A risk assessment-based hierarchical access control system for big data backup.
By extracting multidimensional data features and dynamically perceiving risks, combined with hierarchical access decision-making and refined authorization execution, the static problem of access control mechanisms in big data backup is solved, enabling real-time dynamic access control and improving the system's defense capabilities and data security.
Patent Information
- Application Number
- CN202610462577.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-09
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-04-09
AI Technical Summary
In the current big data backup field, access control mechanisms are too static and cannot detect changes in the security situation in real time. This results in insufficient defense strength during dynamic attacks, posing a risk of data leakage and damage.
Employing a multi-dimensional data feature extraction module, a dynamic risk perception module, a hierarchical access decision-making module, and a refined authorization execution module, this system achieves real-time protection of backup data through in-depth data deconstruction, real-time risk assessment, and dynamic adjustment of access control strategies, combined with data entropy monitoring, defense linkage mechanisms, and multi-factor verification.
It enables real-time dynamic adjustment of access control policies, enhances the big data backup system's ability to resist new dynamic attacks, ensures data security and availability, and prevents data contamination and internal unauthorized behavior.
Smart Images

Figure CN122333499B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data processing and information security, and specifically relates to a risk assessment-based hierarchical access control system for big data backup data. Background Technology
[0002] In the fields of big data storage and network security, data backup is a core means of ensuring data integrity and business continuity. As the last line of defense against hardware and software failures and malicious attacks, it holds an irreplaceable strategic position in modern information technology construction. With the explosive growth of massive amounts of data, how to build an efficient, stable, and secure backup mechanism has become crucial to safeguarding the core competitiveness of digital assets. Access control is a vital technical pillar to ensure that backup data is not illegally obtained or maliciously tampered with during storage and retrieval.
[0003] Among them, the backup data hierarchical access control technology for big data environments aims to achieve differentiated protection for large-scale datasets by classifying the importance and sensitivity of backup data from multiple dimensions and establishing a fine-grained permission mapping system accordingly. The core objective of this technology is to meet the needs of concurrent access by multiple users while minimizing unauthorized operations. Through strict verification and filtering of access requests, it enhances the backup system's ability to resist internal and external security threats.
[0004] Existing technologies primarily employ role-based or attribute-based access control models, with permission allocation typically completed during system initialization. This results in long lifecycles and a pronounced static nature to the granting and revocation of access permissions. However, in increasingly complex network environments, backup systems face highly dynamic and evolving external attacks and internal risks. Traditional static control mechanisms cannot detect fluctuations in environmental risks in real time. Especially during large-scale ransomware or distributed denial-of-service attacks, attackers can bypass existing defenses and damage backup data once they gain legitimate account access. Currently, the backup management field generally lacks a linkage mechanism that can dynamically adjust access thresholds based on real-time security conditions. This severe disconnect between permission control and risk perception directly leads to insufficient defense strength for backup data in extreme security events, posing a significant risk of data leakage and damage. Summary of the Invention
[0005] The purpose of this invention is to provide a risk assessment-based hierarchical access control system for big data backup, in order to solve the technical problems in the existing big data backup field where access control mechanisms are too static, unable to perceive changes in security status in real time, and have insufficient defense strength when subjected to dynamic attacks.
[0006] The technical solution of the present invention includes: The multidimensional data feature extraction module is used to perform deep deconstruction of the large-scale dataset to be backed up in order to extract multidimensional features and determine the security level of the backup data. The dynamic risk perception module is used to capture network environment, user behavior and external threat intelligence in real time, and perform nonlinear fusion through a deep neural network model to output a real-time risk assessment score. The tiered access decision module is used to dynamically calculate the access threshold required for the current access request based on the security level of the backup data and the real-time risk assessment score. The fine-grained authorization execution module, deployed on the data plane, is used to execute access control actions based on admission thresholds; The multidimensional data feature extraction module includes a data entropy monitoring unit. The data entropy monitoring unit is used to monitor the randomness change of the dataset to be backed up in real time before the data backup enters the pipeline. When an abnormal increase in data entropy is detected, the protection level of the batch of data is automatically upgraded to the highest level, and a data integrity verification alarm is triggered synchronously in the dynamic risk perception module. The refined authorization execution module includes a defense linkage mechanism. When the defense linkage mechanism determines that the current access behavior conforms to a high-risk attack mode, it triggers physical isolation protection and sends an isolation command to the physical controller of the storage array, so that the storage medium of the core backup volume is physically disconnected from the data bus and enters an unreadable and unwritable state.
[0007] Furthermore, the data entropy monitoring unit is specifically used for: By calculating the Shannon entropy of a data block, it can be determined whether the data has been processed by an unknown encryption algorithm. When the data entropy value exceeds a preset threshold within a short period of time, it is determined that the data may have been illegally encrypted by ransomware, and the protection level of the batch of data is upgraded to the highest level.
[0008] Furthermore, the defense linkage mechanism is specifically used for: Cut off access sessions and implement a global block, adding the source IP address of the access subject identified as a high-risk attack to the global blacklist; When the current access behavior is determined to be in line with a high-risk attack pattern, an isolation command is sent to the physical controller of the storage array to physically disconnect the storage media of the core backup volume from the data bus.
[0009] Furthermore, the dynamic risk perception module specifically includes: The behavior baseline construction unit is used to track and analyze the behavior sequence of the access subject based on the sliding window technique, and uses the dimensional average of the behavior vector of each time window as a stable behavior baseline. The behavior offset calculation unit is used to calculate the complement of the cosine similarity between the behavior vector of the current time window and the stable behavior baseline, as a behavior heterogeneity indicator to identify insider threats or account hijacking behavior.
[0010] Furthermore, the hierarchical access decision module specifically includes: A non-linear mapping matrix is used to define the admission thresholds for backup data of different security levels under different risk ranges; The time constraint unit is used to automatically apply a time constraint factor to the entry threshold during preset sensitive maintenance cycles or late-night periods to raise the entry threshold.
[0011] Furthermore, the refined authorization execution module also includes: Differentiated encryption units are used to execute differentiated encryption transmission strategies based on the security level of the accessed data. For data with extremely high security levels, the differentiated encryption unit implements a dynamic key destruction mechanism based on clock synchronization. The encryption key is bound to the current access session, and the key information is physically erased from the client's memory after the data transmission is completed or the session times out.
[0012] Furthermore, the refined authorization execution module also includes a multi-factor authentication engine, which dynamically requires access subjects to provide identity credentials of different strengths based on the access threshold. Identity credentials include one or more combinations of static passwords, biometric recognition, hardware token dynamic passwords, or manual authorization review.
[0013] Furthermore, it also includes a state feedback self-optimization module, which is used for: Record historical access behavior patterns and security incident outcomes to construct a sample set containing quantitative values of damage consequences; An incremental learning mechanism is adopted to iteratively optimize the weight parameters of the deep neural network model in the dynamic risk perception module based on the sample set, so that the risk assessment model can continuously learn from historical errors and self-correct.
[0014] Furthermore, it also includes an audit traceability unit, which is used to encapsulate the risk assessment process, access threshold change logs, authorization execution details, and defense linkage mechanism trigger records into blocks that conform to the blockchain data structure in real time, and write them into the consortium chain composed of multiple peer nodes through the consensus mechanism to achieve distributed storage and immutability of audit logs.
[0015] Furthermore, the system's logical architecture is divided into a control plane and a data plane; The control plane includes a multi-dimensional data feature extraction module, a dynamic risk perception module, a hierarchical access decision module, and a state feedback self-optimization module, which are used for the orchestration of global security policies, real-time assessment of risk status, and dynamic generation of access thresholds. The data plane includes a fine-grained authorization execution module, deployed on each storage node, used to perform specific traffic filtering, authentication, and read / write operations on backup data; The control plane and the data plane communicate through a pre-defined encrypted control channel.
[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention, through the collaborative work of multi-dimensional data feature extraction and dynamic risk perception modules, completely breaks through the static limitations of traditional access control mechanisms. The system can capture micro-changes in network traffic, user behavior, and environmental threats in real time, and transform these dynamic factors into quantified risk assessment scores. This makes access control policies no longer preset rigid rules, but an intelligent defense system that can evolve in real time with fluctuations in the security situation, greatly improving the ability of big data backup systems to resist new dynamic attacks.
[0017] 2. This invention introduces a non-linear mapping mechanism between risk scores and access thresholds to achieve deep coupling between data classification protection and risk environment. By dynamically adjusting the access thresholds for backup data of different levels within different risk ranges, this invention can ensure a convenient high-concurrency access experience in low-risk environments, while automatically switching to the ultimate security protection mode in high-risk environments. This differentiated control strategy not only ensures business continuity but also builds a robust isolation zone for core assets in extreme security incidents, effectively solving the dilemma between security and efficiency.
[0018] 3. This invention constructs a closed-loop architecture covering the entire chain from feature extraction, risk perception, decision-making, to refined execution, supplemented by a state feedback self-optimization mechanism. The system can not only execute strategies in real time, but also continuously correct its own defense logic through deep learning of historical access data. This self-evolving design ensures that the access control system can cope with the challenges of unknown threats. At the same time, by combining hardware isolation and blockchain auditing technology, this invention simultaneously strengthens the security of backup data at both the technical and legal levels, providing comprehensive and multi-dimensional security protection for enterprise-level big data assets.
[0019] 4. This invention achieves risk monitoring throughout the entire data lifecycle through sophisticated technical means such as data entropy monitoring and dynamic permission revocation. From the moment data enters the backup process to every stage of its retention and retrieval in the repository, the system implements real-time status verification. This full-process supervision model effectively prevents data pollution and internal unauthorized behavior, ensuring the absolute purity and usability of backup data as the bottom line for recovery. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of a risk-assessed, hierarchical access control system for big data backup data. Figure 2 This is a schematic diagram illustrating the core principles of the hierarchical access decision module; Figure 3 This is the logic flowchart of the multidimensional data feature extraction module; Figure 4 It is a schematic diagram of the collaborative interaction between the control plane and the data plane and the data flow. Detailed Implementation
[0021] This embodiment discloses a risk-assessment-based hierarchical access control system for big data backup, as shown in the attached figure. Figures 1-4 The system runs on a distributed architecture platform, with its logical architecture divided into a control plane and a data plane. The control plane is responsible for orchestrating global security policies, real-time risk assessment, and dynamic generation of access thresholds, while the data plane is deployed on each storage node, responsible for executing specific traffic filtering, authentication, and read / write operations on backup data. The control plane and data plane communicate through a pre-defined encrypted control channel, ensuring that the synchronization latency of flow table entries across the entire network is at the nanosecond level. The risk assessment-based big data backup data hierarchical access control system is implemented according to the following scheme: The first module is the multidimensional data feature extraction module, which serves as the system's input and is responsible for the deep deconstruction of the large-scale dataset to be backed up. This module employs a combination of full scanning and incremental probing techniques to perform semantic analysis on the raw bitstream of the storage system. During execution, the multidimensional data feature extraction module first initiates a metadata scanning program to extract basic attributes such as file creation time, owner permissions, file size, storage path, and access frequency.
[0022] Subsequently, the module invokes the content sensitivity recognition engine, utilizing a pre-defined regular expression library and natural language processing algorithms to identify personal identification information, core financial data, production process parameters, and corporate strategic plans contained within the data. The system incorporates a sensitivity weight matrix, pre-constructed based on industry data classification and grading standards and expert experience, assigning a sensitivity score between 1 and 10 to each type of sensitive information.
[0023] For example, personal identification information containing citizen ID numbers or bank card numbers is assigned a sensitivity score of 9, while general public business announcements are assigned a sensitivity score of 1. During operation, this weight matrix can be dynamically adjusted by the status feedback self-optimization module based on historical risk events.
[0024] The multidimensional data feature extraction module also includes a business relevance index calculation unit. This unit first constructs a business topology graph based on the enterprise's data flow relationships, where nodes represent data blocks and directed edges represent the data calling relationships between different business systems.
[0025] Subsequently, the unit calculates the in-degree of each data block, i.e., the number of times it is called by other business systems, as a quantitative indicator of its logical dependency. If the in-degree of a data block is greater than a preset association threshold, such as being called by three or more key business systems, then the business association index of that data block is set to a higher level, with the specific value determined according to a normalization formula. Calculation, where The in-degree of the current data block. This represents the maximum in-degree of all data blocks in the system.
[0026] The multidimensional data feature extraction module performs dimensionality normalization on the aforementioned metadata attributes, content sensitivity weights, and business relevance indicators, ultimately constructing a 128-dimensional spatial vector representing the characteristics of the backup data. This feature vector not only contains static attribute information but also reflects the importance of the data in dynamic business processes.
[0027] In addition, the multidimensional data feature extraction module integrates a data entropy monitoring unit. This unit monitors the randomness changes of the dataset to be backed up in real time before the data backup enters the pipeline. By calculating the Shannon entropy of the data blocks, the system can identify whether the data has been processed by unknown encryption algorithms.
[0028] When an abnormal increase in data entropy is detected within a short period of time, such as exceeding the preset threshold of 7.5, the system determines that the data may have been illegally encrypted by ransomware. At this time, the multi-dimensional data feature extraction module will automatically upgrade the protection level of the batch of data to the highest level and simultaneously trigger a data integrity verification alarm in the dynamic risk perception module, thereby preventing the compromised data from contaminating the backup repository.
[0029] Through the above mechanism, the multidimensional data feature extraction module completes the multidimensional feature extraction and initial classification of the backup data, providing a data foundation for subsequent dynamic risk perception and hierarchical access decisions.
[0030] Next, we have the dynamic risk awareness module, whose core function is to construct a quantitative map that reflects the security posture of the backup storage environment in real time. This module uses multi-source data fusion and deep learning technology to transform abstract network environment, user behavior, and external threat intelligence into dynamic risk assessment scores, providing real-time basis for subsequent access control decisions. Its specific implementation is as follows.
[0031] The module deploys high-performance traffic probes at the edge nodes of the backup storage network. These probes capture key characteristics of underlying communication packets in real time with microsecond-level precision, including the time interval between packet arrivals, the length distribution of data packets, and abnormal states of flag bits in the protocol header. This allows for the quantification of raw network characteristics into network environment risk values. The system further defines the following three quantifiable indicators: Variance of message arrival time interval It is used to reflect the burstiness and stability of network traffic; Shannon entropy of packet length distribution This is used to detect whether there are abnormal packet size patterns. Number of times the protocol header error flag appears per unit time These are used to identify potential scanning or attack behaviors. The three indicators mentioned above are processed using the min-max normalization method, ensuring their values are all between 0 and 1:
[0032] in, , The system's preset upper and lower bounds for variance. , The upper and lower bounds of the entropy value. , These serve as the upper and lower bounds for the anomaly flag count. Ultimately, the network environment risk value is calculated through weighted fusion:
[0033] In the formula, , , For the pre-set weighting coefficients, satisfying The initial value can be set based on expert experience and dynamically corrected by the system's subsequent self-optimization module. The value ranges from 0 to 1, with higher values indicating a less trustworthy network environment.
[0034] The module introduces sliding window technology to dynamically track and analyze the behavioral sequences of accessing subjects. The sliding window step size is set to 1 hour, and the total window length is 24 hours, ensuring that it can capture both short-term behavioral changes and reflect long-term behavioral trends. The system continuously records three types of behavioral elements for each accessing subject: operation type (such as read, write, delete), accessed backup data path, and executed command sequence.
[0035] To transform the sequence of actions into a computable vector, the system employs the following encoding method: predefined This operation type is converted using one-hot encoding. dimensional vector; predefined Types of access paths, converted to one-hot encoding dimensional vector; The instruction sequence uses word embedding to map each instruction to a fixed-length vector, and the average of all instruction vectors within the window is taken to obtain a... Dimensional vector.
[0036] By piecing together the three parts mentioned above, a final product is formed. dimensional behavior vector For each access subject, the system uses the one-dimensional average of its behavior vectors for each time window over the past 30 days as a stable behavioral baseline. .
[0037] Behavioral heterogeneity index Defined as the behavior vector of the current time window With baseline vector The complement of the cosine similarity:
[0038] In the formula, the numerator represents the dot product of two vectors, and the denominator represents the product of the magnitudes of the two vectors. When the current behavior is completely consistent with the historical baseline, the cosine similarity is 1. When the behaviors are completely divergent, the cosine similarity approaches 0. Approaching 1. For example, if a user suddenly accesses a large number of highly sensitive data volumes they have never touched before during non-working hours, the similarity between their behavior vector and the baseline vector will significantly decrease. The value spiked rapidly, enabling precise identification of potential insider threats or account hijacking.
[0039] The module accesses and analyzes global security alerts in real time through a standardized external threat intelligence interface. This information includes, but is not limited to, ransomware signatures, botnet IP address blacklists, and zero-day vulnerability alerts for storage systems.
[0040] The system maintains a threat signature database to store the latest acquired intelligence. For each access request, the system checks whether its source IP address matches the IP blacklist and whether the characteristics of the accessed data match ransomware signatures. External threat intelligence risk value. The quantification method is as follows: If the access request directly hits the blacklist or the signature, then If a zero-day vulnerability alert exists but is not directly triggered, the system sets a base risk value (e.g., 0.3) based on the severity of the alert; if there are no matches and no relevant alerts, then... .
[0041] Through this simple rule mapping Similarly, the values are normalized to between 0 and 1, with higher values indicating a higher risk from external intelligence.
[0042] The core of the dynamic risk perception module is its internally integrated deep neural network risk assessment model. This model is responsible for applying the aforementioned... , , The features from the three dimensions are nonlinearly fused to generate a comprehensive real-time risk assessment score. The model's construction and training process is as follows.
[0043] The model employs a multi-layer fully connected neural network. The input layer contains three neurons, each corresponding to a normalized input neuron. , , The model contains two hidden layers: the first hidden layer contains 64 neurons, and the second hidden layer contains 32 neurons. Both layers use the Corrected Linear Unit (ReLU) activation function. The output layer is a single neuron; the output value is compressed to between 0 and 1 using the Sigmoid activation function, which becomes the final risk assessment score. .
[0044] The model's training data comes from access logs recorded during the system's historical operation. Each training sample consists of a feature vector. and corresponding tags composition.
[0045] Feature vector Extracted from historical logs The triples are calculated using the methods described above.
[0046] Label :Label This is the model's "standard answer," with a value between 0 and 1. For each historical access record, the system labels it based on post-audit results. If the access is confirmed as a security incident, such as a successful ransomware intrusion or data breach, the sample's label... It is marked as 1; if this access is a normal business operation, then the tag is... The value is 0; for abnormal behavior in between, security experts assign a value between 0 and 1 based on the severity of the event, such as 0.6 indicating "suspicious but not causing substantial harm".
[0047] The goal of model training is to improve the predicted risk value output by the network. To approximate the true label value as closely as possible. Therefore, mean squared error is used as the loss function. :
[0048] in, This represents the number of samples in a single batch used for training. Representing the The true label of each sample Representative model for the first The predicted risk value output for each sample.
[0049] The training process employs the backpropagation algorithm and the Adam optimizer, iteratively updating the weights and biases in the network to optimize the loss function. The value of ...
[0050] Once the model is trained, it can be put into real-time operation. For each access request, the module will calculate the required parameters in real time. , , The three dimensions are used as input features and fed into a trained deep neural network model for forward propagation computation. The model outputs the result. This is the real-time risk assessment score for the current access request. The score ranges from 0 to 1. The closer the value is to 1, the higher the risk the system determines the current access behavior faces.
[0051] In summary, the dynamic risk perception module generates network environment risk values by deploying traffic probes, building behavioral baselines, and integrating external threat intelligence. Behavioral heterogeneity indicators External threat intelligence risk value These three values serve as input to the deep neural network model, and after nonlinear fusion, they output a comprehensive real-time risk assessment score. This score provides accurate and real-time decision-making basis for the downstream "hierarchical access decision module," enabling access control policies to dynamically evolve with the security situation, thereby achieving efficient and adaptive protection of backup data.
[0052] Next is the hierarchical access decision module, which is the decision-making center of the entire system. Its core responsibility is to dynamically calculate the access threshold required for each access request based on the security level of the backup data itself and the real-time risk status of the system. The decision results of this module will directly guide the downstream fine-grained authorization execution module in performing access control. Its specific implementation process is as follows.
[0053] The hierarchical access decision module first receives a feature vector output from the multidimensional data feature extraction module. This feature vector is a 128-dimensional spatial vector that comprehensively reflects the meta-attributes, content sensitivity, and business relevance of the backup data.
[0054] The module employs a clustering algorithm based on Euclidean distance to classify the feature vector into different security levels. The system pre-defines four cluster centers in a 128-dimensional feature space, corresponding to four security levels: extremely high, high, medium, and low. These four cluster centers are obtained through unsupervised learning from a large amount of historical data and are periodically corrected by a state feedback self-optimization module during system operation. For backup data currently awaiting classification, the module calculates the Euclidean distance between its feature vector and the four cluster centers and assigns it to the security level corresponding to the nearest cluster center.
[0055] Through the above methods, backup data is clearly divided into four security levels: extremely high, high, medium, and low, laying the foundation for subsequent differentiated access control.
[0056] After obtaining the security level of the backup data, the hierarchical access decision module combines the real-time risk assessment score output by the dynamic risk perception module. It dynamically calculates the admission threshold required for the current access request.
[0057] The module internally maintains a non-linear mapping matrix, which defines the access thresholds for backup data at different security levels under different risk ranges. The access threshold is a dimensionless value ranging from 0 to 10; a higher value indicates a stronger requirement for authentication of the accessing entity. The specific mapping rules are as follows: When the real-time risk assessment score When the risk level is in the low-risk range of 0 to 0.3, the system determines that the current security situation is relatively stable. At this time, for backup data with low and medium security levels, the access threshold is kept at a low basic level (e.g., set to 2 and 4 respectively) to ensure normal business access efficiency; for backup data with high and very high security levels, the access threshold is set to 6 and 8 respectively to achieve differentiated protection.
[0058] When the real-time risk assessment score When the risk level rises to the medium risk range of 0.3 to 0.7, the system determines that a certain security threat exists. At this point, the module automatically increases the access threshold for backup data of medium security level and above. Specifically, the access threshold for medium security level data is increased from the basic level 4 to 6, the access threshold for high security level data is increased from 6 to 8, and the access threshold for very high security level data is increased from 8 to 9. The access threshold for low security level data remains unchanged to maintain the availability of basic business operations.
[0059] When the real-time risk assessment score When the threshold exceeds 0.7, the system enters a high-risk defense mode. In this mode, the module forcibly raises the access threshold for all security level backup data to the maximum limit of 10. This means that regardless of the level of data requested by the accessing subject, it must accept the highest strength of authentication that the system can provide.
[0060] The above mapping relationship is achieved through a nonlinear mapping matrix. Implementation, in which The output is the admission threshold. This is a real-time risk assessment score. This indicates the security level for backup data. The specific values in this matrix can be configured and adjusted according to the security policies of the actual deployment environment.
[0061] To further enhance security during periods of low staffing, the tiered access decision module incorporates a time-based constraint factor into the calculation of access thresholds. The system predefines two time intervals: a sensitive maintenance period, for example, from 2:00 AM to 4:00 AM every Sunday, and a late-night period, for example, from 12:00 AM to 6:00 AM daily.
[0062] When an access request occurs within any of the preset time intervals mentioned above, the module will automatically apply a time constraint factor, with a value of 1.5, to the admission threshold calculated in step 2. Specifically, the final admission threshold... Calculate using the following formula:
[0063] in, For step 2, based on risk score and data level The calculated admission threshold, This is a time constraint factor, with a value of 1 during normal time periods and a value of 1.5 during preset sensitive maintenance cycles or late-night periods. The function ensures that the final admission threshold does not exceed the upper limit of 10. For example, if an access request is initiated late at night, and the admission threshold is calculated based on its risk level and data grade... If it is 6, then for .
[0064] In this way, the system is more sensitive to any subtle abnormal behavior during off-peak hours, thus enabling stricter access restrictions to be implemented during periods when personnel are less on duty, effectively reducing the risk of internal leaks or latent attacks.
[0065] After completing the above calculations, the hierarchical access decision module will finalize the admission threshold. The request is then sent to the refined authorization execution module. Simultaneously, key information in this decision-making process, including the data security level corresponding to the current access request, the real-time risk assessment score, the effectiveness status of the time constraint factor, and the final output admission threshold, is recorded and transmitted to the status feedback self-optimization module as the basis for subsequent optimization decision logic.
[0066] In summary, the hierarchical access decision module determines the security level of backup data by receiving multi-dimensional feature vectors, and combines this with the real-time risk assessment score output by the dynamic risk perception module. By using a nonlinear mapping matrix to calculate the basic access threshold, and then introducing a time-dimensional constraint factor for correction, the final output is an access threshold that can dynamically change with the security situation and time window. This mechanism ensures that access control policies can adapt to environmental changes in real time, guaranteeing business efficiency during low-risk periods and building a robust protective barrier for core assets during high-risk or sensitive periods, providing accurate and dynamic execution basis for downstream refined authorization execution modules.
[0067] The refined authorization execution module is the core execution unit for policy implementation. Deployed on the data plane, it is responsible for receiving access thresholds issued by the hierarchical access decision module and translating them into specific access control actions. This module covers core functions such as multi-factor authentication, differentiated data encryption, defense-linked response, and dynamic permission revocation. Its specific implementation is as follows.
[0068] This module includes a multi-factor authentication engine, which is used to determine the admission threshold issued by the hierarchical access decision module. The system dynamically requires users to provide authentication credentials of varying strengths. (Access threshold) The value ranges from 0 to 10, with higher values indicating stronger authentication requirements. The engine performs verification according to the following three-tiered strategy: when At this point, the system determines that the current risk level is low or low-to-medium risk. In this case, the verification engine only requires the accessing entity to provide a static password or digital certificate.
[0069] when At this point, the system determines that the system is currently in a medium-to-high risk state. In this situation, the verification engine forcibly adds a second factor verification on top of the static password or digital certificate. The second factor verification options include biometric recognition or a hardware token dynamic password. Biometric recognition supports both fingerprint and facial recognition, while the hardware token dynamic password uses a time-synchronized one-time password algorithm.
[0070] when At this point, the system determines that the current situation is extremely high-risk. The verification engine then activates the highest level of manual authorization review mode. In this mode, the access request is first suspended, and the system pushes an authorization request to the preset security administrator terminal, including the access subject's identity information, the requested data digest, the current risk assessment score, and contextual information. The security administrator manually approves the request through a dedicated management interface, and only after approval can access be granted. Simultaneously, the system automatically triggers a full-process video audit linkage, recording the access subject's operations in real time. The recorded files are stored in conjunction with the access logs to ensure that every operation involving core data has strong non-repudiation.
[0071] When performing backup data read operations, the fine-grained authorization execution module employs differentiated encryption transmission strategies based on the security level of the accessed data. The system predefines four encryption levels, corresponding to low, medium, high, and very high data security levels: For low-security data, standard transport layer security protocols are used for encrypted transmission. For medium-security data, transport layer security protocols are used, and a two-way certificate verification mechanism is enabled. For high-security data, application layer end-to-end encryption is superimposed on the transport layer security protocol, with the key generated and managed independently by the data plane.
[0072] For data with extremely high security requirements, the system employs end-to-end encrypted transmission and implements a dynamic key destruction mechanism based on clock synchronization. Specifically, the encryption key is not persistently stored on the client side but is dynamically generated by the control plane after authorization. The key's lifespan is tied to the current access session. The client uses the key to decrypt data within its validity period. Once data transmission is complete or the session times out, the client immediately physically erases the key information from memory. This mechanism ensures that even if the client system is compromised after access is complete, attackers cannot obtain historical keys from persistent storage or memory remnants.
[0073] The refined authorization execution module continuously monitors the behavioral characteristics of the access subject during the verification process. If the module detects that the access subject's real-time credit credentials are significantly lower than the current access threshold, such as repeatedly entering incorrect static passwords, or if its behavioral characteristics match known attack patterns, such as initiating a large number of failed verification requests from the same IP address within a short period of time, the system immediately triggers the defense linkage mechanism.
[0074] The execution of the defense linkage mechanism consists of three steps: The first step is to terminate the access session. The module immediately terminates the connection of the current access subject and issues a drop rule to the forwarding unit in the data plane, rejecting all subsequent packets of this session.
[0075] The second step is to implement a global block. The module reports the source IP address of the accessing entity to the control plane, which adds the IP address to the global blacklist and synchronizes the blacklist entry with all data plane nodes. Any access requests initiated by the blocked IP address within the next 48 hours will be directly rejected by the data plane and will not proceed to the verification process.
[0076] The third step is to trigger physical isolation protection. When the module determines that the current access behavior conforms to a high-risk attack mode, the system automatically sends an isolation command to the physical controller of the storage array. The physical controller of the storage array drives an offline switch, physically disconnecting the storage medium of the core backup volume from the data bus and putting it into a non-readable and non-writable state. This physical isolation operation is independent of the upper-layer operating system; even if the control plane is compromised, the isolation state is still maintained by the hardware logic of the storage array.
[0077] In the instant before physical isolation takes effect, the system generates a read-only snapshot in the protected security isolation zone. This snapshot is stored in a separate storage area physically isolated from the main storage volume, serving as the last line of defense for disaster recovery and ensuring that even in extreme ransomware penetration scenarios, the enterprise can still obtain a clean copy of the original data.
[0078] The refined authorization execution module has continuous behavior monitoring and dynamic permission revocation capabilities. During the authorization access process, the module continuously monitors the subject's subsequent behavior, including the accessed data objects, operation types, operation frequency, and data flow paths.
[0079] The module maintains a set of rules for judging abnormal behavior. The set of rules includes the following judgment conditions: the subject attempts to access multiple data volumes with different security levels within a short period of time after authorization; the total amount of data downloaded by the subject after authorization exceeds a preset threshold; the subject attempts to write data to a target storage location not declared in the access request; the subject's subsequent behavioral characteristics deviate significantly from the behavioral baseline established in the verification phase.
[0080] When any of the above conditions are triggered, the module immediately determines that the subject is a controlled account or an insider threat and performs dynamic privilege revokement. The revokement action includes: instantly revoking the access token currently held by the subject, making the token immediately invalid in the token cache; synchronizing the token invalidation information to all data plane nodes to ensure consistency of privilege revokement in a distributed environment; and locking the storage port associated with the subject to prevent any subsequent access requests on that port.
[0081] In summary, the refined authorization execution module dynamically adapts access thresholds through a multi-factor authentication engine, ensures data confidentiality during transmission through differentiated encryption strategies, rapidly disconnects risky links and implements physical isolation upon detecting an attack through a defense linkage mechanism, and continuously monitors and handles anomalies in authorized sessions through a dynamic permission revocation function. This module transforms the abstract access thresholds issued by the hierarchical access decision module into specific, executable, and multi-layered access control actions, forming the final link in the entire system's defense architecture, ensuring that only legitimate and compliant access requests can access core backup data.
[0082] Finally, there is the state feedback self-optimization module, which implements the system's closed-loop evolution mechanism. This module continuously learns from historical access behavior, dynamically adjusting the weight parameters in the risk assessment model and the business relevance indicators in data feature extraction. This allows the system to continuously improve its accuracy in identifying new threats over time. Its specific implementation is as follows.
[0083] The module continuously records the complete behavioral trajectory of each access process, including: the identity of the accessing entity, the data object requested for access, the operation type, and the real-time risk assessment score output by the dynamic risk perception module. The access threshold calculated by the hierarchical access decision module, the execution result of the fine-grained authorization execution module, and the final security event handling result.
[0084] The system stores the above information in the form of structured logs, with each log record corresponding to a complete access request. Simultaneously, the module labels the final result of each access, forming the sample labels required for supervised learning. The labeling method is as follows: If the access is confirmed as a security incident in a subsequent audit, such as a successful ransomware intrusion, data breach, or unauthorized internal operation, then the quantified damage consequence value for this sample is... Marked as 1; If the access was a normal business operation and did not cause any security incidents, then Marked as 0; For abnormal behavior falling between these two extremes, security experts assign it a midpoint value between 0 and 1 based on the severity of the event; for example, 0.6 indicates "suspicious but not causing substantial harm".
[0085] The core optimization logic of the state feedback self-optimization module targets the deep neural network model in the dynamic risk perception module. The input to this model consists of three-dimensional feature values. , , The output is a real-time risk assessment score. The weight parameters of the model constitute the coefficient matrix. .
[0086] The module employs an incremental learning mechanism, periodically retrieving a batch of abnormal access samples and their corresponding quantitative values of damage consequences from historical logs. The model parameters are iteratively optimized. The optimization process follows the formula:
[0087] in: This represents the updated risk assessment coefficient matrix, which is the set of all weight parameters in the deep neural network; This represents the original coefficient matrix before the update; This is the learning rate coefficient, used to control the step size of each update. Its value ranges from 0 to 1, with an initial value set to 0.01, and can be dynamically adjusted according to the training convergence. The value is a quantitative value for the actual damage consequences, ranging from 0 to 1, and is determined by the aforementioned labeling process; The real-time risk assessment score output by the dynamic risk perception module; Risk score Relative to each dimension of features , , The gradient vector, which is the partial derivative of the loss function with respect to the weights of each layer of the model, is calculated using the backpropagation algorithm.
[0088] The essence of this formula is: when the system predicts the risk probability... Less than the actual damage occurred When the system underestimates the risk, When the gradient is positive, the update direction causes the model to output a higher risk score when encountering similar features in the future; conversely, when the system overestimates the risk, the update direction causes the model to output a lower risk score. Through this closed-loop iteration, the system can automatically correct the weight allocation in the dynamic risk perception module, so that its accuracy in identifying new threats continuously improves with increasing runtime.
[0089] In addition to optimizing the risk perception weights, the state feedback self-optimization module is also responsible for dynamically adjusting the business relevance index in the multi-dimensional data feature extraction module. This index reflects the core nature of a data block in the business flow, and its calculation depends on the in-degree of each data block in the business topology graph.
[0090] The module continuously monitors the actual frequency of backup data access by various business systems and performs statistical analysis on a monthly basis. For a given data block, if the number of actual calls within the past 30 days increases by more than 20% compared to the previous 30 days, the module determines that its business relevance has increased and automatically increases the business relevance weight of that data block in the feature vector by 0.1, with a maximum of 1.0. Conversely, if the number of calls decreases by more than 20%, the weight is reduced accordingly.
[0091] Through the above adjustments, the feature vectors output by the multidimensional data feature extraction module can more accurately reflect the actual importance of data in a dynamic business environment, avoiding classification deviations caused by initial setting delays.
[0092] In summary, the state feedback self-optimization module constructs a sample set containing quantified damage consequences by recording access behavior trajectories and security event outcomes. Based on this, the module uses incremental learning and gradient descent algorithms to iteratively optimize the weights of the deep neural network in the dynamic risk perception module, enabling the risk assessment model to continuously learn from historical errors and self-correct. Simultaneously, the module dynamically adjusts the business relevance indicators in the multi-dimensional data feature extraction module based on changes in the actual frequency of data access by business systems. This closed-loop design of the module enables the entire access control system to self-evolve, adapting to constantly changing security threat environments and business needs, providing crucial assurance for the long-term stable operation of the system.
[0093] To ensure the absolute authenticity of audit information, the system integrates an audit traceability unit. This unit encapsulates all risk assessment processes, access threshold change logs, execution details of each authorization, and trigger records of defense linkage mechanisms into blocks conforming to a blockchain data structure in real time. Each block contains a block header and a block body. The block body stores the hash values of the aforementioned audit events, while the block header records the hash value of the previous block, a timestamp, and the root hash calculated using the Merkle tree algorithm.
[0094] The system constructs a consortium blockchain consisting of 5 peer nodes as a notarization platform. These 5 nodes are deployed on independent servers in different geographical locations and managed by different institutions or departments. When the audit traceability unit generates a new block, it broadcasts it to all peer nodes through a practical Byzantine fault-tolerant consensus mechanism. The new block must be verified by at least 4 nodes before it can be officially written into the on-chain ledger of each node, achieving distributed storage and consistency assurance.
[0095] Through the above mechanism, audit logs are permanently stored in the blockchain evidence storage platform. Based on the chain structure, consensus verification, and timestamp characteristics of blockchain, even if the system administrator's privileges are hijacked, attackers cannot unilaterally tamper with or erase the audit records already uploaded to the chain. This provides legal-level technical evidence for determining responsibility and collecting evidence after a security incident.
[0096] The system constructed in this embodiment solves the problem of sluggishness in traditional access control when facing complex security situations through deep collaboration of multi-dimensional feature extraction, real-time risk perception, and dynamic decision execution. The system no longer relies on static, preset blacklist and whitelist rules, but instead constructs an intelligent defense barrier that can perceive micro-changes in the environment and has the ability to evolve itself.
[0097] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.
[0098] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A big data backup data tiered access control system based on risk assessment, characterized in that, include: The multidimensional data feature extraction module is used to perform deep deconstruction of the large-scale dataset to be backed up in order to extract multidimensional features and determine the security level of the backup data. The dynamic risk perception module is used to capture network environment, user behavior and external threat intelligence in real time, and perform nonlinear fusion through a deep neural network model to output a real-time risk assessment score. The tiered access decision module is used to dynamically calculate the access threshold required for the current access request based on the security level of the backup data and the real-time risk assessment score. The fine-grained authorization execution module, deployed on the data plane, is used to execute access control actions based on admission thresholds; The multidimensional data feature extraction module includes a data entropy monitoring unit. This unit monitors the randomness changes of the dataset to be backed up in real time before the data backup enters the pipeline. By calculating the Shannon entropy of the data blocks, it identifies whether the data has been processed by an unknown encryption algorithm. When the data entropy value exceeds a preset threshold within a short period of time, it determines that the data may have been illegally encrypted by ransomware, automatically upgrades the protection level of the batch of data to the highest level, and synchronously triggers a data integrity verification alarm in the dynamic risk perception module. The dynamic risk perception module specifically includes: a behavior baseline construction unit, used to track and analyze the behavior sequence of the access subject based on sliding window technology, using the dimensional average of the behavior vector of each time window as a stable behavior baseline; and a behavior offset calculation unit, used to calculate the complement of the cosine similarity between the behavior vector of the current time window and the stable behavior baseline, as a behavior heterogeneity indicator to identify internal threats or account hijacking behavior. The hierarchical access decision module specifically includes: a nonlinear mapping matrix, used to define the access threshold for backup data of different security levels under different risk intervals; and a time constraint unit, used to automatically apply a time constraint factor to the access threshold during preset sensitive maintenance cycles or late-night periods to increase the access threshold. The refined authorization execution module includes a multi-factor authentication engine, which dynamically requires the access subject to provide identity credentials of different strengths based on the access threshold. The identity credentials include one or more combinations of static passwords, biometric recognition, hardware token dynamic passwords, or manual authorization review. The refined authorization execution module includes a defense linkage mechanism, which is used to: cut off access sessions and implement a global block, adding the source IP address of the access subject identified as a high-risk attack to the global blacklist; when the current access behavior is determined to meet the high-risk attack mode, trigger physical isolation protection, send an isolation command to the physical controller of the storage array, physically disconnect the storage medium of the core backup volume from the data bus, and put it into a non-readable and non-writable state; and before the physical isolation takes effect, generate a read-only instantaneous snapshot of the core backup volume in the protected security isolation area. This snapshot is stored in an independent storage area physically isolated from the main storage volume, serving as the last line of defense for disaster recovery.
2. The system of claim 1, wherein, The refined authorization execution module also includes: Differentiated encryption units are used to execute differentiated encryption transmission strategies based on the security level of the accessed data. For data with extremely high security levels, the differentiated encryption unit implements a dynamic key destruction mechanism based on clock synchronization. The encryption key is bound to the current access session, and the key information is physically erased from the client's memory after data transmission is completed or the session times out.
3. The system of claim 1, wherein, It also includes a state feedback self-optimization module, which is used for: Record historical access behavior patterns and security incident outcomes to construct a sample set containing quantitative values of damage consequences; An incremental learning mechanism is adopted to iteratively optimize the weight parameters of the deep neural network model in the dynamic risk perception module based on the sample set, so that the risk assessment model can continuously learn from historical errors and self-correct.
4. The system of claim 1, wherein, It also includes an audit traceability unit, which is used to encapsulate the risk assessment process, access threshold change logs, authorization execution details and defense linkage mechanism trigger records into blocks that conform to the blockchain data structure in real time, and write them into the consortium chain composed of multiple peer nodes through the consensus mechanism to achieve distributed storage and immutability of audit logs.
5. The system of any one of claims 1 to 4, wherein, The system's logical architecture is divided into a control plane and a data plane; The control plane includes the multi-dimensional data feature extraction module, the dynamic risk perception module, the hierarchical access decision module, and the state feedback self-optimization module, which are used for the orchestration of global security policies, real-time assessment of risk status, and dynamic generation of access thresholds. The data plane includes the fine-grained authorization execution module, which is deployed on each storage node to perform specific traffic filtering, authentication, and read / write operations on backup data. The control plane and the data plane communicate through a preset encrypted control channel.
Citation Information
Patent Citations
Classified protection method and dynamic protection system for computer data
CN120162764A
Enterprise data security management system and method
CN120915571A
Automated ransomware identification and recovery
US20190235973A1
