and implicit demands to obtain configuration requirements.

[0017] The risk assessment model is constructed based on a multimodal input module, a feature extraction module, a feature fusion module, and a risk output module. The multimodal input module is used to input multimodal data. The feature extraction module extracts multimodal features from the multimodal data, including semantic conflict level, frequency of abnormal operations, geographical distribution of request sources, and resource over-allocation ratio. The feature fusion module fuses the various multimodal features using a random forest classifier to obtain fused features. The risk output module outputs a risk assessment result carrying a risk level based on the fused features; the risk level is low risk, medium risk, or high risk.

[0018] The dynamic resource allocation model is built on a dynamic graph neural network and is used to generate resource allocation schemes based on configuration requirements and resource heatmaps.

[0019] Furthermore, step S2 specifically includes:

[0020] The cloud platform acquires a large amount of historical multimodal data, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. The historical virtual machine purchase orders include at least the order number, order details, purchase time, purchaser information, and terms of service. The historical purchase behavior data includes at least user attributes, order placement time, purchase frequency, payment method, billing method, IP address, purchase channel, browsing behavior, and interaction behavior. The historical resource data includes at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. The historical activation logs include at least transaction records, activation time, virtual machine name, virtual machine configuration, and user account.

[0021] The historical multimodal data are preprocessed in at least the following ways: stop word removal, outlier removal, missing value imputation, word segmentation, and format standardization.

[0022] The preprocessed historical multimodal data are labeled with configuration requirements, risk levels, resource heatmaps, and resource allocation schemes. A dataset is constructed based on the labeled historical multimodal data.

[0023] Furthermore, step S3 specifically includes:

[0024] The cloud platform divides the dataset into training, validation, and test sets in an 8:1:1 ratio. The order semantic recognition model, risk assessment model, and resource dynamic allocation model are trained using the corresponding data in the dataset. During training, hyperparameters including learning rate, learning decay rate, random dropout rate, batch size, and number of iterations are continuously optimized. The weights and biases of the order semantic recognition model, risk assessment model, and resource dynamic allocation model are updated using the backpropagation algorithm until the preset convergence conditions are met.

[0025] The trained order semantic recognition model, risk assessment model, and resource dynamic allocation model are validated sequentially using the validation set. The validated order semantic recognition model, risk assessment model, and resource dynamic allocation model are tested sequentially using the test set. The tested order semantic recognition model, risk assessment model, and resource dynamic allocation model are then converted to ONNX format and deployed.

[0026] Furthermore, step S4 specifically includes:

[0027] The cloud platform collects computing resource data in real time, including at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. It performs preprocessing on each computing resource data, including at least cleaning, transformation, and aggregation. Based on a preset display method, it maps the preprocessed computing resource data to obtain a resource heatmap. The resource heatmap is updated through a preset sliding time window.

[0028] Step S5 specifically involves:

[0029] The cloud platform obtains real-time virtual machine purchase orders carrying encrypted messages, hardware fingerprints, timestamps, and hash values. The encrypted messages are obtained by encrypting the order content and user account. The specific encryption rules are as follows: the order content and user account are encrypted into a first layer of encrypted data using the AES algorithm; this first layer is converted to hexadecimal data; the numbers 6 and A, and 9 and B are swapped in the hexadecimal data to obtain a second layer of encrypted data; this second layer is encrypted into a third layer using the 3DES algorithm; each character in the third layer is shifted 8 bits to the right to obtain a fourth layer of encrypted data; and this fourth layer is encrypted into an encrypted message using the RSA algorithm. The hardware fingerprint is obtained by concatenating the device serial number and MAC address. The hash value is obtained by hashing the encrypted message, hardware fingerprint, and timestamp.

[0030] The cloud platform parses the received real-time virtual machine purchase order to obtain an encrypted message, hardware fingerprint, timestamp, and hash value. Integrity is verified using the hash value, timeliness is verified using the timestamp, and legality is verified using the hardware fingerprint. Then, the encrypted message is decrypted into four layers of encrypted data using the RSA algorithm. Each character in the four layers of encrypted data is shifted 8 bits to the left to obtain three layers of encrypted data. The three layers of encrypted data are then decrypted into two layers of encrypted data using the 3DES algorithm. The number 6 and the letter A, and the number 9 and the letter B in the two layers of encrypted data are swapped to obtain hexadecimal data. The hexadecimal data undergoes a base inversion to obtain one layer of encrypted data. The one layer of encrypted data is then decrypted into order details and user account information using the AES algorithm. Real-time multimodal data is then obtained based on the user account.

[0031] Furthermore, step S7 specifically includes:

[0032] The cloud platform analyzes the risk assessment results. When the risk assessment result is low risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, and automatically executes the virtual machine resource activation operation based on the resource allocation scheme. When the risk assessment result is medium risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, encrypts the resource allocation scheme, and pushes it to the pre-associated management terminal for approval in real time. When the risk assessment result is high risk, it refuses to execute the virtual machine resource activation operation.

[0033] Step S8 specifically involves:

[0034] The cloud platform generates activation results, indicating whether the activation was successful or failed. The activation results are then encrypted into first-level encrypted data using the RC6 algorithm. The first-level encrypted data is then divided into two levels of encrypted data by splitting the data in a 1:4 ratio and rearranging the order. Finally, the second-level encrypted data is encrypted into an encrypted notification using the Twofish algorithm, and the encrypted notification is then sent back.

[0035] The cloud platform records real-time activation logs, encrypts the real-time activation logs into first encrypted data using the IDEA algorithm, adds a random string of a specified length to a specified position in the first encrypted data to obtain second encrypted data, encrypts the second encrypted data into an encrypted log using the SM4 algorithm, and stores and performs distributed backup of the encrypted log.

[0036] Secondly, the present invention provides a virtual machine resource provisioning system that incorporates multimodal risk assessment, comprising the following modules:

[0037] The model creation module is used to create an order semantic recognition model, a risk assessment model, and a resource dynamic allocation model on the cloud platform.

[0038] The dataset construction module is used to acquire a large amount of historical multimodal data from the cloud platform, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. The dataset is constructed after preprocessing and labeling each of the historical multimodal data.

[0039] The model training module is used by the cloud platform to train the order semantic recognition model, risk assessment model, and resource dynamic allocation model using the dataset, and to deploy the trained order semantic recognition model, risk assessment model, and resource dynamic allocation model.

[0040] The resource heatmap construction module is used to collect computing resource data in real time on the cloud platform and construct a resource heatmap based on the collected computing resource data.

[0041] The purchase order acquisition module is used to acquire real-time virtual machine purchase orders from the cloud platform, verify and parse the real-time virtual machine purchase orders to obtain the order content and user account, and acquire real-time multimodal data based on the user account;

[0042] The semantic recognition risk assessment module is used by the cloud platform to input the order content into the deployed order semantic recognition model to obtain configuration requirements, and to input the real-time multimodal data into the risk assessment model to obtain risk assessment results;

[0043] The virtual machine resource activation module is used by the cloud platform to call the resource dynamic allocation model to allocate computing resources based on the risk assessment results, configuration requirements and resource heatmap, so as to perform virtual machine resource activation operations.

[0044] The activation result feedback module is used by the cloud platform to provide feedback on the activation results and record real-time activation logs.

[0045] In the model creation module, the order semantic recognition model is constructed based on an input representation module, a context encoding module, a multi-task decoding module, and a post-processing enhancement module. The input representation module converts virtual machine purchase orders into order vectors. The context encoding module is constructed based on a BiLSTM network and a Transformer network; the BiLSTM network extracts local dependency features from the order vectors, and the Transformer network extracts global dependency features from the order vectors. The multi-task decoding module extracts demand parameters, demand strength, and implicit demands from the local and global dependency features. The post-processing enhancement module performs semantic normalization on the demand parameters, demand strength, and implicit demands to obtain configuration requirements.

[0046] The risk assessment model is constructed based on a multimodal input module, a feature extraction module, a feature fusion module, and a risk output module. The multimodal input module is used to input multimodal data. The feature extraction module extracts multimodal features from the multimodal data, including semantic conflict level, frequency of abnormal operations, geographical distribution of request sources, and resource over-allocation ratio. The feature fusion module fuses the various multimodal features using a random forest classifier to obtain fused features. The risk output module outputs a risk assessment result carrying a risk level based on the fused features; the risk level is low risk, medium risk, or high risk.

[0047] The dynamic resource allocation model is built on a dynamic graph neural network and is used to generate resource allocation schemes based on configuration requirements and resource heatmaps.

[0048] Furthermore, the dataset construction module is specifically used for:

[0049] The cloud platform acquires a large amount of historical multimodal data, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. The historical virtual machine purchase orders include at least the order number, order details, purchase time, purchaser information, and terms of service. The historical purchase behavior data includes at least user attributes, order placement time, purchase frequency, payment method, billing method, IP address, purchase channel, browsing behavior, and interaction behavior. The historical resource data includes at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. The historical activation logs include at least transaction records, activation time, virtual machine name, virtual machine configuration, and user account.

[0050] The historical multimodal data are preprocessed in at least the following ways: stop word removal, outlier removal, missing value imputation, word segmentation, and format standardization.

[0051] The preprocessed historical multimodal data are labeled with configuration requirements, risk levels, resource heatmaps, and resource allocation schemes. A dataset is constructed based on the labeled historical multimodal data.

[0052] Furthermore, the model training module is specifically used for:

[0053] The cloud platform divides the dataset into training, validation, and test sets in an 8:1:1 ratio. The order semantic recognition model, risk assessment model, and resource dynamic allocation model are trained using the corresponding data in the dataset. During training, hyperparameters including learning rate, learning decay rate, random dropout rate, batch size, and number of iterations are continuously optimized. The weights and biases of the order semantic recognition model, risk assessment model, and resource dynamic allocation model are updated using the backpropagation algorithm until the preset convergence conditions are met.

[0054] The trained order semantic recognition model, risk assessment model, and resource dynamic allocation model are validated sequentially using the validation set. The validated order semantic recognition model, risk assessment model, and resource dynamic allocation model are tested sequentially using the test set. The tested order semantic recognition model, risk assessment model, and resource dynamic allocation model are then converted to ONNX format and deployed.

[0055] Furthermore, the resource heatmap construction module is specifically used for:

[0056] The cloud platform collects computing resource data in real time, including at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. It performs preprocessing on each computing resource data, including at least cleaning, transformation, and aggregation. Based on a preset display method, it maps the preprocessed computing resource data to obtain a resource heatmap. The resource heatmap is updated through a preset sliding time window.

[0057] The purchase order acquisition module is specifically used for:

[0058] The cloud platform obtains real-time virtual machine purchase orders carrying encrypted messages, hardware fingerprints, timestamps, and hash values. The encrypted messages are obtained by encrypting the order content and user account. The specific encryption rules are as follows: the order content and user account are encrypted into a first layer of encrypted data using the AES algorithm; this first layer is converted to hexadecimal data; the numbers 6 and A, and 9 and B are swapped in the hexadecimal data to obtain a second layer of encrypted data; this second layer is encrypted into a third layer using the 3DES algorithm; each character in the third layer is shifted 8 bits to the right to obtain a fourth layer of encrypted data; and this fourth layer is encrypted into an encrypted message using the RSA algorithm. The hardware fingerprint is obtained by concatenating the device serial number and MAC address. The hash value is obtained by hashing the encrypted message, hardware fingerprint, and timestamp.

[0059] The cloud platform parses the received real-time virtual machine purchase order to obtain an encrypted message, hardware fingerprint, timestamp, and hash value. Integrity is verified using the hash value, timeliness is verified using the timestamp, and legality is verified using the hardware fingerprint. Then, the encrypted message is decrypted into four layers of encrypted data using the RSA algorithm. Each character in the four layers of encrypted data is shifted 8 bits to the left to obtain three layers of encrypted data. The three layers of encrypted data are then decrypted into two layers of encrypted data using the 3DES algorithm. The number 6 and the letter A, and the number 9 and the letter B in the two layers of encrypted data are swapped to obtain hexadecimal data. The hexadecimal data undergoes a base inversion to obtain one layer of encrypted data. The one layer of encrypted data is then decrypted into order details and user account information using the AES algorithm. Real-time multimodal data is then obtained based on the user account.

[0060] Furthermore, the virtual machine resource enabling module is specifically used for:

[0061] The cloud platform analyzes the risk assessment results. When the risk assessment result is low risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, and automatically executes the virtual machine resource activation operation based on the resource allocation scheme. When the risk assessment result is medium risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, encrypts the resource allocation scheme, and pushes it to the pre-associated management terminal for approval in real time. When the risk assessment result is high risk, it refuses to execute the virtual machine resource activation operation.

[0062] The activation result feedback module is specifically used for:

[0063] The cloud platform generates activation results, indicating whether the activation was successful or failed. The activation results are then encrypted into first-level encrypted data using the RC6 algorithm. The first-level encrypted data is then divided into two levels of encrypted data by splitting the data in a 1:4 ratio and rearranging the order. Finally, the second-level encrypted data is encrypted into an encrypted notification using the Twofish algorithm, and the encrypted notification is then sent back.

[0064] The cloud platform records real-time activation logs, encrypts the real-time activation logs into first encrypted data using the IDEA algorithm, adds a random string of a specified length to a specified position in the first encrypted data to obtain second encrypted data, encrypts the second encrypted data into an encrypted log using the SM4 algorithm, and stores and performs distributed backup of the encrypted log.

[0065] The advantages of this invention are:

[0066] 1. An order semantic recognition model, risk assessment model, and resource dynamic allocation model are created through a cloud platform. A large amount of historical multimodal data is acquired, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. After preprocessing and labeling each historical multimodal data point, a dataset is constructed. This dataset is then used to train and deploy the order semantic recognition model, risk assessment model, and resource dynamic allocation model. Next, the cloud platform collects real-time computing resource data to construct a resource heatmap. The acquired real-time virtual machine purchase orders are verified and parsed to obtain order content and user accounts. Based on the user accounts, real-time multimodal data is obtained. The order content is input into the deployed order semantic recognition model to obtain configuration requirements, and the real-time multimodal data is input into the risk assessment model to obtain risk assessment results. Finally, based on the risk assessment results, configuration requirements, and resource heatmap, resources are allocated. The dynamic allocation model executes the allocation of computing resources to perform virtual machine resource activation operations. The cloud platform provides feedback on the activation results and records real-time activation logs. Specifically, it performs virtual machine resource activation operations based on a pre-trained order semantic recognition model, risk assessment model, and dynamic resource allocation model. When the risk assessment result is low risk, the entire process is automated and requires no manual intervention. The order semantic recognition model can effectively identify semantic conflicts in order text, improving the accuracy of configuration requirement identification and minimizing manual intervention. The resource heatmap can effectively integrate resource fragments, and combined with the dynamic resource allocation model, it can effectively utilize computing resources. Furthermore, during the virtual machine resource activation process, the verification of real-time virtual machine purchase orders prevents order tampering. Combined with the risk assessment of the risk assessment model, this ultimately greatly improves the efficiency, resource utilization, and security of virtual machine resource activation.

[0067] 2. By setting up an order semantic recognition model based on an input representation module, a context encoding module, a multi-task decoding module, and a post-processing enhancement module, and combining hierarchical feature extraction (local → global), multi-task joint optimization (explicit requirement parameters and requirement intensity + implicit requirements), heterogeneous model fusion (BiLSTM + Transformer), and output standardization, deep semantic parsing of virtual machine purchase orders is achieved, thereby greatly improving the accuracy of configuration requirement recognition.

[0068] 3. The feature extraction module of the risk assessment model focuses on four major features: semantic conflict level, frequency of abnormal operations, geographical distribution, and resource overprovision ratio. It covers key dimensions such as risk behavior patterns (abnormal frequency), intent recognition (semantic conflict), environmental anomalies (geographical anomalies), and system vulnerability (resource overprovision). It has clear business interpretability. For example, abnormal geographical distribution can identify credential stuffing attacks, and resource overprovision warning system overload risk can be used to comprehensively assess the risks in the virtual machine resource provisioning process, thereby greatly improving the security of virtual machine resource provisioning.

[0069] 4. By employing a dynamic graph neural network to analyze resource heatmaps in real time (based on real-time metrics such as CPU / memory / network), combined with a sliding window update strategy, elastic scheduling of resource allocation is achieved, which greatly improves the utilization rate of computing resources compared to static allocation strategies.

[0070] 5. By designing a multi-layered hybrid encryption mechanism (AES+3DES+RSA+character substitution) in the real-time virtual machine purchase order, combined with hardware fingerprint (device serial number + MAC address), timestamp and hash value verification, it can resist man-in-the-middle attacks and replay attacks, and increase the cost of cracking encrypted messages to the order of 2^128, so that the security reaches the financial grade standard.

[0071] 6. By setting up a three-tiered risk handling strategy: low-risk automatic execution (response time <500ms), medium-risk manual review (encrypted push to management terminal), and high-risk immediate blocking, business efficiency and risk control are effectively balanced.

[0072] 7. The activation result is encrypted into first-level encrypted data using the RC6 algorithm. The first-level encrypted data is then divided into two-level encrypted data by splitting the data in a 1:4 ratio and rearranging the order. The second-level encrypted data is then encrypted into an encrypted notification using the Twofish algorithm. At least three security measures are taken (RC6 algorithm, data splitting and order rearranging, and Twofish algorithm), which greatly improves the security of the activation result transmission.

[0073] 8. The real-time activation log is encrypted into first encrypted data using the IDEA algorithm. A random string of a specified length is added to a specified position in the first encrypted data to obtain second encrypted data. The second encrypted data is then encrypted into an encrypted log using the SM4 algorithm. The encrypted log is stored and backed up in a distributed manner. At least four security measures are taken (IDEA algorithm, random string, SM4 algorithm, distributed backup), which greatly improves the security of the activation log storage and facilitates later traceability.

[0074] 9. By using an 8:1:1 split ratio, the dataset is proportionally allocated to the training set, validation set, and test set, which ensures sufficient training data (80%) while reserving an effective sample size for model tuning (10% validation set) and independent evaluation (10% test set).

[0075] 10. By dynamically adjusting key parameters during model training, covering core parameters such as learning rate (controlling gradient step size), learning decay rate (adaptively adjusting learning pace), random dropout rate (Dropout to prevent overfitting), batch size (balancing training speed and stability), and number of iterations (training sufficiency), a multi-dimensional optimization coverage is formed, which greatly improves model performance. Attached Figure Description

[0076] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0077] Figure 1 This is a flowchart of a virtual machine resource provisioning method that incorporates multimodal risk assessment according to the present invention.

[0078] Figure 2 This is a schematic diagram of the structure of a virtual machine resource provisioning system that combines multimodal risk assessment according to the present invention. Detailed Implementation

[0079] The overall approach of the technical solution in this application is as follows: Virtual machine resource activation is performed based on a pre-trained order semantic recognition model, risk assessment model, and dynamic resource allocation model. When the risk assessment result is low risk, the entire process is automated without human intervention. The order semantic recognition model effectively identifies semantic conflicts in order text, improving the accuracy of configuration requirement identification and minimizing human intervention. Resource heatmaps effectively integrate resource fragments, and the dynamic resource allocation model enables dynamic allocation of computing resources, facilitating efficient utilization. During virtual machine resource activation, real-time verification of virtual machine purchase orders prevents order tampering, and the risk assessment, combined with the risk assessment model, improves the efficiency, resource utilization, and security of virtual machine resource activation.

[0080] Please refer to Figures 1 to 2As shown, a preferred embodiment of the virtual machine resource provisioning method combining multimodal risk assessment of the present invention includes the following steps:

[0081] Step S1: The cloud platform creates an order semantic recognition model, a risk assessment model, and a resource dynamic allocation model;

[0082] Step S2: The cloud platform acquires a large amount of historical multimodal data, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. After preprocessing and labeling each type of historical multimodal data, a dataset is constructed.

[0083] Step S3: The cloud platform trains the order semantic recognition model, risk assessment model, and resource dynamic allocation model using the dataset, and then deploys the trained order semantic recognition model, risk assessment model, and resource dynamic allocation model.

[0084] Step S4: The cloud platform collects computing resource data in real time and constructs a resource heat map based on the collected computing resource data;

[0085] Step S5: The cloud platform obtains real-time virtual machine purchase orders, verifies and parses the real-time virtual machine purchase orders to obtain order content and user accounts, and obtains real-time multimodal data based on the user accounts; that is, it obtains all data related to the user accounts for subsequent risk assessment.

[0086] Step S6: The cloud platform inputs the order content into the deployed order semantic recognition model to obtain configuration requirements, and inputs the real-time multimodal data into the risk assessment model to obtain risk assessment results;

[0087] Step S7: Based on the risk assessment results, configuration requirements, and resource heatmap, the cloud platform calls the resource dynamic allocation model to allocate computing resources and perform virtual machine resource activation. Subsequently, the virtual machine can be elastically expanded and shrunk based on the resource heatmap and resource dynamic allocation model to cope with sudden loads and improve resource utilization.

[0088] Step S8: The cloud platform provides feedback on the activation result and records the real-time activation log;

[0089] In step S1, the order semantic recognition model is constructed based on an input representation module, a context encoding module, a multi-task decoding module, and a post-processing enhancement module. The input representation module converts virtual machine purchase orders into order vectors. The context encoding module is constructed based on a BiLSTM network and a Transformer network; the BiLSTM network extracts local dependency features from the order vectors, and the Transformer network extracts global dependency features from the order vectors. The multi-task decoding module extracts demand parameters, demand strength, and implicit demand from the local and global dependency features. The post-processing enhancement module further refines the demand parameters, demand strength, and implicit demand. Semantic normalization yields configuration requirements; the BiLSTM network can capture serialized configuration descriptions (e.g., "no less than 4 cores"); the Transformer network can identify cross-sentence associations (e.g., "same configuration as the previous order"); in specific implementations, constraint rules can also be configured (e.g., "SSD must be paired with ≥4 core CPUs"); requirement parameters include: {type: CPU, value: 8 cores}, {type: storage, value: 500GB SSD}; requirement intensity includes: hard requirements ("must be equipped with GPU"), soft requirements ("preferably with backup storage"); implicit requirements include: identifying "for deep learning" -> automatically associating GPU / TensorCore requirements; semantic normalization includes: unifying "16G memory" and "16GB RAM" into a standard format;

[0090] By setting up an order semantic recognition model based on an input representation module, a context encoding module, a multi-task decoding module, and a post-processing enhancement module, and combining hierarchical feature extraction (local → global), multi-task joint optimization (explicit requirement parameters and requirement intensity + implicit requirements), heterogeneous model fusion (BiLSTM + Transformer), and output standardization, deep semantic parsing of virtual machine purchase orders is achieved, thereby greatly improving the accuracy of configuration requirement recognition.

[0091] The risk assessment model is constructed based on a multimodal input module, a feature extraction module, a feature fusion module, and a risk output module. The multimodal input module is used to input multimodal data. The feature extraction module extracts multimodal features from the multimodal data, including semantic conflict level, abnormal operation frequency, request source geographical distribution, and resource over-provisioning ratio. The feature fusion module fuses the various multimodal features using a random forest classifier to obtain fused features. The risk output module outputs a risk assessment result carrying a risk level based on the fused features. The risk level is low, medium, or high risk. The semantic conflict level is used to assess the rationality of the configuration; the abnormal operation frequency is used to assess account credibility; the request source geographical distribution is used to assess fraud risk; and the resource over-provisioning ratio is used for stability prediction. The fusion of various multimodal features using a random forest classifier can be selected as follows:

[0092] Risk_Score = 0.3 * Semantic Risk (Semantic Conflict Level) + 0.2 * Behavioral Anomaly (Frequency of Abnormal Operations) + 0.25 * Device Suspicion (Geographical Distribution of Request Sources) + 0.25 * System Load (Ratio of Resource Over-provisioning).

[0093] The feature extraction module of the risk assessment model focuses on four major features: semantic conflict level, frequency of abnormal operations, geographical distribution, and resource overprovision ratio. It covers key dimensions such as risk behavior patterns (abnormal frequency), intent recognition (semantic conflict), environmental anomalies (geographical anomalies), and system vulnerability (resource overprovision). It has clear business interpretability. For example, abnormal geographical distribution can identify credential stuffing attacks, and resource overprovision warnings can identify system overload risks. It can comprehensively assess the risks in the virtual machine resource provisioning process, thereby greatly improving the security of virtual machine resource provisioning.

[0094] The dynamic resource allocation model is built on a dynamic graph neural network and is used to generate resource allocation schemes based on configuration requirements and resource heatmaps.

[0095] By employing a dynamic graph neural network to analyze resource heatmaps in real time (based on real-time metrics such as CPU, memory, and network), and combining this with a sliding window update strategy, elastic scheduling of resource allocation is achieved, which greatly improves the utilization rate of computing resources compared to static allocation strategies.

[0096] Resource heatmaps are visualization tools that use color intensity to represent resource distribution. They can visually display the density of computing resources in different regions or nodes, providing important references for resource allocation. Dynamic graph neural networks can handle dynamically changing graph structure data. They can not only capture the relationships between nodes but also adapt to changes in the graph's topology over time, making them suitable for resource allocation scenarios where resource demand and availability change dynamically over time.

[0097] Using data from resource heatmaps as input, including information such as resource location, quantity, and demand, this data can be transformed into node features in a dynamic graph neural network (Graph Neural Network). The connections between nodes are constructed based on the dependencies or interactions between resources. The Graph Neural Network can learn the complex relationships between node features and graph structure. Through training, the network can identify key patterns and rules in resource allocation, such as which regions have high resource demand but insufficient supply, and which regions have resource surplus. Based on the learned patterns and rules, the Graph Neural Network can predict the optimal resource allocation scheme, taking into account dynamic resource changes and adjusting allocation strategies in real time to meet the needs of different regions while optimizing resource utilization.

[0098] Furthermore, step S2 specifically includes:

[0099] The cloud platform acquires a large amount of historical multimodal data, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. The historical virtual machine purchase orders include at least the order number, order details, purchase time, purchaser information, and terms of service. The historical purchase behavior data includes at least user attributes, order placement time, purchase frequency, payment method, billing method, IP address, purchase channel, browsing behavior, and interaction behavior. The historical resource data includes at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. The historical activation logs include at least transaction records, activation time, virtual machine name, virtual machine configuration, and user account.

[0100] The historical multimodal data are preprocessed in at least the following ways: stop word removal, outlier removal, missing value imputation, word segmentation, and format standardization.

[0101] The preprocessed historical multimodal data are labeled with configuration requirements, risk levels, resource heatmaps, and resource allocation schemes. A dataset is constructed based on the labeled historical multimodal data.

[0102] Furthermore, step S3 specifically includes:

[0103] The cloud platform divides the dataset into training, validation, and test sets in an 8:1:1 ratio. The order semantic recognition model, risk assessment model, and resource dynamic allocation model are trained using the corresponding data in the dataset. During training, hyperparameters including learning rate, learning decay rate, random dropout rate, batch size, and number of iterations are continuously optimized. The weights and biases of the order semantic recognition model, risk assessment model, and resource dynamic allocation model are updated using the backpropagation algorithm until the preset convergence conditions are met.

[0104] By using an 8:1:1 split ratio, the dataset is proportionally allocated to the training set, validation set, and test set, which ensures sufficient training data (80%) while reserving an effective sample size for model tuning (10% validation set) and independent evaluation (10% test set).

[0105] By dynamically adjusting key parameters during model training, covering core parameters such as learning rate (controlling gradient step size), learning decay rate (adaptively adjusting learning pace), random dropout rate (Dropout to prevent overfitting), batch size (balancing training speed and stability), and number of iterations (training sufficiency), a multi-dimensional optimization coverage is formed, which greatly improves model performance.

[0106] The trained order semantic recognition model, risk assessment model, and resource dynamic allocation model are validated sequentially using the validation set. The validated order semantic recognition model, risk assessment model, and resource dynamic allocation model are tested sequentially using the test set. The tested order semantic recognition model, risk assessment model, and resource dynamic allocation model are then converted to ONNX format and deployed.

[0107] Furthermore, step S4 specifically includes:

[0108] The cloud platform collects computing resource data in real time, including at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. It performs preprocessing on each computing resource data, including at least cleaning, transformation, and aggregation. Based on a preset display method, it maps the preprocessed computing resource data to obtain a resource heatmap, and updates the resource heatmap through a preset sliding time window. In other words, it maps each computing resource data to the corresponding nodes in the network topology to obtain the resource heatmap.

[0109] To achieve dynamic updates to the resource heatmap, a pre-defined sliding time window technique is introduced. The sliding time window creates training samples using a fixed-size window to capture patterns in the time series. During the resource heatmap update process, the sliding time window moves at preset time intervals (e.g., per second, per minute), acquiring the latest computing resource data within the window with each movement and updating the heatmap accordingly. This ensures that the heatmap reflects the latest status of computing resources in real time, helping users understand resource usage and trends promptly.

[0110] Step S5 specifically involves:

[0111] The cloud platform obtains real-time virtual machine purchase orders carrying encrypted messages, hardware fingerprints, timestamps, and hash values. The encrypted messages are obtained by encrypting the order content and user account. The specific encryption rules are as follows: the order content and user account are encrypted into a first layer of encrypted data using the AES algorithm; this first layer is converted to hexadecimal data; the numbers 6 and A, and 9 and B are swapped in the hexadecimal data to obtain a second layer of encrypted data; this second layer is encrypted into a third layer using the 3DES algorithm; each character in the third layer is shifted 8 bits to the right to obtain a fourth layer of encrypted data; and this fourth layer is encrypted into an encrypted message using the RSA algorithm. The hardware fingerprint is obtained by concatenating the device serial number and MAC address. The hash value is obtained by hashing the encrypted message, hardware fingerprint, and timestamp.

[0112] By designing a multi-layered hybrid encryption mechanism (AES+3DES+RSA+character substitution) in the real-time virtual machine purchase order, combined with hardware fingerprint (device serial number + MAC address), timestamp and hash value verification, it can resist man-in-the-middle attacks and replay attacks, and increase the cost of cracking encrypted messages to the order of 2^128, so that the security reaches the financial grade standard.

[0113] The cloud platform parses the received real-time virtual machine purchase order to obtain an encrypted message, hardware fingerprint, timestamp, and hash value. Integrity is verified using the hash value, timeliness is verified using the timestamp, and legality is verified using the hardware fingerprint. Then, the encrypted message is decrypted into four layers of encrypted data using the RSA algorithm. Each character in the four layers of encrypted data is shifted 8 bits to the left to obtain three layers of encrypted data. The three layers of encrypted data are then decrypted into two layers of encrypted data using the 3DES algorithm. The number 6 and the letter A, and the number 9 and the letter B in the two layers of encrypted data are swapped to obtain hexadecimal data. The hexadecimal data undergoes a base inversion to obtain one layer of encrypted data. The one layer of encrypted data is then decrypted into order details and user account information using the AES algorithm. Real-time multimodal data is then obtained based on the user account.

[0114] Furthermore, step S7 specifically includes:

[0115] The cloud platform analyzes the risk assessment results. When the risk assessment result is low risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, and automatically executes the virtual machine resource activation operation based on the resource allocation scheme. When the risk assessment result is medium risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, encrypts the resource allocation scheme, and pushes it to the pre-associated management terminal for approval in real time. When the risk assessment result is high risk, it refuses to execute the virtual machine resource activation operation.

[0116] By setting up a three-tiered risk management strategy: low-risk automatic execution (response time <500ms), medium-risk manual review (encrypted push to management terminal), and high-risk immediate blocking, business efficiency and risk control are effectively balanced.

[0117] Step S8 specifically involves:

[0118] The cloud platform generates activation results, indicating whether the activation was successful or failed. The activation results are then encrypted into first-level encrypted data using the RC6 algorithm. The first-level encrypted data is then divided into two levels of encrypted data by splitting the data in a 1:4 ratio and rearranging the order. Finally, the second-level encrypted data is encrypted into an encrypted notification using the Twofish algorithm, and the encrypted notification is then sent back.

[0119] The activation result is encrypted into first-level encrypted data using the RC6 algorithm. The first-level encrypted data is then split into second-level encrypted data by a 1:4 ratio and the order of the data is reversed. Finally, the second-level encrypted data is encrypted into an encrypted notification using the Twofish algorithm. With at least three security measures (RC6 algorithm, data splitting and reordering, and Twofish algorithm), the security of the activation result transmission is greatly improved.

[0120] The cloud platform records real-time activation logs, encrypts the real-time activation logs into first encrypted data using the IDEA algorithm, adds a random string of a specified length to a specified position in the first encrypted data to obtain second encrypted data, encrypts the second encrypted data into an encrypted log using the SM4 algorithm, and stores and performs distributed backup of the encrypted log.

[0121] The real-time activation log is encrypted into first encrypted data using the IDEA algorithm. A random string of a specified length is added to a specified position in the first encrypted data to obtain second encrypted data. The second encrypted data is then encrypted into an encrypted log using the SM4 algorithm. The encrypted log is stored and backed up in a distributed manner. At least four security measures are taken (IDEA algorithm, random string, SM4 algorithm, and distributed backup), which greatly improves the security of the activation log storage and facilitates later traceability.

[0122] A preferred embodiment of the virtual machine resource provisioning system combining multimodal risk assessment of the present invention includes the following modules:

[0123] The model creation module is used to create an order semantic recognition model, a risk assessment model, and a resource dynamic allocation model on the cloud platform.

[0124] The dataset construction module is used to acquire a large amount of historical multimodal data from the cloud platform, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. The dataset is constructed after preprocessing and labeling each of the historical multimodal data.

[0125] The model training module is used by the cloud platform to train the order semantic recognition model, risk assessment model, and resource dynamic allocation model using the dataset, and to deploy the trained order semantic recognition model, risk assessment model, and resource dynamic allocation model.

[0126] The resource heatmap construction module is used to collect computing resource data in real time on the cloud platform and construct a resource heatmap based on the collected computing resource data.

[0127] The purchase order acquisition module is used to obtain real-time virtual machine purchase orders from the cloud platform, verify and parse the real-time virtual machine purchase orders to obtain the order content and user account, and obtain real-time multimodal data based on the user account; that is, to obtain all data related to the user account for subsequent risk assessment.

[0128] The semantic recognition risk assessment module is used by the cloud platform to input the order content into the deployed order semantic recognition model to obtain configuration requirements, and to input the real-time multimodal data into the risk assessment model to obtain risk assessment results;

[0129] The virtual machine resource activation module is used by the cloud platform to call the resource dynamic allocation model to allocate computing resources based on the risk assessment results, configuration requirements, and resource heatmap, so as to perform virtual machine resource activation operations; subsequently, the virtual machine can be elastically expanded and shrunk based on the resource heatmap and resource dynamic allocation model to cope with sudden load and improve resource utilization.

[0130] The activation result feedback module is used by the cloud platform to provide feedback on the activation results and record real-time activation logs.

[0131] In the model creation module, the order semantic recognition model is constructed based on an input representation module, a context encoding module, a multi-task decoding module, and a post-processing enhancement module. The input representation module converts virtual machine purchase orders into order vectors. The context encoding module is constructed based on a BiLSTM network and a Transformer network. The BiLSTM network extracts local dependency features from the order vectors, and the Transformer network extracts global dependency features from the order vectors. The multi-task decoding module extracts demand parameters, demand strength, and implicit demands from the local and global dependency features. The post-processing enhancement module performs semantic normalization on the demand parameters, demand strength, and implicit demands to obtain configuration requirements. The BiLSTM network can capture serialized configuration descriptions (e.g., "Do not use less than 4 cores"). The Transformer network can identify cross-sentence associations (e.g., "Same configuration as the previous order"). In specific implementations, constraint rules can also be configured (e.g., "SSD must be equipped with ≥4 core CPUs"). Demand parameters include, for example: {Type: CPU, Value: 8 cores}, {Type: Storage, Value: 500GB}. SSD}; Demand intensity includes: hard requirements ("GPU must be equipped"), soft requirements ("backup storage is preferred"); Implicit requirements, such as: identifying "for deep learning" -> automatically associating GPU / TensorCore requirements; Semantic normalization, such as: unifying "16G memory" and "16GB RAM" into a standard format;

[0132] By setting up an order semantic recognition model based on an input representation module, a context encoding module, a multi-task decoding module, and a post-processing enhancement module, and combining hierarchical feature extraction (local → global), multi-task joint optimization (explicit requirement parameters and requirement intensity + implicit requirements), heterogeneous model fusion (BiLSTM + Transformer), and output standardization, deep semantic parsing of virtual machine purchase orders is achieved, thereby greatly improving the accuracy of configuration requirement recognition.

[0133] The risk assessment model is constructed based on a multimodal input module, a feature extraction module, a feature fusion module, and a risk output module. The multimodal input module is used to input multimodal data. The feature extraction module extracts multimodal features from the multimodal data, including semantic conflict level, abnormal operation frequency, request source geographical distribution, and resource over-provisioning ratio. The feature fusion module fuses the various multimodal features using a random forest classifier to obtain fused features. The risk output module outputs a risk assessment result carrying a risk level based on the fused features. The risk level is low, medium, or high risk. The semantic conflict level is used to assess the rationality of the configuration; the abnormal operation frequency is used to assess account credibility; the request source geographical distribution is used to assess fraud risk; and the resource over-provisioning ratio is used for stability prediction. The fusion of various multimodal features using a random forest classifier can be selected as follows:

[0134] Risk_Score = 0.3 * Semantic Risk (Semantic Conflict Level) + 0.2 * Behavioral Anomaly (Frequency of Abnormal Operations) + 0.25 * Device Suspicion (Geographical Distribution of Request Sources) + 0.25 * System Load (Ratio of Resource Over-provisioning).

[0135] The feature extraction module of the risk assessment model focuses on four major features: semantic conflict level, frequency of abnormal operations, geographical distribution, and resource overprovision ratio. It covers key dimensions such as risk behavior patterns (abnormal frequency), intent recognition (semantic conflict), environmental anomalies (geographical anomalies), and system vulnerability (resource overprovision). It has clear business interpretability. For example, abnormal geographical distribution can identify credential stuffing attacks, and resource overprovision warnings can identify system overload risks. It can comprehensively assess the risks in the virtual machine resource provisioning process, thereby greatly improving the security of virtual machine resource provisioning.

[0136] The dynamic resource allocation model is built on a dynamic graph neural network and is used to generate resource allocation schemes based on configuration requirements and resource heatmaps.

[0137] By employing a dynamic graph neural network to analyze resource heatmaps in real time (based on real-time metrics such as CPU, memory, and network), and combining this with a sliding window update strategy, elastic scheduling of resource allocation is achieved, which greatly improves the utilization rate of computing resources compared to static allocation strategies.

[0138] Resource heatmaps are visualization tools that use color intensity to represent resource distribution. They can visually display the density of computing resources in different regions or nodes, providing important references for resource allocation. Dynamic graph neural networks can handle dynamically changing graph structure data. They can not only capture the relationships between nodes but also adapt to changes in the graph's topology over time, making them suitable for resource allocation scenarios where resource demand and availability change dynamically over time.

[0139] Using data from resource heatmaps as input, including information such as resource location, quantity, and demand, this data can be transformed into node features in a dynamic graph neural network (Graph Neural Network). The connections between nodes are constructed based on the dependencies or interactions between resources. The Graph Neural Network can learn the complex relationships between node features and graph structure. Through training, the network can identify key patterns and rules in resource allocation, such as which regions have high resource demand but insufficient supply, and which regions have resource surplus. Based on the learned patterns and rules, the Graph Neural Network can predict the optimal resource allocation scheme, taking into account dynamic resource changes and adjusting allocation strategies in real time to meet the needs of different regions while optimizing resource utilization.

[0140] The dataset construction module is specifically used for:

[0141] The cloud platform acquires a large amount of historical multimodal data, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. The historical virtual machine purchase orders include at least the order number, order details, purchase time, purchaser information, and terms of service. The historical purchase behavior data includes at least user attributes, order placement time, purchase frequency, payment method, billing method, IP address, purchase channel, browsing behavior, and interaction behavior. The historical resource data includes at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. The historical activation logs include at least transaction records, activation time, virtual machine name, virtual machine configuration, and user account.

[0142] The historical multimodal data are preprocessed in at least the following ways: stop word removal, outlier removal, missing value imputation, word segmentation, and format standardization.

[0143] The preprocessed historical multimodal data are labeled with configuration requirements, risk levels, resource heatmaps, and resource allocation schemes. A dataset is constructed based on the labeled historical multimodal data.

[0144] The model training module is specifically used for:

[0145] The cloud platform divides the dataset into training, validation, and test sets in an 8:1:1 ratio. The order semantic recognition model, risk assessment model, and resource dynamic allocation model are trained using the corresponding data in the dataset. During training, hyperparameters including learning rate, learning decay rate, random dropout rate, batch size, and number of iterations are continuously optimized. The weights and biases of the order semantic recognition model, risk assessment model, and resource dynamic allocation model are updated using the backpropagation algorithm until the preset convergence conditions are met.

[0146] By using an 8:1:1 split ratio, the dataset is proportionally allocated to the training set, validation set, and test set, which ensures sufficient training data (80%) while reserving an effective sample size for model tuning (10% validation set) and independent evaluation (10% test set).

[0147] By dynamically adjusting key parameters during model training, covering core parameters such as learning rate (controlling gradient step size), learning decay rate (adaptively adjusting learning pace), random dropout rate (Dropout to prevent overfitting), batch size (balancing training speed and stability), and number of iterations (training sufficiency), a multi-dimensional optimization coverage is formed, which greatly improves model performance.

[0148] The trained order semantic recognition model, risk assessment model, and resource dynamic allocation model are validated sequentially using the validation set. The validated order semantic recognition model, risk assessment model, and resource dynamic allocation model are tested sequentially using the test set. The tested order semantic recognition model, risk assessment model, and resource dynamic allocation model are then converted to ONNX format and deployed.

[0149] The resource heatmap construction module is specifically used for:

[0150] The cloud platform collects computing resource data in real time, including at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. It performs preprocessing on each computing resource data, including at least cleaning, transformation, and aggregation. Based on a preset display method, it maps the preprocessed computing resource data to obtain a resource heatmap, and updates the resource heatmap through a preset sliding time window. In other words, it maps each computing resource data to the corresponding nodes in the network topology to obtain the resource heatmap.

[0151] To achieve dynamic updates to the resource heatmap, a pre-defined sliding time window technique is introduced. The sliding time window creates training samples using a fixed-size window to capture patterns in the time series. During the resource heatmap update process, the sliding time window moves at preset time intervals (e.g., per second, per minute), acquiring the latest computing resource data within the window with each movement and updating the heatmap accordingly. This ensures that the heatmap reflects the latest status of computing resources in real time, helping users understand resource usage and trends promptly.

[0152] The purchase order acquisition module is specifically used for:

[0153] The cloud platform obtains real-time virtual machine purchase orders carrying encrypted messages, hardware fingerprints, timestamps, and hash values. The encrypted messages are obtained by encrypting the order content and user account. The specific encryption rules are as follows: the order content and user account are encrypted into a first layer of encrypted data using the AES algorithm; this first layer is converted to hexadecimal data; the numbers 6 and A, and 9 and B are swapped in the hexadecimal data to obtain a second layer of encrypted data; this second layer is encrypted into a third layer using the 3DES algorithm; each character in the third layer is shifted 8 bits to the right to obtain a fourth layer of encrypted data; and this fourth layer is encrypted into an encrypted message using the RSA algorithm. The hardware fingerprint is obtained by concatenating the device serial number and MAC address. The hash value is obtained by hashing the encrypted message, hardware fingerprint, and timestamp.

[0154] By designing a multi-layered hybrid encryption mechanism (AES+3DES+RSA+character substitution) in the real-time virtual machine purchase order, combined with hardware fingerprint (device serial number + MAC address), timestamp and hash value verification, it can resist man-in-the-middle attacks and replay attacks, and increase the cost of cracking encrypted messages to the order of 2^128, so that the security reaches the financial grade standard.

[0155] The cloud platform parses the received real-time virtual machine purchase order to obtain an encrypted message, hardware fingerprint, timestamp, and hash value. Integrity is verified using the hash value, timeliness is verified using the timestamp, and legality is verified using the hardware fingerprint. Then, the encrypted message is decrypted into four layers of encrypted data using the RSA algorithm. Each character in the four layers of encrypted data is shifted 8 bits to the left to obtain three layers of encrypted data. The three layers of encrypted data are then decrypted into two layers of encrypted data using the 3DES algorithm. The number 6 and the letter A, and the number 9 and the letter B in the two layers of encrypted data are swapped to obtain hexadecimal data. The hexadecimal data undergoes a base inversion to obtain one layer of encrypted data. The one layer of encrypted data is then decrypted into order details and user account information using the AES algorithm. Real-time multimodal data is then obtained based on the user account.

[0156] The virtual machine resource enabling module is specifically used for:

[0157] The cloud platform analyzes the risk assessment results. When the risk assessment result is low risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, and automatically executes the virtual machine resource activation operation based on the resource allocation scheme. When the risk assessment result is medium risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, encrypts the resource allocation scheme, and pushes it to the pre-associated management terminal for approval in real time. When the risk assessment result is high risk, it refuses to execute the virtual machine resource activation operation.

[0158] By setting up a three-tiered risk management strategy: low-risk automatic execution (response time <500ms), medium-risk manual review (encrypted push to management terminal), and high-risk immediate blocking, business efficiency and risk control are effectively balanced.

[0159] The activation result feedback module is specifically used for:

[0160] The cloud platform generates activation results, indicating whether the activation was successful or failed. The activation results are then encrypted into first-level encrypted data using the RC6 algorithm. The first-level encrypted data is then divided into two levels of encrypted data by splitting the data in a 1:4 ratio and rearranging the order. Finally, the second-level encrypted data is encrypted into an encrypted notification using the Twofish algorithm, and the encrypted notification is then sent back.

[0161] The activation result is encrypted into first-level encrypted data using the RC6 algorithm. The first-level encrypted data is then split into second-level encrypted data by a 1:4 ratio and the order of the data is reversed. Finally, the second-level encrypted data is encrypted into an encrypted notification using the Twofish algorithm. With at least three security measures (RC6 algorithm, data splitting and reordering, and Twofish algorithm), the security of the activation result transmission is greatly improved.

[0162] The cloud platform records real-time activation logs, encrypts the real-time activation logs into first encrypted data using the IDEA algorithm, adds a random string of a specified length to a specified position in the first encrypted data to obtain second encrypted data, encrypts the second encrypted data into an encrypted log using the SM4 algorithm, and stores and performs distributed backup of the encrypted log.

[0163] The real-time activation log is encrypted into first encrypted data using the IDEA algorithm. A random string of a specified length is added to a specified position in the first encrypted data to obtain second encrypted data. The second encrypted data is then encrypted into an encrypted log using the SM4 algorithm. The encrypted log is stored and backed up in a distributed manner. At least four security measures are taken (IDEA algorithm, random string, SM4 algorithm, and distributed backup), which greatly improves the security of the activation log storage and facilitates later traceability.

[0164] In summary, the advantages of this invention are:

[0165] 1. An order semantic recognition model, risk assessment model, and resource dynamic allocation model are created through a cloud platform. A large amount of historical multimodal data is acquired, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. After preprocessing and labeling each historical multimodal data point, a dataset is constructed. This dataset is then used to train and deploy the order semantic recognition model, risk assessment model, and resource dynamic allocation model. Next, the cloud platform collects real-time computing resource data to construct a resource heatmap. The acquired real-time virtual machine purchase orders are verified and parsed to obtain order content and user accounts. Based on the user accounts, real-time multimodal data is obtained. The order content is input into the deployed order semantic recognition model to obtain configuration requirements, and the real-time multimodal data is input into the risk assessment model to obtain risk assessment results. Finally, based on the risk assessment results, configuration requirements, and resource heatmap, resources are allocated. The dynamic allocation model executes the allocation of computing resources to perform virtual machine resource activation operations. The cloud platform provides feedback on the activation results and records real-time activation logs. Specifically, it performs virtual machine resource activation operations based on a pre-trained order semantic recognition model, risk assessment model, and dynamic resource allocation model. When the risk assessment result is low risk, the entire process is automated and requires no manual intervention. The order semantic recognition model can effectively identify semantic conflicts in order text, improving the accuracy of configuration requirement identification and minimizing manual intervention. The resource heatmap can effectively integrate resource fragments, and combined with the dynamic resource allocation model, it can effectively utilize computing resources. Furthermore, during the virtual machine resource activation process, the verification of real-time virtual machine purchase orders prevents order tampering. Combined with the risk assessment of the risk assessment model, this ultimately greatly improves the efficiency, resource utilization, and security of virtual machine resource activation.

[0166] 2. By setting up an order semantic recognition model based on an input representation module, a context encoding module, a multi-task decoding module, and a post-processing enhancement module, and combining hierarchical feature extraction (local → global), multi-task joint optimization (explicit requirement parameters and requirement intensity + implicit requirements), heterogeneous model fusion (BiLSTM + Transformer), and output standardization, deep semantic parsing of virtual machine purchase orders is achieved, thereby greatly improving the accuracy of configuration requirement recognition.

[0167] 3. The feature extraction module of the risk assessment model focuses on four major features: semantic conflict level, frequency of abnormal operations, geographical distribution, and resource overprovision ratio. It covers key dimensions such as risk behavior patterns (abnormal frequency), intent recognition (semantic conflict), environmental anomalies (geographical anomalies), and system vulnerability (resource overprovision). It has clear business interpretability. For example, abnormal geographical distribution can identify credential stuffing attacks, and resource overprovision warning system overload risk can be used to comprehensively assess the risks in the virtual machine resource provisioning process, thereby greatly improving the security of virtual machine resource provisioning.

[0168] 4. By employing a dynamic graph neural network to analyze resource heatmaps in real time (based on real-time metrics such as CPU / memory / network), combined with a sliding window update strategy, elastic scheduling of resource allocation is achieved, which greatly improves the utilization rate of computing resources compared to static allocation strategies.

[0169] 5. By designing a multi-layered hybrid encryption mechanism (AES+3DES+RSA+character substitution) in the real-time virtual machine purchase order, combined with hardware fingerprint (device serial number + MAC address), timestamp and hash value verification, it can resist man-in-the-middle attacks and replay attacks, and increase the cost of cracking encrypted messages to the order of 2^128, so that the security reaches the financial grade standard.

[0170] 6. By setting up a three-tiered risk handling strategy: low-risk automatic execution (response time <500ms), medium-risk manual review (encrypted push to management terminal), and high-risk immediate blocking, business efficiency and risk control are effectively balanced.

[0171] 7. The activation result is encrypted into first-level encrypted data using the RC6 algorithm. The first-level encrypted data is then divided into two-level encrypted data by splitting the data in a 1:4 ratio and rearranging the order. The second-level encrypted data is then encrypted into an encrypted notification using the Twofish algorithm. At least three security measures are taken (RC6 algorithm, data splitting and order rearranging, and Twofish algorithm), which greatly improves the security of the activation result transmission.

[0172] 8. The real-time activation log is encrypted into first encrypted data using the IDEA algorithm. A random string of a specified length is added to a specified position in the first encrypted data to obtain second encrypted data. The second encrypted data is then encrypted into an encrypted log using the SM4 algorithm. The encrypted log is stored and backed up in a distributed manner. At least four security measures are taken (IDEA algorithm, random string, SM4 algorithm, distributed backup), which greatly improves the security of the activation log storage and facilitates later traceability.

[0173] 9. By using an 8:1:1 split ratio, the dataset is proportionally allocated to the training set, validation set, and test set, which ensures sufficient training data (80%) while reserving an effective sample size for model tuning (10% validation set) and independent evaluation (10% test set).

[0174] 10. By dynamically adjusting key parameters during model training, covering core parameters such as learning rate (controlling gradient step size), learning decay rate (adaptively adjusting learning pace), random dropout rate (Dropout to prevent overfitting), batch size (balancing training speed and stability), and number of iterations (training sufficiency), a multi-dimensional optimization coverage is formed, which greatly improves model performance.

[0175] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for enabling virtual machine resources by combining multimodal risk assessment, characterized in that: Includes the following steps: Step S1: The cloud platform creates an order semantic recognition model, a risk assessment model, and a resource dynamic allocation model; Step S2: The cloud platform acquires a large amount of historical multimodal data, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. After preprocessing and labeling each type of historical multimodal data, a dataset is constructed. Step S3: The cloud platform trains the order semantic recognition model, risk assessment model, and resource dynamic allocation model using the dataset, and then deploys the trained order semantic recognition model, risk assessment model, and resource dynamic allocation model. Step S4: The cloud platform collects computing resource data in real time and constructs a resource heat map based on the collected computing resource data; Step S5: The cloud platform obtains real-time virtual machine purchase orders, verifies and parses the real-time virtual machine purchase orders to obtain order content and user accounts, and obtains real-time multimodal data based on the user accounts; Step S6: The cloud platform inputs the order content into the deployed order semantic recognition model to obtain configuration requirements, and inputs the real-time multimodal data into the risk assessment model to obtain risk assessment results; Step S7: Based on the risk assessment results, configuration requirements, and resource heatmap, the cloud platform calls the resource dynamic allocation model to allocate computing resources in order to perform virtual machine resource activation operations. Step S8: The cloud platform provides feedback on the activation result and records the real-time activation log.

2. The virtual machine resource provisioning method combining multimodal risk assessment as described in claim 1, characterized in that: In step S1, the order semantic recognition model is constructed based on an input representation module, a context encoding module, a multi-task decoding module, and a post-processing enhancement module. The input representation module converts virtual machine purchase orders into order vectors. The context encoding module is constructed based on a BiLSTM network and a Transformer network. The BiLSTM network is used to extract local dependency features from the order vectors, and the Transformer network is used to extract global dependency features from the order vectors. The multi-task decoding module is used to extract global dependency features from the local dependency features and the global dependency features. The requirement parameters, requirement intensity, and implicit requirements are extracted; the post-processing enhancement module is used to semantically normalize the requirement parameters, requirement intensity, and implicit requirements to obtain the configuration requirements. The risk assessment model is constructed based on a multimodal input module, a feature extraction module, a feature fusion module, and a risk output module. The multimodal input module is used to input multimodal data. The feature extraction module extracts multimodal features from the multimodal data, including semantic conflict level, frequency of abnormal operations, geographical distribution of request sources, and resource over-allocation ratio. The feature fusion module fuses the various multimodal features using a random forest classifier to obtain fused features. The risk output module outputs a risk assessment result carrying a risk level based on the fused features; the risk level is low risk, medium risk, or high risk. The resource dynamic allocation model is built on a dynamic graph neural network and is used to generate resource allocation schemes based on configuration requirements and resource heatmaps. Step S2 specifically involves: The cloud platform acquires a large amount of historical multimodal data, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. The historical virtual machine purchase orders include at least the order number, order details, purchase time, purchaser information, and terms of service. The historical purchase behavior data includes at least user attributes, order placement time, purchase frequency, payment method, billing method, IP address, purchase channel, browsing behavior, and interaction behavior. The historical resource data includes at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. The historical activation logs include at least transaction records, activation time, virtual machine name, virtual machine configuration, and user account. The historical multimodal data are preprocessed in at least the following ways: stop word removal, outlier removal, missing value imputation, word segmentation, and format standardization. The preprocessed historical multimodal data are labeled with configuration requirements, risk levels, resource heatmaps, and resource allocation schemes. A dataset is constructed based on the labeled historical multimodal data.

3. The virtual machine resource provisioning method combining multimodal risk assessment as described in claim 1, characterized in that: Step S3 specifically involves: The cloud platform divides the dataset into training, validation, and test sets in an 8:1:1 ratio. The order semantic recognition model, risk assessment model, and resource dynamic allocation model are trained using the corresponding data in the dataset. During training, hyperparameters including learning rate, learning decay rate, random dropout rate, batch size, and number of iterations are continuously optimized. The weights and biases of the order semantic recognition model, risk assessment model, and resource dynamic allocation model are updated using the backpropagation algorithm until the preset convergence conditions are met. The trained order semantic recognition model, risk assessment model, and resource dynamic allocation model are validated sequentially using the validation set. The validated order semantic recognition model, risk assessment model, and resource dynamic allocation model are tested sequentially using the test set. The tested order semantic recognition model, risk assessment model, and resource dynamic allocation model are then converted to ONNX format and deployed.

4. The virtual machine resource provisioning method combining multimodal risk assessment as described in claim 1, characterized in that: Step S4 specifically involves: The cloud platform collects computing resource data in real time, including at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. It performs preprocessing on each computing resource data, including at least cleaning, transformation, and aggregation. Based on a preset display method, it maps the preprocessed computing resource data to obtain a resource heatmap. The resource heatmap is updated through a preset sliding time window. Step S5 specifically involves: The cloud platform obtains real-time virtual machine purchase orders carrying encrypted messages, hardware fingerprints, timestamps, and hash values. The encrypted messages are obtained by encrypting the order content and user account. The specific encryption rules are as follows: the order content and user account are encrypted into a first layer of encrypted data using the AES algorithm; this first layer is converted to hexadecimal data; the numbers 6 and A, and 9 and B are swapped in the hexadecimal data to obtain a second layer of encrypted data; this second layer is encrypted into a third layer using the 3DES algorithm; each character in the third layer is shifted 8 bits to the right to obtain a fourth layer of encrypted data; and this fourth layer is encrypted into an encrypted message using the RSA algorithm. The hardware fingerprint is obtained by concatenating the device serial number and MAC address. The hash value is obtained by hashing the encrypted message, hardware fingerprint, and timestamp. The cloud platform parses the received real-time virtual machine purchase order to obtain an encrypted message, hardware fingerprint, timestamp, and hash value. Integrity is verified using the hash value, timeliness is verified using the timestamp, and legality is verified using the hardware fingerprint. Then, the encrypted message is decrypted into four layers of encrypted data using the RSA algorithm. Each character in the four layers of encrypted data is shifted 8 bits to the left to obtain three layers of encrypted data. The three layers of encrypted data are then decrypted into two layers of encrypted data using the 3DES algorithm. The number 6 and the letter A, and the number 9 and the letter B in the two layers of encrypted data are swapped to obtain hexadecimal data. The hexadecimal data undergoes a base inversion to obtain one layer of encrypted data. The one layer of encrypted data is then decrypted into order details and user account information using the AES algorithm. Real-time multimodal data is then obtained based on the user account.

5. The virtual machine resource provisioning method combining multimodal risk assessment as described in claim 1, characterized in that: Step S7 specifically involves: The cloud platform analyzes the risk assessment results. When the risk assessment result is low risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, and automatically executes the virtual machine resource activation operation based on the resource allocation scheme. When the risk assessment result is medium risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, encrypts the resource allocation scheme, and pushes it to the pre-associated management terminal for approval in real time. When the risk assessment result is high risk, it refuses to execute the virtual machine resource activation operation. Step S8 specifically involves: The cloud platform generates activation results, indicating whether the activation was successful or failed. The activation results are then encrypted into first-level encrypted data using the RC6 algorithm. The first-level encrypted data is then divided into two levels of encrypted data by splitting the data in a 1:4 ratio and rearranging the order. Finally, the second-level encrypted data is encrypted into an encrypted notification using the Twofish algorithm, and the encrypted notification is then sent back. The cloud platform records real-time activation logs, encrypts the real-time activation logs into first encrypted data using the IDEA algorithm, adds a random string of a specified length to a specified position in the first encrypted data to obtain second encrypted data, encrypts the second encrypted data into an encrypted log using the SM4 algorithm, and stores and performs distributed backup of the encrypted log.

6. A virtual machine resource provisioning system combining multimodal risk assessment, characterized in that: Includes the following modules: The model creation module is used to create an order semantic recognition model, a risk assessment model, and a resource dynamic allocation model on the cloud platform. The dataset construction module is used to acquire a large amount of historical multimodal data from the cloud platform, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. The dataset is constructed after preprocessing and labeling each of the historical multimodal data. The model training module is used by the cloud platform to train the order semantic recognition model, risk assessment model, and resource dynamic allocation model using the dataset, and to deploy the trained order semantic recognition model, risk assessment model, and resource dynamic allocation model. The resource heatmap construction module is used to collect computing resource data in real time on the cloud platform and construct a resource heatmap based on the collected computing resource data. The purchase order acquisition module is used to acquire real-time virtual machine purchase orders from the cloud platform, verify and parse the real-time virtual machine purchase orders to obtain the order content and user account, and acquire real-time multimodal data based on the user account; The semantic recognition risk assessment module is used by the cloud platform to input the order content into the deployed order semantic recognition model to obtain configuration requirements, and to input the real-time multimodal data into the risk assessment model to obtain risk assessment results; The virtual machine resource activation module is used by the cloud platform to call the resource dynamic allocation model to allocate computing resources based on the risk assessment results, configuration requirements and resource heatmap, so as to perform virtual machine resource activation operations. The activation result feedback module is used by the cloud platform to provide feedback on the activation results and record real-time activation logs.

7. A virtual machine resource provisioning system combining multimodal risk assessment as described in claim 6, characterized in that: In the model creation module, the order semantic recognition model is constructed based on an input representation module, a context encoding module, a multi-task decoding module, and a post-processing enhancement module. The input representation module converts virtual machine purchase orders into order vectors. The context encoding module is constructed based on a BiLSTM network and a Transformer network; the BiLSTM network extracts local dependency features from the order vectors, and the Transformer network extracts global dependency features from the order vectors. The multi-task decoding module extracts global dependency features from both local and global dependency features. The requirement parameters, requirement intensity, and implicit requirements are extracted; the post-processing enhancement module is used to semantically normalize the requirement parameters, requirement intensity, and implicit requirements to obtain the configuration requirements. The risk assessment model is constructed based on a multimodal input module, a feature extraction module, a feature fusion module, and a risk output module. The multimodal input module is used to input multimodal data. The feature extraction module extracts multimodal features from the multimodal data, including semantic conflict level, frequency of abnormal operations, geographical distribution of request sources, and resource over-allocation ratio. The feature fusion module fuses the various multimodal features using a random forest classifier to obtain fused features. The risk output module outputs a risk assessment result carrying a risk level based on the fused features; the risk level is low risk, medium risk, or high risk. The resource dynamic allocation model is built on a dynamic graph neural network and is used to generate resource allocation schemes based on configuration requirements and resource heatmaps. The dataset construction module is specifically used for: The cloud platform acquires a large amount of historical multimodal data, including historical virtual machine purchase orders, historical purchase behavior data, historical resource data, and historical activation logs. The historical virtual machine purchase orders include at least the order number, order details, purchase time, purchaser information, and terms of service. The historical purchase behavior data includes at least user attributes, order placement time, purchase frequency, payment method, billing method, IP address, purchase channel, browsing behavior, and interaction behavior. The historical resource data includes at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. The historical activation logs include at least transaction records, activation time, virtual machine name, virtual machine configuration, and user account. The historical multimodal data are preprocessed in at least the following ways: stop word removal, outlier removal, missing value imputation, word segmentation, and format standardization. The preprocessed historical multimodal data are labeled with configuration requirements, risk levels, resource heatmaps, and resource allocation schemes. A dataset is constructed based on the labeled historical multimodal data.

8. A virtual machine resource provisioning system combining multimodal risk assessment as described in claim 6, characterized in that: The model training module is specifically used for: The cloud platform divides the dataset into training, validation, and test sets in an 8:1:1 ratio. The order semantic recognition model, risk assessment model, and resource dynamic allocation model are trained using the corresponding data in the dataset. During training, hyperparameters including learning rate, learning decay rate, random dropout rate, batch size, and number of iterations are continuously optimized. The weights and biases of the order semantic recognition model, risk assessment model, and resource dynamic allocation model are updated using the backpropagation algorithm until the preset convergence conditions are met. The trained order semantic recognition model, risk assessment model, and resource dynamic allocation model are validated sequentially using the validation set. The validated order semantic recognition model, risk assessment model, and resource dynamic allocation model are tested sequentially using the test set. The tested order semantic recognition model, risk assessment model, and resource dynamic allocation model are then converted to ONNX format and deployed.

9. A virtual machine resource provisioning system combining multimodal risk assessment as described in claim 6, characterized in that: The resource heatmap construction module is specifically used for: The cloud platform collects computing resource data in real time, including at least CPU utilization, memory utilization, storage utilization, storage growth trend, network traffic, network type, network bandwidth, and network topology. It performs preprocessing on each computing resource data, including at least cleaning, transformation, and aggregation. Based on a preset display method, it maps the preprocessed computing resource data to obtain a resource heatmap. The resource heatmap is updated through a preset sliding time window. The purchase order acquisition module is specifically used for: The cloud platform obtains real-time virtual machine purchase orders carrying encrypted messages, hardware fingerprints, timestamps, and hash values. The encrypted messages are obtained by encrypting the order content and user account. The specific encryption rules are as follows: the order content and user account are encrypted into a first layer of encrypted data using the AES algorithm; this first layer is converted to hexadecimal data; the numbers 6 and A, and 9 and B are swapped in the hexadecimal data to obtain a second layer of encrypted data; this second layer is encrypted into a third layer using the 3DES algorithm; each character in the third layer is shifted 8 bits to the right to obtain a fourth layer of encrypted data; and this fourth layer is encrypted into an encrypted message using the RSA algorithm. The hardware fingerprint is obtained by concatenating the device serial number and MAC address. The hash value is obtained by hashing the encrypted message, hardware fingerprint, and timestamp. The cloud platform parses the received real-time virtual machine purchase order to obtain an encrypted message, hardware fingerprint, timestamp, and hash value. Integrity is verified using the hash value, timeliness is verified using the timestamp, and legality is verified using the hardware fingerprint. Then, the encrypted message is decrypted into four layers of encrypted data using the RSA algorithm. Each character in the four layers of encrypted data is shifted 8 bits to the left to obtain three layers of encrypted data. The three layers of encrypted data are then decrypted into two layers of encrypted data using the 3DES algorithm. The number 6 and the letter A, and the number 9 and the letter B in the two layers of encrypted data are swapped to obtain hexadecimal data. The hexadecimal data undergoes a base inversion to obtain one layer of encrypted data. The one layer of encrypted data is then decrypted into order details and user account information using the AES algorithm. Real-time multimodal data is then obtained based on the user account.

10. A virtual machine resource provisioning system combining multimodal risk assessment as described in claim 6, characterized in that: The virtual machine resource enabling module is specifically used for: The cloud platform analyzes the risk assessment results. When the risk assessment result is low risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, and automatically executes the virtual machine resource activation operation based on the resource allocation scheme. When the risk assessment result is medium risk, it inputs the configuration requirements and resource heatmap into the dynamic resource allocation model to obtain a resource allocation scheme, encrypts the resource allocation scheme, and pushes it to the pre-associated management terminal for approval in real time. When the risk assessment result is high risk, it refuses to execute the virtual machine resource activation operation. The activation result feedback module is specifically used for: The cloud platform generates activation results, indicating whether the activation was successful or failed. The activation results are then encrypted into first-level encrypted data using the RC6 algorithm. The first-level encrypted data is then divided into two levels of encrypted data by splitting the data in a 1:4 ratio and rearranging the order. Finally, the second-level encrypted data is encrypted into an encrypted notification using the Twofish algorithm, and the encrypted notification is then sent back. The cloud platform records real-time activation logs, encrypts the real-time activation logs into first encrypted data using the IDEA algorithm, adds a random string of a specified length to a specified position in the first encrypted data to obtain second encrypted data, encrypts the second encrypted data into an encrypted log using the SM4 algorithm, and stores and performs distributed backup of the encrypted log.

Images