Security defense system, method, medium and product for educational training platform

Abstract

The invention provides a security defense system and method for an educational training platform, a medium and a product. The system comprises a traffic protection layer used for processing network traffic to obtain first feature data; the container security layer is used for processing the network traffic to obtain second feature data; the privacy protection layer is used for processing the network traffic to obtain third feature data; the gating router is used for constructing feature vectors based on the three kinds of feature data, determining weights of n expert models according to the feature vectors, arranging the n expert models according to a descending order of the weights, and selecting the first k expert models to identify attack behavior information in the network traffic, k is greater than or equal to 1 and less than or equal to n, and k and n are positive integers; and the intelligent decision-making layer is used for generating a defense strategy based on the attack behavior information, evaluating the effect of the defense strategy to obtain effect information corresponding to each expert model, and feeding back the effect information to the gating router.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of data security, and in particular to a security defense system, method, medium and product for an education training platform. BACKGROUND

[0002] In the field of cloud-native security technology, the technology ecosystem centered on Kubernetes has matured and formed a complete technical system. This system mainly includes three key components: first, microservice governance technology, which realizes secure communication and traffic management between services through Service Mesh; second, container runtime protection technology, including monitoring and anomaly detection of container behavior; and finally, security assurance of continuous delivery pipeline to ensure the security of the software supply chain. These technologies have formed standardized implementation solutions and tool chains in practice, such as Falco (open-source security tool) runtime detection system and Trivy image scanning tool, etc.

[0003] As an important architecture in the field of deep learning, the core features and technical principles of Mixture of Experts (MoE) have been widely researched and applied. In natural language processing, computer vision and other fields, MoE architecture has shown excellent performance.

[0004] In the field of dynamic defense technology, Mobile Target Defense (MTD) has developed various implementation methods and technical routes. The most representative ones include network topology dynamic reconstruction technology, service instance random migration technology and system configuration dynamic change technology, etc. These technologies increase the uncertainty and randomness of the system to increase the difficulty of intrusion for attackers. SUMMARY

[0005] One technical problem solved by the present disclosure is that in the prior art, when facing new or composite attacks, the system is not easy to identify complex threat features.

[0006] In view of this, embodiments of the present disclosure provide a security defense system and / or method for an education training platform to facilitate detection and identification of complex threat features.

[0007] According to one aspect of the present disclosure, a security defense system for an education training platform is provided, comprising: a traffic protection layer configured to process network traffic to obtain first feature data, the first feature data being feature data related to traffic data of the network traffic; a container security layer configured to process the network traffic to obtain second feature data, the second feature data being feature data related to container attack data of the network traffic; a privacy protection layer configured to process the network traffic to obtain third feature data, the third feature data being feature data related to privacy data of the network traffic; a gating router configured to receive the first feature data, the second feature data and the third feature data, construct a feature vector based on the first feature data, the second feature data and the third feature data, determine a weight of each expert model in n expert models according to the feature vector, arrange the n expert models in a descending order of the weight, select a first k expert models to identify attack behavior information in the network traffic, where 1≤k≤n, and k and n are positive integers; and an intelligent decision layer configured to generate a defense strategy based on the attack behavior information, evaluate an effect of the defense strategy after the defense strategy is executed to obtain effect information corresponding to each expert model of the first k expert models, and feed back the effect information to the gating router.

[0008] In some embodiments, the security defense system further comprises: the gating router is further configured to update the weight of each expert model based on the current weight of each expert model and the effect information corresponding to each expert model.

[0009] In some embodiments, the attack behavior information comprises: a first threat feature vector obtained based on the first feature data, a second threat feature vector obtained based on the second feature data, and a third threat feature vector obtained based on the third feature data.

[0010] In some embodiments, the n expert models comprise: a traffic analysis expert model configured to perform space-time feature analysis on the network traffic to detect network layer attack behavior; a container security expert model configured to detect attack behavior of the network traffic on containers; a federated privacy protection expert model configured to detect attack behavior of the network traffic on privacy data, and implement cross-tenant knowledge sharing under privacy protection; and a decision expert model configured to generate an adaptive defense strategy based on multi-source threat intelligence obtained from the network traffic analysis, the multi-source threat intelligence comprising the first threat feature vector, the second threat feature vector and the third threat feature vector.

[0011] In some embodiments, the traffic analysis expert model is further configured to generate a defense strategy against the network layer attack behavior based on single-source threat intelligence of the network layer attack behavior, the container security expert model is further configured to generate a defense strategy against the container attack behavior based on single-source threat intelligence of the attack behavior of the container security layer to the network traffic, and the federated privacy protection expert model is further configured to generate a defense strategy against the privacy data attack behavior based on single-source threat intelligence of the attack behavior of the privacy protection layer to the privacy data.

[0012] In some embodiments, the first feature data includes an entropy value and an IP dispersion degree obtained by the traffic protection layer in processing the network traffic, the second feature data includes a system call frequency and a vulnerability score obtained by the container security layer in processing the network traffic, and the third feature data includes a gradient norm of privacy data obtained by the privacy protection layer in processing the network traffic.

[0013] In some embodiments, the gating router is further configured to obtain a network traffic abnormality degree based on the first feature data, obtain a vulnerability exploitation severity based on the second feature data, obtain a data sensitivity based on the third feature data, calculate a threat evaluation value based on the network traffic abnormality degree and a weight coefficient corresponding to the network traffic abnormality degree, the vulnerability exploitation severity and a weight coefficient corresponding to the vulnerability exploitation severity, and the data sensitivity and a weight coefficient corresponding to the data sensitivity, determine a threat level to which the attack behavior of the network traffic belongs based on the threat evaluation value, and determine the server resource to be used based on the threat level.

[0014] In some embodiments, the threat level includes a low threat level, a medium threat level, and a high threat level, the high threat level is higher than the medium threat level, and the medium threat level is higher than the low threat level; the gating router is configured to determine that the attack behavior of the network traffic belongs to the low threat level when the threat evaluation value is within an evaluation value range corresponding to the low threat level, determine that the attack behavior of the network traffic belongs to the medium threat level when the threat evaluation value is within an evaluation value range corresponding to the medium threat level, and determine that the attack behavior of the network traffic belongs to the high threat level when the threat evaluation value is within an evaluation value range corresponding to the high threat level.

[0015] According to another aspect of this disclosure, a security defense method for an education and training platform is provided, comprising: processing network traffic using a traffic protection layer to obtain first feature data, wherein the first feature data is feature data related to traffic data of the network traffic; processing the network traffic using a container security layer to obtain second feature data, wherein the second feature data is feature data related to container attack data of the network traffic; processing the network traffic using a privacy protection layer to obtain third feature data, wherein the third feature data is feature data related to privacy data of the network traffic; and receiving the first feature data and the second feature data using a gated router. Based on the first feature data, the second feature data, and the third feature data, a feature vector is constructed. The weight of each of the n expert models is determined according to the feature vector. The n expert models are arranged in descending order of weight. The first k expert models are selected to identify attack behavior information in the network traffic, where 1 ≤ k ≤ n, and k and n are positive integers. A defense strategy is generated based on the attack behavior information using an intelligent decision layer. After the defense strategy is executed, its effectiveness is evaluated to obtain effect information corresponding to each of the first k expert models. This effect information is then fed back to the gated router.

[0016] In some embodiments, the security defense method further includes: using the gated router to update the weight of each expert model based on the current weight of each expert model and the effect information corresponding to each expert model.

[0017] In some embodiments, the attack behavior information includes: a first threat feature vector obtained based on the first feature data, a second threat feature vector obtained based on the second feature data, and a third threat feature vector obtained based on the third feature data.

[0018] In some embodiments, the n expert models include: a traffic analysis expert model for performing spatiotemporal feature analysis on the network traffic to detect network layer attack behaviors; a container security expert model for detecting network traffic attack behaviors on containers; a federated privacy protection expert model for detecting network traffic attack behaviors on privacy data and implementing cross-tenant knowledge sharing under privacy protection; and a decision expert model for generating adaptive defense strategies based on multi-source threat intelligence obtained from the network traffic analysis, wherein the multi-source threat intelligence includes a first threat feature vector, a second threat feature vector, and a third threat feature vector.

[0019] In some embodiments, the traffic analysis expert model is further used to generate defense strategies against network layer attacks based on single-source threat intelligence of network layer attack behavior; the container security expert model is further used to generate defense strategies against container attacks based on single-source threat intelligence of network traffic attacks on containers; and the federated privacy protection expert model is further used to generate defense strategies against privacy data attacks based on single-source threat intelligence of network traffic attacks on privacy data.

[0020] In some embodiments, the first feature data includes: the entropy value and IP dispersion obtained by the traffic protection layer from processing the network traffic; the second feature data includes: the system call frequency and vulnerability score obtained by the container security layer from processing the network traffic; and the third feature data includes: the gradient norm of the privacy data obtained by the privacy protection layer from processing the network traffic.

[0021] In some embodiments, the security defense method further includes: using the gated router to obtain network traffic anomaly degree based on the first feature data, vulnerability exploitation severity based on the second feature data, and data sensitivity based on the third feature data; calculating a threat assessment value based on the network traffic anomaly degree and its corresponding weight coefficient, the vulnerability exploitation severity and its corresponding weight coefficient, and the data sensitivity and its corresponding weight coefficient; determining the threat level of the attack behavior of the network traffic based on the threat assessment value; and determining the server resources to be used based on the threat level.

[0022] In some embodiments, the threat level includes a low threat level, a medium threat level, and a high threat level, wherein the high threat level is higher than the medium threat level, and the medium threat level is higher than the low threat level. Determining the threat level of the network traffic attack based on the threat assessment value includes: determining that the network traffic attack belongs to the low threat level when the threat assessment value is within the range corresponding to the low threat level; determining that the network traffic attack belongs to the medium threat level when the threat assessment value is within the range corresponding to the medium threat level; and determining that the network traffic attack belongs to the high threat level when the threat assessment value is within the range corresponding to the high threat level.

[0023] According to another aspect of this disclosure, a security defense system for an education and training platform is provided, comprising: a memory; and a processor coupled to the memory, the processor being configured to execute the security defense method for the education and training platform as described above based on instructions stored in the memory.

[0024] According to another aspect of this disclosure, a computer-readable storage medium is provided having computer instructions stored thereon that, when executed by a processor, implement the security defense method for an education and training platform as described above.

[0025] According to another aspect of this disclosure, a computer program product is provided, comprising a computer program or instructions that, when executed by a processor, implement the security defense method for an education and training platform as described above.

[0026] In the above system, different feature data are obtained through different functional layers. Based on the feature vectors constructed from these data, the weight of each of the n expert models is determined. The n expert models are then arranged in descending order of weight, and the top k expert models are selected to identify attack behavior information in network traffic. This allows the intelligent decision layer to generate a defense strategy based on the attack behavior information. After the defense strategy is executed, its effectiveness is evaluated to obtain the effect information corresponding to each of the top k expert models. The effect information is then fed back to the gated router. This facilitates the detection and identification of complex threat characteristics, thereby obtaining an adaptive defense strategy and improving the system's security.

[0027] Other features and advantages of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0028] The accompanying drawings, which form part of this specification, illustrate embodiments of this disclosure and, together with the specification, serve to explain the principles of this disclosure.

[0029] This disclosure will become clearer with reference to the accompanying drawings and the following detailed description, wherein:

[0030] Figure 1 This is a structural block diagram illustrating a security defense system for an education and training platform according to some embodiments of the present disclosure;

[0031] Figure 2 This is a flowchart illustrating a security defense method for an education and training platform according to some embodiments of the present disclosure;

[0032] Figure 3 This is a flowchart illustrating a security defense method for an education and training platform according to other embodiments of this disclosure;

[0033] Figure 4 This is a flowchart illustrating a method for determining the threat level of network traffic attacks according to some embodiments of this disclosure;

[0034] Figure 5 This is a structural block diagram illustrating a security defense system for an education and training platform according to other embodiments of this disclosure;

[0035] Figure 6 This is a structural block diagram illustrating a security defense system for an education and training platform according to other embodiments of the present disclosure. Detailed Implementation

[0036] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present disclosure.

[0037] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0038] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this disclosure or its application or use.

[0039] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0040] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0041] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0042] Security (core area): Involves core security technologies such as dynamic security defense systems, threat detection and response, and privacy protection.

[0043] Artificial Intelligence (Enabling Technologies): Employing AI algorithms such as Hybrid Expert Model (MoE) and Spatiotemporal Graph Neural Network (STGNN) to achieve intelligent threat analysis and adaptive decision-making.

[0044] Vertical industries (application implementation areas): Multi-tenant cloud platform scenarios specifically for the education and training industry.

[0045] Through in-depth research and analysis of existing technologies, the inventors of this disclosure have discovered that current technical solutions in the field of cloud-native security mainly suffer from three key technical limitations, which severely restrict the security protection effect of cloud-native environments.

[0046] First, regarding detection mechanisms, existing solutions generally employ static gating architectures or fixed rule engines. This rigid architecture design demonstrates significant inadequacy in adapting to rapidly evolving attack patterns. Specifically, static detection strategies cannot dynamically adjust according to changes in attack characteristics, leading to a significant decrease in the system's detection effectiveness against new threats and complex attacks. More seriously, this static nature also causes a substantial increase in false positive rates, severely impacting the efficiency of security operations and maintenance.

[0047] Secondly, in terms of system architecture design, existing technologies mostly adopt a serial processing structure and loosely coupled expert module design. This architecture has obvious performance bottlenecks: on the one hand, the serial processing flow causes threat response latency to exceed the real-time requirements of cloud-native environments; on the other hand, the lack of effective coordination mechanisms between loosely coupled expert modules makes the system less effective at detecting complex attack types that require multi-dimensional correlation analysis. These architectural defects directly affect the overall protection effectiveness.

[0048] Furthermore, in terms of resource utilization efficiency, existing solutions generally adopt a static resource allocation model. This model cannot flexibly respond to dynamic changes in attack intensity, resulting in two efficiency problems: during non-attack periods, a large number of security resources are idle; while during peak attack periods, there is often a shortage of resources. This rigidity in resource allocation not only increases operating costs but may also affect the effectiveness of protection during critical periods.

[0049] Through more in-depth technical analysis, the inventors of this disclosure have discovered that the core contradictions faced by existing technical solutions are mainly reflected in the following three aspects, and these contradictions are particularly prominent in the cloud-native environment.

[0050] The first contradiction is between static models and dynamic threats. Most existing security detection models are statically designed, with their parameters and structure essentially fixed after deployment. However, attack methods in cloud environments are rapidly evolving, and this mismatch makes it difficult for systems to respond to new threats in a timely and effective manner. Especially when facing zero-day attacks and multi-stage APT (Advanced Persistent Threat) attacks, the detection effectiveness of static models decreases significantly.

[0051] The second contradiction lies between privacy protection and collaborative defense. In a cloud-native multi-tenant environment, ensuring the data privacy of each tenant while simultaneously achieving effective collaborative defense places high technical demands. Existing solutions often compromise on one aspect: either severely sacrificing system real-time performance to guarantee data privacy, or neglecting privacy protection requirements in pursuit of collaborative efficiency. This lack of balance limits the improvement of overall protection capabilities.

[0052] The third contradiction lies between resource allocation and protection requirements. Cloud-native environments exhibit significant dynamic workloads, while existing security resource allocation strategies remain relatively fixed. This mismatch leads to inefficient resource utilization: resource waste during off-peak periods and insufficient protection during peak periods. More seriously, this contradiction also impacts system scalability and cost-effectiveness.

[0053] Based on the above analysis, the inventors of this disclosure have discovered that existing technical solutions have several key technical defects in the field of cloud-native security protection, mainly in the following aspects:

[0054] First, in terms of detection mechanisms, existing solutions generally employ static gated networks or fixed rule engines. This rigid architecture struggles to dynamically adapt to rapidly evolving attack patterns. When faced with new or complex attacks, the system's detection performance significantly degrades, making it difficult to identify complex threat characteristics.

[0055] Secondly, in terms of system architecture design, existing technologies mostly adopt serial processing structures and loosely coupled expert modules. This architecture has obvious performance bottlenecks, causing threat response delays to exceed the real-time requirements of cloud-native environments. At the same time, it is not effective in detecting attack types that require multi-dimensional correlation analysis.

[0056] Furthermore, in terms of resource utilization efficiency, the static resource allocation mode of the existing solution cannot flexibly cope with the dynamic changes in attack intensity, resulting in low resource utilization. It is neither able to effectively save resources during non-attack periods nor can it guarantee protection effectiveness during peak attack periods.

[0057] The inventors of this disclosure have discovered problems in the prior art described above, and therefore propose a new technical solution to address at least one of the problems.

[0058] For example, in view of the problem in the prior art that the system is not easy to identify complex threat characteristics when facing new or complex attacks, the embodiments of this disclosure provide a security defense system and / or method for education and training platforms to facilitate the detection and identification of complex threat characteristics.

[0059] The following describes in detail, with reference to the accompanying drawings, a security defense system for an education and training platform according to some embodiments of the present disclosure.

[0060] Figure 1 This is a structural block diagram illustrating a security defense system for an education and training platform according to some embodiments of the present disclosure.

[0061] like Figure 1 As shown, the security defense system includes a traffic protection layer 110, a container security layer 120, a privacy protection layer 130, an intelligent decision-making layer 140, and a gated router 150.

[0062] The traffic protection layer 110 is used to process network traffic to obtain first characteristic data. The first characteristic data is characteristic data related to the traffic data of the network traffic. For example, the first characteristic data includes: the entropy value and IP (Internet Protocol) dispersion obtained by the traffic protection layer from processing the network traffic. For example, the entropy value and IP dispersion can be obtained using known techniques.

[0063] Container security layer 120 is used to process network traffic to obtain second characteristic data. The second characteristic data is characteristic data related to container attack data in the network traffic. For example, the second characteristic data includes: system call frequency and vulnerability score obtained by the container security layer from processing the network traffic. For example, the system call frequency and vulnerability score can be obtained using known techniques.

[0064] The privacy protection layer 130 is used to process network traffic to obtain third feature data. The third feature data is feature data related to the privacy data of the network traffic. For example, the third feature data includes the gradient norm of the privacy data obtained by the privacy protection layer from processing the network traffic. For example, the gradient norm of the privacy data can be obtained using known techniques.

[0065] The gated router 150 is used to receive first feature data, second feature data and third feature data, construct feature vectors based on the first feature data, second feature data and third feature data, determine the weight of each of the n expert models according to the feature vectors, arrange the n expert models in descending order of weight, and select the first k expert models to identify attack behavior information in network traffic, where 1≤k≤n, and k and n are positive integers.

[0066] The intelligent decision layer 140 is used to generate defense strategies based on attack behavior information, evaluate the effectiveness of the defense strategies after they are executed to obtain the effect information corresponding to each of the top k expert models, and feed the effect information back to the gated router.

[0067] For example, the intelligent decision layer 140 can use known evaluation methods to evaluate the effect of the defense strategy after its execution to obtain effect information. This effect information may include, for example, accuracy. For instance, if expert models 1 and 2 are currently selected, and the accuracy of the defense strategy (as effect information) obtained after execution is 0.7, then 0.7 is used as the accuracy of expert models 1 and 2 respectively. The intelligent decision layer can then feed back the accuracy of expert model 1 (0.7) and the accuracy of expert model 2 (0.7) to the gated router.

[0068] In some embodiments, the attack behavior information includes: a first threat feature vector obtained based on first feature data, a second threat feature vector obtained based on second feature data, and a third threat feature vector obtained based on third feature data.

[0069] For example, a first feature value is obtained by weighting and summing the entropy value and the preset weight corresponding to the entropy value, as well as the IP dispersion and the preset weight corresponding to the IP dispersion. The first feature values ​​at different times are then combined to form a first threat feature vector.

[0070] For example, a second feature value is obtained by weighting and summing the system call frequency and the preset weight corresponding to the system call frequency, as well as the vulnerability score and the preset weight corresponding to the vulnerability score. The second feature values ​​at different times are then combined to form a second threat feature vector.

[0071] For example, the gradient norm of privacy data can be used as the third feature value, and the third feature values ​​at different times can be combined to form a third threat feature vector.

[0072] This provides a security defense system for an education and training platform according to some embodiments of the present disclosure. The security defense system includes: a traffic protection layer for processing network traffic to obtain first feature data; a container security layer for processing network traffic to obtain second feature data; a privacy protection layer for processing network traffic to obtain third feature data; a gated router for receiving the first, second, and third feature data, constructing feature vectors based on the first, second, and third feature data, determining the weight of each of n expert models according to the feature vectors, arranging the n expert models in descending order of weight, and selecting the top k expert models to identify attack behavior information in the network traffic, where 1 ≤ k ≤ n, and k and n are positive integers; and an intelligent decision layer for generating a defense strategy based on the attack behavior information, evaluating the effectiveness of the defense strategy after its execution to obtain effect information corresponding to each of the top k expert models, and feeding back the effect information to the gated router. In this embodiment, different feature data are obtained through different functional layers. Based on the feature vector constructed from these data, the weight of each of the n expert models is determined. The n expert models are then arranged in descending order of weight, and the top k expert models are selected to identify attack behavior information in network traffic. This allows the intelligent decision layer to generate a defense strategy based on the attack behavior information. After the defense strategy is executed, its effectiveness is evaluated to obtain the effect information corresponding to each of the top k expert models. The effect information is then fed back to the gated router. This facilitates the detection and identification of complex threat characteristics, thereby obtaining an adaptive defense strategy and improving the system's security.

[0073] In some embodiments, the gated router is further configured to update the weights of each expert model based on the current weights of each expert model and the effect information corresponding to each expert model.

[0074] For example, in the process of updating the weights of each expert model, the new weights are calculated. for

[0075]

[0076] in, Let be the weights of the i-th expert model at time t+1 (i.e., the updated weights). Let α be the weights of the i-th expert model at time t (i.e., the weights before the update), α be the pre-set learning rate, and Accuracy(E) be the weights of the i-th expert model at time t. i ) represents the accuracy (i.e., performance information) corresponding to the i-th expert model. Here, Accuracy(E) iThe initial value of ) is 0, and the effect information is updated after the defense strategy is executed, that is, the Accuracy(E) is updated. i The value of ).

[0077] The above method enables dynamic updates to the weights of each expert model. Addressing the challenge of existing static defense strategies adapting to rapid changes in cloud environments, the embodiments of this disclosure implement this dynamic response mechanism. This allows for adaptive optimization from threat detection to response. Compared to fixed-period update protection systems in related technologies, the embodiments of this disclosure significantly improve the real-time response capability against new threats and mutated attacks.

[0078] In some embodiments, the n expert models include: a traffic analysis expert model, a container security expert model, a federal privacy protection expert model, and a decision-making expert model.

[0079] Traffic analysis expert models are used to perform spatiotemporal feature analysis on network traffic to detect network layer attack behaviors.

[0080] Container security expert models are used to detect network traffic attacks on containers.

[0081] The Federal Privacy Protection Expert Model is used to detect network traffic attacks on privacy data and to implement cross-tenant knowledge sharing under privacy protection.

[0082] The decision expert model is used to generate adaptive defense strategies based on multi-source threat intelligence obtained from network traffic analysis. The multi-source threat intelligence includes a first threat feature vector, a second threat feature vector, and a third threat feature vector.

[0083] In other embodiments, traffic analysis expert models can also be used to generate defense strategies against network layer attacks based on single-source threat intelligence. Container security expert models can also be used to generate defense strategies against container attacks based on single-source threat intelligence based on network traffic attacks on containers. Federated privacy protection expert models can also be used to generate defense strategies against privacy data attacks based on single-source threat intelligence based on network traffic attacks on privacy data.

[0084] In other words, in addition to detecting corresponding attack behaviors, traffic analysis expert models, container security expert models, and federated privacy protection expert models can each provide corresponding defense strategies for single-source threat intelligence such as the attack behaviors they detect. Meanwhile, decision expert models can generate adaptive defense strategies for multi-source threat intelligence. This allows for the integration of defense strategies from various expert models to execute an appropriate defense strategy.

[0085] The above system is a dynamic security defense architecture based on MoE, which will be described in further detail below.

[0086] (1) System design objectives and constraints

[0087] Addressing the highly dynamic and multi-tenant characteristics of cloud-native training platforms, this system meets core constraints such as real-time response, privacy compliance, and resource efficiency. It ensures that the end-to-end latency from attack detection to response meets the requirements of critical business SLAs (Service Level Agreements), that tenant sensitive data is strictly processed locally in accordance with GDPR (General Data Protection Regulation) / CCPA (California Consumer Privacy Act), and that the defense module maintains low resource consumption to ensure critical business priorities. Based on this, four principles of dynamic defense are proposed: 1. Covering the entire attack chain—traffic-container-privacy-decision—through heterogeneous expert collaboration. 2. Adaptively selecting the optimal expert model based on threat characteristics. 3. Achieving cross-tenant knowledge sharing under privacy protection through federated learning and homomorphic encryption. 4. Establishing a resource elastic linkage mechanism that intelligently adjusts defense strength according to the attack scale, forming a closed-loop control system.

[0088] (2) Overall architecture design

[0089] The dynamic security defense system adopts a four-layer vertical protection and MoE horizontal collaborative architecture, such as Figure 1 As shown.

[0090] This security architecture employs a layered design. The traffic protection layer is responsible for real-time analysis of the inter-host traffic matrix (source / destination IP, packet size, protocol type), utilizing traffic analysis expert models (e.g., STGNN (Spatial-Temporal Graph Neural Network) expert models) to detect network layer attacks such as DDoS (Distributed Denial of Service) and port scanning, and can dynamically update the traffic graph topology to adapt to IP changes caused by container scaling. The container security layer focuses on container runtime security, fusing and analyzing multimodal data such as container standard output / error logs (text), system call sequences (time-series), and image vulnerability reports (structured), and employing container security expert models (e.g., multimodal CNN (Convolutional Neural Network) expert models) to identify escape attacks and abnormal processes. The privacy protection layer, while ensuring data isolation, leverages locally generated security event feature vectors (not raw data) from each tenant, utilizing federated privacy protection expert models (e.g., federated learning expert models) (combining differential privacy and knowledge distillation) to achieve privacy-preserving sharing of threat knowledge across tenants. Ultimately, the intelligent decision layer aggregates the threat feature vectors output from each layer, generates executable defense strategies through a fine-tuned Security-LLM (Security Large Language Model) decision engine, and evaluates their effectiveness, forming a closed-loop intelligent response.

[0091] (3) Gating network design

[0092] As the central hub of a dynamic defense system, the gating network performs adaptive routing from attack features to expert models. Through a three-tiered linkage of feature encoding, sparse routing, and dynamic optimization, it ensures low latency (e.g., activating only two experts) while continuously evolving expert selection accuracy as the system runs, effectively adapting to the dynamic threat landscape of the cloud environment. Its design comprises three core mechanisms. The first is input feature encoding. Heterogeneous feature vectors are extracted from the four-layer defense system (e.g., entropy and IP dispersion at the traffic layer, system call frequency and vulnerability scores at the container layer, gradient norm at the privacy layer, etc.), and standardized using the Min-Max algorithm to eliminate dimensional differences, forming a unified representation. The second is sparse routing decision-making. A Top-K sparse gating algorithm (e.g., K=2) is used to balance efficiency and accuracy, calculating the weight w of each expert. i =softmax(W g ·x), where w i W represents the weight of the i-th expert model, where i is a positive integer. gThe weights represent the parameters used to linearly transform the input features, x represents the input feature vector, and the softmax function is a commonly used normalization function. For example, it can transform a vector containing arbitrary real numbers into a vector with element values ​​between 0 and 11, and the sum of all elements is 11. It is often used in multi-class classification problems to represent the probability distribution of different classes. The top K expert models with the highest weights are selected to perform the task (e.g., activating the STGNN expert model and the LLM expert model during a DDoS attack). An expert capacity factor C is introduced. i Constrain instance load and avoid resource contention through weighted round-robin scheduling to ensure response latency remains within a specified time. Then, dynamic weight adjustment is implemented, using a feedback loop mechanism to iteratively optimize routing weights based on historical expert detection accuracy.

[0093]

[0094] During the cold start phase, uniform weight allocation is adopted, and the routing strategy gradually converges to the optimal routing strategy as attack samples accumulate.

[0095] (4) Expert Collaboration Mechanism

[0096] This system achieves deep collaboration among multiple experts through event-driven chain activation and cross-layer feature fusion, and dynamically constructs expert collaboration chains based on attack type features (Table 2-1):

[0097] Table 1 Expert Collaboration Paths in Typical Attack Scenarios

[0098]

[0099]

[0100] In Table 1 above, the STGNN expert model is used as a traffic analysis expert model, the multimodal CNN expert model as a container security expert model, the federated learning expert model as a federated privacy protection expert model, and the LLM decision expert model as a decision expert model.

[0101] DDoS flood attack: Activate the STGNN expert at the traffic layer to identify spatiotemporal patterns and link it with the LLM at the decision layer to generate traffic cleaning strategies.

[0102] Container privilege escalation attack: Anomaly detection by multimodal CNN experts at the container layer triggers federated learning experts at the privacy layer to analyze cross-tenant association risks.

[0103] Data theft attack: After privacy layer experts locate the leakage path, they report back to the container layer to revoke access permissions.

[0104] This mechanism avoids redundant expert calls and reduces average response latency by 32%.

[0105] Cross-layer feature enhancement fusion:

[0106] Vertical transmission mechanism: Lower-layer expert output serves as upper-layer input (e.g., container layer detection results trigger deep auditing of the privacy layer);

[0107] Attention-guided alignment: The decision-level LLM uses a multi-head attention module to dynamically weight the importance of features at each layer.

[0108]

[0109] Wherein, Attention(Q_decision, K_expert, V_expert) is the final attention output calculated by the decision layer LLM through a multi-head attention module, used to dynamically weight the importance of features of each expert layer; Q_decision is the query vector of the decision layer, representing the information or task objective currently being focused on by the decision layer; K_expert is the key vector of the expert layer, used to match with the query vector and measure the relevance of expert layer features to decision requirements; V_expert is the value vector of the expert layer, containing specific feature information of the expert layer, and is the object of attention weighting; ∑ represents summing the attention results of multiple heads; softmax represents transforming the similarity scores between the query and the key into a weight distribution, so that the sum of the weights is 1; QK T The dot product of the query vector and the key vector is used to calculate their similarity; d k Let QK be the dimension of the key vector. T Scaling (divided by) To prevent the softmax gradient from vanishing due to excessively large inner product values; V is the expert layer value vector after weighting.

[0110] Enable multi-dimensional correlation analysis of threat characteristics (such as correlating abnormal traffic with container vulnerabilities).

[0111] (5) Secure Data Flow Design

[0112] To ensure the efficient operation of the dynamic defense system, this system has built a standardized, low-latency, and scalable security data flow framework, covering the entire process of data acquisition, preprocessing, expert routing, and decision feedback.

[0113] Multi-source data acquisition and standardization utilize lightweight probes to achieve efficient collection of heterogeneous data. The network traffic layer captures traffic from the privacy protection / intelligent decision-making layer using the Envoy sidecar proxy, dynamically adjusting the sampling rate (e.g., a baseline of 10%, increasing to 30% during attacks), outputting a spatiotemporal sequence of five tuples (source / destination IP, port, protocol, packet size). The container runtime layer collects system call events (key calls such as open and execve) through eBPF (extended Berkeley Packet Filter) kernel probes, aggregating standard container outputs with the Fluentd log collector to form a behavioral time-series chain. In the privacy protection layer, each tenant deploys a feature extractor locally, converting raw logs into 32-dimensional feature vectors (such as API (Application Programming Interface) call frequency and sensitive file access counts), ensuring data remains within the domain.

[0114] A pipelined processing mechanism is established, with data flowing through four stages of processing. The first stage is dynamic preprocessing: traffic data is used to construct a spatiotemporal graph (nodes = Pod labels, edges = traffic entropy weights); container logs are vectorized using fine-tuned BERT (Bidirectional Encoder Representations from Transformers), and system call sequences are encoded into One-Hot vectors. The second stage is gated routing decision-making: preprocessed features are input into the gating network, which selects expert instances through Top-2 sparse routing (e.g., for DDoS attacks, STGNN and LLM experts are selected). The third stage is parallel expert execution: selected expert instances receive tasks via gRPC (a high-performance, open-source remote procedure call framework), with forced interruption upon timeout (e.g., a threshold of 150ms), and results are output in a Unified Threat Description Framework (UTD) format. The fourth stage is closed-loop feedback optimization: after the decision layer generates a strategy (e.g., container isolation), execution performance metrics (false positive / false negative rate, response latency) are fed back to the gating network, dynamically adjusting routing weights.

[0115] In some embodiments, the gated router is further configured to obtain network traffic anomaly level based on first feature data, vulnerability exploitation severity based on second feature data, and data sensitivity based on third feature data; calculate a threat assessment value based on network traffic anomaly level and its corresponding weight coefficient, vulnerability exploitation severity and its corresponding weight coefficient, and data sensitivity and its corresponding weight coefficient; determine the threat level of the network traffic attack behavior based on the threat assessment value; and determine the server resources required based on the threat level. Here, network traffic anomaly level, vulnerability exploitation severity, and data sensitivity are calculated using known techniques.

[0116] For example, the network traffic anomaly degree can be obtained by weighting and summing the entropy value and the preset weight corresponding to the entropy value, as well as the IP dispersion degree and the preset weight corresponding to the IP dispersion degree.

[0117] For example, the severity of vulnerability exploitation can be obtained by weighting and summing the system call frequency and the preset weight corresponding to the system call frequency, as well as the vulnerability score and the preset weight corresponding to the vulnerability score.

[0118] For example, the gradient norm of privacy data can be used as a measure of data sensitivity.

[0119] In related technologies, security solutions often suffer from system overload or insufficient protection due to unreasonable resource allocation. However, the embodiments of this disclosure establish a precise mapping relationship between defense strength and resource requirements, dynamically scheduling computing resources based on threat levels. This can reduce system resource overhead.

[0120] Furthermore, the sparse activation characteristic of the MoE architecture can be leveraged to invoke relevant expert models when necessary. This design significantly reduces system resource overhead while ensuring effective protection.

[0121] In some embodiments, the threat level includes a low threat level, a medium threat level, and a high threat level, wherein the high threat level is higher than the medium threat level, and the medium threat level is higher than the low threat level.

[0122] Gated routers are used to determine that network traffic attacks belong to the low threat level when the threat assessment value is within the range corresponding to the low threat level, to the medium threat level when the threat assessment value is within the range corresponding to the medium threat level, and to the high threat level when the threat assessment value is within the range corresponding to the high threat level.

[0123] (6) Flexible resource scheduling mechanism

[0124] To address the dynamic cybersecurity threats in cloud-native training platforms, embodiments of this disclosure propose a threat-aware elastic resource scheduling mechanism that achieves an optimal balance between defense effectiveness and computational cost by quantifying attack intensity and dynamically adjusting the allocation of defense resources.

[0125] The embodiments of this disclosure provide a threat-driven resource scheduling model. An Attack Intensity Index (TII) is constructed based on multi-dimensional attack characteristics, comprehensively considering factors such as network traffic anomalies, vulnerability exploitation severity, and data sensitivity. A quantitative threat assessment value is generated through weighted fusion. This index is divided into three levels: low threat level (TII < 0.4), medium threat level (0.4 ≤ TII < 0.7), and high threat level (TII ≥ 0.7), each triggering differentiated resource scheduling strategies.

[0126] Figure 4 This is a flowchart illustrating a method for determining the threat level of network traffic attacks according to some embodiments of this disclosure. Figure 4 The dynamic transition logic of elastic resource scheduling is formally described using a Finite State Machine (FSM), which includes three core states and triggering conditions. For example, ... Figure 4 As shown, the assessment value range for low threat level is threat assessment value < 0.4, the assessment value range for medium threat level is 0.4 ≤ threat assessment value < 0.7, and the assessment value range for high threat level is threat assessment value ≥ 0.7.

[0127] Table 2 State Definitions and Transition Rules

[0128]

[0129]

[0130] It should be noted that in Table 2 above, sudden traffic increases exceeding 50% of the baseline and the exploitation of critical vulnerabilities are considered sudden events. Additionally, the number of expert instances represents the server resources (i.e., computing resources) required to run a certain type of expert model. For example, "expert instances × 3" indicates the server resources required to run 3 expert models of a certain type.

[0131] The dynamic resource adaptation mechanism is described below.

[0132] Elastic scaling of expert instances: In low-threat scenarios, basic defense resources are maintained and only a single expert instance is run; in medium- and high-threat scenarios, the number of expert instance replicas is dynamically expanded (2-3 times) and prioritized for scheduling to high-computing-power nodes.

[0133] Load balancing and degradation protection: Load balancing among expert instances is achieved through a distributed task queue; when resource consumption exceeds the safety threshold, it automatically switches to a lightweight model to ensure business continuity.

[0134] Real-time performance guarantee: Preloading and fast instance startup technologies are used to keep resource adjustment latency within seconds (≤10 seconds).

[0135] The previous section described four expert models; the following section will provide a detailed description of each of the four expert models.

[0136] (1) Traffic Analysis Expert Model

[0137] The traffic analysis expert model is an expert model built on a spatiotemporal graph neural network. In this network, Pod tags (components in the Kubernetes (an open-source container orchestration platform) architecture) are used as graph nodes, and directed edges are constructed based on 5-tuple traffic data. This 5-tuple traffic data includes: source IP address, destination IP address, port, network protocol, and the size of the traffic data packet.

[0138] Key technologies for expert model implementation

[0139] For example, the traffic analysis expert model is the STGNN traffic analysis expert model.

[0140] To address the spatiotemporal dynamics of network traffic in cloud-native environments, embodiments of this disclosure propose a traffic analysis expert model based on a spatiotemporal graph neural network (STGNN). By integrating topology and temporal dependency features, it achieves high-precision DDoS attack and abnormal traffic detection.

[0141] (1.1) Dynamic flow graph modeling

[0142] To adapt to the dynamic IP allocation characteristics of cloud environments, a tag-driven graph construction method is adopted.

[0143] Node definition: Kubernetes Pod tags (such as app=training-web) are used as graph nodes to replace traditional IP addresses and solve the node drift problem caused by container scaling.

[0144] Edge weight calculation: Directed edges are constructed based on quintuple traffic data. The weights are combined with the number of packets, protocol entropy value and time series fluctuations to reflect the communication strength and anomaly probability between hosts.

[0145] (1.2) Joint extraction of spatiotemporal features

[0146] Design a GCN-GRU hybrid architecture (a hybrid model of Graph Convolutional Network (GCN) and Gated Recurrent Unit (GRU)) to achieve multi-level feature learning:

[0147] Spatial dependency modeling: By aggregating the traffic features of neighboring nodes through Graph Convolutional Networks (GCNs), the topological patterns of attack propagation can be captured.

[0148]

[0149] Formula (3) is a form of the propagation formula for Graph Convolutional Layers (GCN) commonly used in Graph Neural Networks (GNNs). This formula describes the update process of node features in a graph, and the specific parameters have the following meanings:

[0150] H (l+1) : Represents the node feature matrix of the (l+1)th layer, that is, the node features output after the current layer's graph convolution operation. Its dimension is usually N×D′, where N is the number of nodes in the graph and D′ is the dimension of the current layer's output features.

[0151] σ(·): Activation function, used to perform nonlinear mapping on features after linear transformation. Common activation functions include ReLU and Sigmoid.

[0152] The normalization matrix, where D is a transformation of the degree matrix of the graph (usually the degree matrix after adding self-loops), and the superscript -1 / 2 indicates that the matrix is ​​square rooted and then inverted. It is used to normalize the node features and avoid the eigenvalue scaling problem caused by the difference in node degree.

[0153] The normalized adjacency matrix is ​​usually obtained by adding an identity matrix (to include information about the nodes themselves) to the original adjacency matrix A. Where I N It is an N×N identity matrix.

[0154] H (l) : Represents the node feature matrix of the l-th layer, that is, the node features input to the current layer, with a dimension of N×D, where D is the dimension of the input features.

[0155] W (l) : represents the weight matrix of the l-th layer, used to perform linear transformation on the input features. Its dimension is D×D′, where D is the input feature dimension and D′ is the output feature dimension. This matrix is ​​obtained through training.

[0156] Timing dynamic analysis: Gated cyclic units (GRUs) are used to process continuous time slices to identify the timing characteristics of DDoS attacks, such as their impulsiveness and periodicity.

[0157] Spatiotemporal attention fusion: Introducing a cross-layer attention mechanism to dynamically weight threat indicators of key nodes and time steps (such as the mutation rate of SYN (Synchronize) requests).

[0158] (1.3) Dynamic Detection and Optimization

[0159] Adaptive threshold mechanism: The anomaly detection threshold is dynamically adjusted based on the historical traffic baseline to reduce the false alarm rate (measured FPR (False Positive Rate) < 1.2%).

[0160] Lightweight deployment: The model is compressed to 30% of its original size through knowledge distillation, and the inference latency is controlled within 50ms.

[0161] (2) Container safety expert model

[0162] The container security expert model is an expert model built on a multimodal deep learning framework based on cross-modal attention mechanism, targeting multi-source heterogeneous security data in the container runtime environment.

[0163] For example, the container security expert model is a multimodal container security expert.

[0164] To address the multi-source, heterogeneous security data in container runtime environments, embodiments of this disclosure construct a multimodal deep learning framework based on a cross-modal attention mechanism, enabling accurate detection of advanced threats such as container escape and privilege escalation. This framework achieves end-to-end analysis through a three-level processing flow.

[0165] Unified representation of heterogeneous data

[0166] A domain-adaptive encoder is used to align the feature spaces of text-based logs (BERT fine-tuning), temporal system calls (1D-CNN+BiLSTM, a hybrid deep learning model combining one-dimensional convolutional neural networks and bidirectional long short-term memory networks) and structured mirrored metadata (graph neural networks) to address the semantic gap caused by modality differences.

[0167] Gated cross-modal attention fusion

[0168] A dynamic weight allocation mechanism is designed to automatically adjust the contribution of each modality through learnable gating units, significantly improving the detection rate of covert attacks such as container weight escalation. Gated Cross-Attention (GCA) is introduced.

[0169] Intermodal attention:

[0170]

[0171] Formula (4) is the mathematical expression for the intermodal attention mechanism, based on the fundamental principles of the attention mechanism. In the attention mechanism, the similarity is usually calculated between the query vector and the key vector, and the weights are obtained through softmax before the value vectors are summed with weights.

[0172] In formula (4), Attention(...) represents the result of intermodal attention calculation. Q log It is a query vector, representing a feature vector from the log modality, used to query relevant information in the syscall modality. K syscall This is the key vector, derived from the system call modality, used to calculate similarity with the query vector to determine attention weights. V syscall It is a value vector, also derived from the system call modality, and is the feature vector ultimately used for weighted summation. d in k It is the dimension of the key vector, for Q log and The product of the inner product is scaled to prevent the gradient vanishing in the softmax function due to an excessively large inner product, thus making the weight distribution more reasonable. The softmax function converts the scaled inner product into a probability distribution, i.e., attention weights. These weights reflect the importance of different parts of the system call modality to the log modality query. Finally, the weights are scaled using the value vector V. syscall Multiplying these together yields an attention output that incorporates important information from both modalities.

[0173] Gating mechanism filters valid information:

[0174] g=σ(W g ·[Attn log-syscall Attn syscall-log (5)

[0175] Where g represents the weight coefficient output by the gating mechanism, used to control the fusion ratio of log features and system call features; σ is the activation function, usually the sigmoid function, which maps the input to between 0 and 1; W g It is the weight matrix of the gating mechanism, used to perform a linear transformation on the input attention features; Attnlog-syscall This represents the attention characteristics from logs to system calls, reflecting the degree of attention log information pays to system calls; Attn syscall-log This represents the attention feature from system calls to logs, reflecting the degree of attention the system call information pays to the logs; [·; ·] indicates the concatenation operation of two attention features.

[0176] Weighted output fusion features:

[0177] F fusion =g·F log +(1-g)·F syscall (6)

[0178] Among them, F fusion It is the final feature after fusion; g is the weight coefficient calculated by formula (5); F log Represents the features extracted from log data; 1-g is the complementary weight coefficient corresponding to g; F syscall This indicates the characteristics of data extracted from system calls.

[0179] Lightweight inference optimization: A hierarchical early termination strategy was developed, which terminates the calculation in advance when the confidence of the intermediate layer exceeds the threshold, achieving a 42% reduction in average inference latency while maintaining a detection accuracy of 98.3%.

[0180] (3) Federal Privacy Protection Expert Model

[0181] The Federal Privacy Protection Expert Model is an expert model that performs local differential training and global knowledge distillation on network data.

[0182] Privacy-Performance Collaborative Optimization Framework

[0183] To address the multi-tenant data isolation requirements of cloud-native training platforms, this disclosure proposes a two-stage federated learning enhancement architecture. This architecture achieves efficient sharing and collaborative defense of cross-tenant security knowledge while strictly meeting GDPR / CCPA compliance requirements. The expert system overcomes the challenge of balancing privacy protection and detection efficiency through three core technologies.

[0184] Phase 1: Local Differential Training

[0185] Each tenant injects Laplace noise (∈=0.5) locally to train the threat detection model:

[0186] M local =Train(D private Lap(0,β). (7)

[0187] Formula (7) addresses the scenario of combining differential privacy and knowledge distillation in federated learning. Where M... localThis represents the model obtained through local training; Train() is the model training function; D private Represents a local private dataset; Lap(0,β) is a Laplace distribution used to add noise to local data or model parameters to achieve differential privacy protection. Here, the parameter β is the scale parameter of the Laplace distribution. The smaller the value of β, the greater the noise and the higher the degree of privacy protection. Conversely, the degree of privacy protection decreases, but the model performance may be better.

[0188] Phase Two: Global Knowledge Distillation

[0189] The server aggregates the local model's output logits (non-parametric) and distills them into a lightweight global model:

[0190]

[0191] In the global knowledge distillation stage, logits refer to the model's raw output (output before undergoing activation functions such as softmax); logits global It is the global logits obtained after server aggregation; This represents the logits output by the i-th local model; The loss function is called knowledge distillation; KL(·||·) is the KL divergence (Kullback-Leibler divergence), which measures the difference between two probability distributions (obtained here by performing a softmax transformation on logits). By minimizing this loss, the output distribution of the local model is aligned with the output distribution of the global model, thereby achieving knowledge transfer and aggregation.

[0192] Differentiated privacy protection mechanism

[0193] Adaptive Laplacian noise (ε = 0.3-0.7) is introduced during the local training phase. Through gradient clipping and dynamic adjustment of noise variance, the privacy budget consumption of the model is reduced by 62%.

[0194] Design a secure aggregation protocol based on homomorphic encryption to ensure that the intermediate parameter transmission process meets the IND-CPA (Indistinguishability under Chosen-Plaintext Attack) security standard.

[0195] Knowledge distillation-enhanced global modeling: Adopting a teacher-student paradigm, the output logits of each tenant's local model are used as soft labels, and knowledge transfer is achieved by minimizing KL divergence; An importance-weighted distillation strategy is proposed, which assigns greater weight to high-confidence samples, and still maintains a global model accuracy of 91.2% under Non-IID data distribution.

[0196] Dynamic participation control: Based on the Shapley value, the contribution of each tenant to the global model is quantified, and incentive-compatible participation scheduling is implemented; a fault-tolerant asynchronous aggregation algorithm is developed, which allows the model to be updated even when 30% of the nodes are offline.

[0197] (4) Decision expert model

[0198] The decision expert model is an expert system based on a large language model. It generates defense strategies based on multi-source threat intelligence through domain-adaptive pre-training and executability constraint decoding techniques.

[0199] For example, the decision expert model is a large-scale intelligent decision expert model.

[0200] The embodiments of this disclosure design an intelligent decision-making expert system based on Large Language Model (LLM), which achieves end-to-end generation from multi-source threat intelligence to defense strategies through domain-adaptive pre-training and executability constraint decoding techniques. This expert system comprises three core modules.

[0201] Enhanced pre-training with security domain knowledge: A cross-modal security corpus (215GB in size) was constructed, integrating structured knowledge such as ATT&CK (Adversarial Tactics, Techniques, and Common Knowledge) tactical manuals, CVE (Common Vulnerabilities and Exposures) vulnerability descriptions, and cloud security incident reports; a progressive domain adaptation method was adopted, and a two-stage fine-tuning (domain adaptation + task specificity) was performed on a general base model (such as LLaMA-2 (LargeLanguage ModelMeta AI 2)) to improve the accuracy of security-related intent recognition to 93.7%.

[0202] Constrained policy generation mechanism: Design decoding constraints based on finite state automata to ensure that the output policy conforms to the predefined action set (isolation / rollback / alarm, etc.); develop a policy feasibility verification module to generate schemes by simulating the execution environment and eliminating 87.6% of the illusion output.

[0203] Real-time inference acceleration system: adopts LoRA low-rank adaptation technology to compress the 175B parameter model to an 8-bit quantized version; realizes dynamic batch processing and continuous inference optimization, and controls the single policy generation latency within 210ms (meeting SLA requirements).

[0204] In the above embodiments, in order to meet the requirements of accuracy and executability in generating network security defense strategies, an intelligent decision expert system based on Large Language Model (LLM) is proposed. Through domain adaptive pre-training and executability constraint decoding technology, the system can achieve end-to-end generation from multi-source heterogeneous threat intelligence to defense strategies.

[0205] (4.1) Enhanced pre-training in security domain knowledge

[0206] To overcome the semantic understanding bottleneck of general large models in the field of cybersecurity, embodiments of this disclosure construct a cross-modal security corpus D. sec The corpus integrates four key data sources: the MITRE Adversarial Tactics, Techniques and Common Knowledge (ATT&CK), common vulnerability disclosure (CVE) description texts from the China National Vulnerability Database, cloud platform security event logs, and attack and defense exercise scripts. Based on the LLaMA-3-8B base model, a progressive two-stage fine-tuning strategy is adopted. The first stage involves domain-adaptive training, using a masked language modeling task to enable the model to deeply understand security entity relationships and attack pattern characteristics. In D... sec The above training uses Masked Language Modeling (MLM) as the objective to learn embedded representations of secure entities and attack patterns:

[0207]

[0208] Formula (9) is the loss function expression for the Masked Language Model (MLM). Wherein, represents the loss function of the masked language model, used to measure the model's performance in predicting masked tokens; M is the set of random mask positions, i.e., the set of masked tokens, which are also the position indices in the input sequence that are artificially masked during training; x i The real token at the masked position i; x M This represents contextual information other than the masked position i, i.e., the information the model uses to predict x. i The known sequence content; θ is the model parameters, including the model weights, biases, etc., which are continuously optimized through the training process to minimize the loss function; logP(x i |x M ;θ) is the prediction of the true token given the context and model parameters θ. xi The overall loss is obtained by taking the logarithm of the probability, taking the negative sign, and summing the results. The goal is to minimize this loss to make the model's predictions more accurate.

[0209] The second phase involves task-specific fine-tuning, leveraging structured instruction templates to enhance strategy generation capabilities. The instruction template is as follows:

[0210]

[0211] (4.2) Constraint Generation Mechanism

[0212] To address the problem of infeasible policy output caused by the inherent hallucination problem of large models, a two-layer constraint mechanism is designed. The structural constraint layer mechanism constructs a decoding controller based on a finite state automaton, strictly restricting that the output actions must belong to a predefined operation set (including five core operations: isolation, rollback, alarm, permission restriction, and traffic traction). During the text generation process, the state transition path is monitored in real time, and hard masking is applied to candidate tokens (tokens / lemmas) that exceed the legal action space. The conditional probability distribution is as follows:

[0213]

[0214] In the conditional probability distribution P(π|v I ), π represents the policy, which is a sequence of actions; v I represents the input context information, providing the background for model decision-making. P(a t |v I , a<t) is the LLM policy generation function, where a t refers to the action selected at time t, and a<t represents all actions selected before time t. This function reflects the probability of generating the current action based on historical actions and context.

[0215] Among them, the LLM policy generation function is:

[0216] P(a t |v I , a<t). (12)

[0217] The finite state automaton constraint is:

[0218]

[0219] Finite State Automaton Constraint In it, q t is the state at time t, and Alegal(q) is the set of legal actions in state q. This indicator function is used to ensure that the actions conform to the state transition rules.

[0220] The action risk assessment function is:

[0221] Risk(a) = |W r ·Embed(a)|2. (14)

[0222] In the action risk assessment function Risk(a) = |W rIn Embed(a)|2, Embed(a) is the embedding vector of action a, W r The weight matrix is ​​used for risk assessment. The risk value of an action is obtained by calculating the vector magnitude.

[0223] The executability verification layer mechanism deploys a lightweight simulated execution sandbox environment to dynamically pre-evaluate the initially generated strategy draft. It quantifies the success rate of strategy execution in the simulated environment, for example, setting a feasibility threshold of 0.85. When the evaluation score falls below the threshold, a strategy regeneration process is automatically triggered. This mechanism can eliminate 80% of invalid operation instructions. The pre-evaluation formula is:

[0224] Score feas =f sim (π;ε sim )∈[0,1]. (15)

[0225] Among them, Score feas This represents the feasibility score, with values ​​ranging from [0, 1], used to measure the degree of feasibility of an object or plan; f sim The similarity function is used to calculate the similarity between related objects; π typically represents the strategy, approach, or path used to evaluate the object; ε sim A threshold or error parameter representing similarity, used to control the standard or range of similarity calculation. ε sim To simulate the execution environment.

[0226] (4.3) End-to-end policy generation process

[0227] The workflow follows a four-stage closed-loop logic. First, the input threat intelligence (including vulnerability features, log alerts, and network traffic data) undergoes feature alignment and vectorization encoding. Second, an initial policy draft is generated from a large model. Subsequently, the draft undergoes structural compliance correction through a finite state automaton constraint layer. Finally, the final policy is output after feasibility verification in a simulated environment. Let the input threat intelligence be a triple: I = CVE ,I log ,I net Log refers to log information in threat intelligence. log CVE is a vector obtained after extracting or processing certain log features; it represents information in threat intelligence. CVE Encoding is performed to convert CVE-related text information into corresponding vector representations, and network information I in threat intelligence is then processed. net Process it.

[0228] The feature encoding function is:

[0229] v I =ReLU(W c [BERT(I​CVE Log(I) log ); CNN(I net )]+b c ). 16)

[0230] v I The threat intelligence feature vector is the final feature representation output by the formula.

[0231] ReLU: An activation function that performs nonlinear processing on the result of a linear transformation, setting negative values ​​to 0 and retaining positive values.

[0232] W c : The fusion weight matrix has dimension R 256×d 256 is the output feature dimension, and d is the input concatenation vector dimension, used for linear feature fusion. This is the fusion weight matrix.

[0233] b c : Bias term, used to adjust the result of linear transformation.

[0234] BERT(I CVE ): Using the BERT model to analyze CVE information CVE The encoded vector.

[0235] Log(I log ): For log information I log The vector after feature extraction or encoding.

[0236] CNN(I net ): Using CNN to analyze network information CNN(I net The vector obtained by processing.

[0237] [·;·;·]: Vector concatenation operation, which concatenates the output vectors of BERT, Log, and CNN into a joint vector in sequence.

[0238] The strategy generation process is as follows:

[0239]

[0240] Among them, a t ∈A represents the action at step t (A is a predefined set of actions), and θ represents the model parameters after LoRA fine-tuning. Formula (17) presents an expression for policy probability in probabilistic modeling scenarios such as sequential decision-making or reinforcement learning.

[0241] P(π|v I ): represents the situation in a given observation sequence v I Under the given conditions, the probability of strategy π. Simply put, it's the probability that strategy π holds when observation information v is obtained.I The probability of adopting policy π afterwards.

[0242] ∏ T t＝1 : This is the symbol for continuous multiplication, which means that from time step t = 1 to time step t = T, the following terms are continuously multiplied. Its function is to combine the probability values of each time step to obtain the joint probability of the entire sequence.

[0243] P(a t |v I , a<t; θ): It represents the probability of selecting action a at the t-th time step under specific conditions. Among them, a t is the specific action taken at the t-th time step; v t is the observation sequence, which is all the observation information based on when making decisions; a<t represents the sequence of actions that have been executed before the t-th time step (that is, from time step 1 to t - 1); θ is the parameter of the model, and these parameters will affect the shape of the probability distribution and can be adjusted through learning, so that the model can better fit the actual data.

[0244] The following introduces the expert collaboration interface specification.

[0245] To achieve the efficient collaboration of heterogeneous security expert models, the embodiments of this disclosure have formulated a unified threat description framework (Unified Threat Description, UTD) and a lightweight communication protocol to build a standardized interaction system between cross-layer defense components. This specification contains three core elements. Semantically unified threat representation: Define a structured threat description template, including 7 mandatory fields (such as threat type, confidence level, timestamp, etc.) and 5 extended fields (such as attack phase marker, associated metrics, etc.). Serialization encoding is implemented using ProtocolBuffers, and the volume of a single message is compressed to an average of 2.3KB (68% less than JSON). Performance-optimized communication mechanism: High-concurrency data transmission is achieved based on gRPC-streaming, supporting the processing of more than 1500 threat events per second; Design a dual-channel communication mode: control channel (meta-instruction transmission) and data channel (feature vector stream), with the delay jitter controlled within ±15ms.

[0246] The dynamic load balancing strategy monitors the CPU / memory utilization rate of each expert instance in real time (sampling interval 500ms); An improved Consistent Hashing algorithm is used for request routing, achieving a load deviation of <8% (measured data).

[0247] The following introduces the dynamic security defense mechanism and system practice. ​​

[0248] (1) Closed-loop optimization mechanism of defense strategy

[0249] The embodiments of this disclosure construct a dynamic defense strategy optimization framework based on reinforcement learning, which continuously improves protection effectiveness through a three-layer progressive mechanism:

[0250] Multi-dimensional evaluation index system: An evaluation matrix with 12 quantitative indicators is designed, covering three dimensions: detection effectiveness (accuracy, recall), resource efficiency (CPU / memory usage), and timeliness (end-to-end latency). A dynamic weighting algorithm based on entropy weighting is introduced to automatically adjust the index weights according to the attack situation (e.g., recall is emphasized in high-threat scenarios).

[0251] Policy generation-execution-feedback closed loop: An Actor-Critic architecture is used for policy optimization, where the Actor network generates defensive actions (weight adjustment, instance scaling, etc.); the Critic network evaluates the state-action value function. A priority experience replay mechanism is designed, improving the sample reuse rate for key attack scenarios by 40%.

[0252] Online incremental learning: Develop a hot update interface for models to support parameter updates without interrupting services; mitigate catastrophic forgetting and retain historical knowledge through the Elastic Weight Consolidation (EWC) algorithm.

[0253] (2) System Implementation and Deployment Architecture

[0254] This system is implemented using cloud-native technology stack, and its main components include: microservice-based expert modules, event-driven architecture, and security enhancement design.

[0255] Microservice-based expert module: Each expert model is encapsulated as an independent Operator, and resource configuration is defined through Kubernetes CRD (Custom Resource Definition); it achieves fine-grained resource isolation (cgroup v2) and elastic scaling (Vertical Pod Autoscaler (VPA)).

[0256] Event-driven architecture: The message bus is built based on Knative Eventing, supporting an event throughput of 20,000+ EPS; a seven-layer priority queue is designed to ensure that the processing latency of critical threat events is <100ms.

[0257] Security Enhancement Design: Implements a zero-trust architecture, enforces mTLS authentication for all inter-component communication; integrates a Hardware Security Module (HSM) to protect model parameters and communication keys.

[0258] (3) Key technology innovation

[0259] First, dynamic defense strategy optimization. A defense payoff function based on game theory is proposed to quantify the cost-benefit of different defense actions; experiments show that the average defense efficiency is significantly improved compared to static strategies.

[0260] Second, resource-efficiency collaborative management. Develop joint optimization algorithms to significantly reduce resource consumption under SLA constraints; support Pareto optimal trade-offs between defense strength and computational cost.

[0261] Third, interpretability is enhanced. A knowledge graph of defense decisions is constructed to visualize the strategy generation logic; local interpretability is provided through the LIME (Local Interpretable Model-agnostic Explanations) method, improving the satisfaction of operations and maintenance personnel.

[0262] Figure 2 This is a flowchart illustrating a security defense method for an education and training platform according to some embodiments of the present disclosure. Figure 2 As shown, the security defense method includes steps S202 to S210.

[0263] In step S202, the network traffic is processed using a traffic protection layer to obtain first feature data. The first feature data is feature data related to the traffic data of the network traffic.

[0264] For example, the first feature data includes: the entropy value and IP dispersion obtained by the traffic protection layer from processing network traffic.

[0265] In step S204, the network traffic is processed using the container security layer to obtain second feature data. The second feature data is feature data related to container attack data in the network traffic.

[0266] For example, the second feature data includes: the system call frequency and vulnerability score obtained by the container security layer from processing network traffic.

[0267] In step S206, the network traffic is processed using a privacy protection layer to obtain third feature data. The third feature data is feature data related to the privacy data of the network traffic.

[0268] For example, the third feature data includes the gradient norm of the privacy data obtained by the privacy protection layer from processing network traffic.

[0269] In step S208, the first feature data, the second feature data, and the third feature data are received using a gated router. A feature vector is constructed based on the first feature data, the second feature data, and the third feature data. The weight of each of the n expert models is determined according to the feature vector. The n expert models are arranged in descending order of weight. The first k expert models are selected to identify attack behavior information in network traffic, where 1≤k≤n and k and n are positive integers.

[0270] For example, attack behavior information includes: a first threat feature vector obtained based on first feature data, a second threat feature vector obtained based on second feature data, and a third threat feature vector obtained based on third feature data.

[0271] In step S210, the intelligent decision layer generates a defense strategy based on the attack behavior information. After the defense strategy is executed, the effect of the defense strategy is evaluated to obtain the effect information corresponding to each of the top k expert models, and the effect information is fed back to the gated router.

[0272] This provides a security defense method for an education and training platform according to some embodiments of the present disclosure. In this method, different feature data are obtained through different functional layers. Based on these data, a feature vector is constructed to determine the weight of each of the n expert models. The n expert models are then arranged in descending order of weight, and the top k expert models are selected to identify attack behavior information in network traffic. This allows the intelligent decision layer to generate a defense strategy based on the attack behavior information. After the defense strategy is executed, its effectiveness is evaluated to obtain the effect information corresponding to each of the top k expert models. This effect information is then fed back to the gated router. This facilitates the detection and identification of complex threat characteristics, thereby obtaining an adaptive defense strategy and improving system security.

[0273] In some embodiments, the security defense method further includes: using a gated router to update the weights of each expert model based on the current weights of each expert model and the effect information corresponding to each expert model.

[0274] In some embodiments, the n expert models include: a traffic analysis expert model for performing spatiotemporal feature analysis on network traffic to detect network layer attack behavior; a container security expert model for detecting network traffic attack behavior on containers; a federated privacy protection expert model for detecting network traffic attack behavior on privacy data and implementing cross-tenant knowledge sharing under privacy protection; and a decision expert model for generating adaptive defense strategies based on multi-source threat intelligence obtained from network traffic analysis, wherein the multi-source threat intelligence includes a first threat feature vector, a second threat feature vector, and a third threat feature vector.

[0275] In some embodiments, the traffic analysis expert model is also used to generate defense strategies against network layer attacks based on single-source threat intelligence of network layer attack behavior; the container security expert model is also used to generate defense strategies against container attacks based on single-source threat intelligence of network traffic attacks on containers; and the federal privacy protection expert model is also used to generate defense strategies against privacy data attacks based on single-source threat intelligence of network traffic attacks on privacy data.

[0276] In some embodiments, the security defense method further includes: using a gated router to obtain network traffic anomaly based on first feature data, vulnerability exploitation severity based on second feature data, and data sensitivity based on third feature data; calculating a threat assessment value based on network traffic anomaly and its corresponding weight coefficient, vulnerability exploitation severity and its corresponding weight coefficient, and data sensitivity and its corresponding weight coefficient; determining the threat level of the attack behavior of the network traffic based on the threat assessment value; and determining the server resources to be used based on the threat level.

[0277] In some embodiments, the threat level includes a low threat level, a medium threat level, and a high threat level, wherein the high threat level is higher than the medium threat level, and the medium threat level is higher than the low threat level.

[0278] In some embodiments, determining the threat level of an attack on network traffic based on a threat assessment value includes: determining that the attack on network traffic belongs to a low threat level when the threat assessment value is within the range corresponding to a low threat level; determining that the attack on network traffic belongs to a medium threat level when the threat assessment value is within the range corresponding to a medium threat level; and determining that the attack on network traffic belongs to a high threat level when the threat assessment value is within the range corresponding to a high threat level.

[0279] Figure 3 This is a flowchart illustrating a security defense method for an education and training platform according to other embodiments of the present disclosure.

[0280] like Figure 3 As shown, firstly, feature encoding is obtained. Heterogeneous feature vectors are extracted from the four-layer defense system (e.g., entropy and IP dispersion of the traffic layer, system call frequency and vulnerability score of the container layer, gradient norm of the privacy layer, etc.), and the dimensional differences are eliminated by Min-Max (minimum algorithm) to form a unified representation.

[0281] Next, sparse routing decision-making is performed. A Top-K sparse gating algorithm is used to balance efficiency and accuracy, and the weights w of each expert are calculated.i =softmax(W g ·x), select the top K expert models with the highest weights to perform the task (e.g., a DDoS attack activates the STGNN expert model and the LLM expert model). For example, the STGNN expert model performs traffic analysis, the multimodal CNN expert model performs container detection, and the federated learning model performs privacy analysis.

[0282] Next, the threat features obtained from the analysis of the top K expert models are aggregated and fed into the LLM decision engine (i.e., the LLM decision expert model). The LLM decision engine outputs a defense strategy and provides feedback signals on the execution of the defense strategy to the expert weight matrix. This achieves dynamic weight adjustment.

[0283] Thus, a security defense method for an education and training platform according to other embodiments of this disclosure is provided, which facilitates the detection and identification of complex threat signatures.

[0284] In the systems and methods of this disclosure, regarding the detection mechanism, a dynamically adaptive detection technique can be developed. This technique can adjust the detection strategy in real time according to changes in attack characteristics, improving the adaptability to new threats. Regarding the system architecture, a more efficient parallel processing framework is designed. This framework can support deep collaboration among multiple expert models while ensuring the system's real-time response capability. Regarding resource management, an intelligent elastic scheduling algorithm is developed. This algorithm can dynamically adjust resource allocation based on real-time threat assessment, achieving the optimal balance between protective effectiveness and resource consumption.

[0285] In the systems and methods of this disclosure, a dynamic adaptive mechanism is used to achieve real-time optimization of attack situation awareness and defense strategies; a parallel architecture and deep feature fusion are employed to improve system processing efficiency; and an intelligent scheduling algorithm is used to achieve elastic resource allocation. Furthermore, a privacy protection framework can be used to achieve secure collaboration while meeting privacy requirements.

[0286] Some embodiments of this disclosure provide a dynamic security defense system architecture based on hybrid expert models (MoE). This architecture applies MoE technology to cloud security by dynamically scheduling four expert models through a gated network: the STGNN expert model handles spatiotemporal feature analysis of network traffic; the multimodal CNN expert model processes container behavior detection; the federated learning expert model enables cross-tenant knowledge sharing under privacy protection; and the LLM decision expert model generates adaptive defense strategies. This architecture overcomes the limitations of traditional single-model defense schemes, achieving collaborative detection and handling of multi-dimensional threats.

[0287] Some embodiments of this disclosure provide a four-layer dynamic defense system encompassing "traffic-container-privacy-decision". This system organically integrates previously isolated defense layers into a collaborative defense chain by establishing a cross-layer threat intelligence sharing mechanism. Specifically, it includes: real-time attack detection at the network traffic layer, abnormal behavior analysis at the container layer, data compliance protection at the privacy layer, and dynamic policy generation at the decision layer. Each layer interacts with the others through standardized interfaces for data exchange and coordinated responses.

[0288] Some embodiments of this disclosure can also provide a dynamic linkage mechanism for defense resources. This mechanism, based on real-time threat assessment results, achieves precise matching between defense strength and resource allocation through an optimized Kubernetes resource scheduling algorithm. It transforms security protection requirements into quantifiable resource scheduling parameters, achieving an optimal balance between defense effectiveness and resource utilization efficiency while ensuring business SLAs.

[0289] Some embodiments of this disclosure also provide a scheme for synergistic optimization of privacy protection and model performance. Addressing the unique data compliance requirements of the education sector, an improved federated learning scheme integrating homomorphic encryption and knowledge distillation is designed, ensuring data privacy and security while keeping model accuracy loss at an industry-leading level. This scheme places particular emphasis on the localization requirements for sensitive educational data.

[0290] Some embodiments of this disclosure can also provide adaptive policy generation technology based on the ATT&CK framework. By constructing an attack knowledge graph and a large language model policy engine, dynamic generation and continuous optimization of defense policies can be achieved, enabling the protection system to evolve to counter new threats.

[0291] The system and method of this disclosure have the following significant technical advantages compared with existing technical solutions:

[0292] First, it offers multi-dimensional threat collaborative protection capabilities. Existing technical solutions often employ independent models to handle specific types of attacks, resulting in limited protection scope. This solution utilizes a hybrid expert model (MoE) architecture, organically integrating four expert systems: spatiotemporal graph neural networks, multimodal convolutional networks, federated learning, and large language models. This constructs a comprehensive protection system covering network traffic, container behavior, data privacy, and policy decision-making. The gating network dynamically schedules the optimal expert combination based on real-time threat characteristics, enabling collaborative detection and handling across attack dimensions, thus addressing the narrow protection scope of existing solutions.

[0293] Second, a dynamic adaptive defense mechanism. Addressing the challenge of existing static defense strategies failing to adapt to the rapidly changing cloud environment, this solution employs a threat-driven dynamic response mechanism. By sensing real-time changes in the attack landscape, the system automatically adjusts its expert model combinations and resource allocation strategies, achieving adaptive optimization throughout the entire process from threat detection to response. Compared to related technologies with fixed-period updates, this solution significantly enhances the real-time response capability against new and mutated threats and attacks.

[0294] Third, the solution optimizes the balance between privacy protection and detection performance. Existing federated learning schemes often suffer from a significant drop in model accuracy due to an overemphasis on privacy protection. This scheme adopts a technical approach that combines local data processing with global knowledge distillation. Under the premise of strictly adhering to data privacy regulations, it effectively controls accuracy loss by optimizing model aggregation methods and feature extraction algorithms, achieving the optimal balance between privacy security and detection performance.

[0295] Fourth, resource utilization efficiency is improved. Security solutions using related technologies often suffer from system overload or insufficient protection due to unreasonable resource allocation. This solution establishes a precise mapping between defense strength and resource requirements, dynamically schedules computing resources based on threat levels, and leverages the sparse activation characteristics of the MoE architecture to invoke relevant expert models only when necessary. This design significantly reduces system resource overhead while ensuring effective protection.

[0296] Fifth, intelligent strategy evolution capability. Existing defense strategies mostly rely on manual rule maintenance, which is difficult to cope with rapidly evolving attack methods. This solution integrates ATT&CK attack knowledge graph and large language model technology to build a strategy generation engine with autonomous learning and evolution capabilities. It can automatically deduce and optimize the defense tactical chain, enabling the system to continuously maintain its ability to counter new threats.

[0297] The aforementioned technological advantages enable this solution to form a complete, dynamic, and efficient cloud-native security protection solution, systematically addressing the key shortcomings of existing technologies in terms of protection dimensions, response speed, privacy balance, resource efficiency, and policy intelligence.

[0298] With the accelerated digital transformation of education, cloud-native architecture has become the standard technology for online education platforms and teaching management systems, highlighting the growing need for dynamic security protection. This solution addresses the specific needs of education and training scenarios, such as multi-tenant isolation, teaching data privacy protection, and business continuity assurance, providing customized solutions that effectively meet the core security requirements of online education platforms, vocational training institutions, and educational information service providers.

[0299] The technical architecture boasts excellent scalability and adaptability. The modular defense system designed in this solution is not only suitable for the education sector, but its core technical framework can also be extended to other vertical industries with stringent requirements for data privacy and system stability, such as healthcare and financial services.

[0300] This solution can be quickly deployed as a standardized SaaS service to meet the needs of small and medium-sized educational institutions; it can also provide customized protection for large educational groups in a private form; and it can empower educational technology developers to build a more secure application ecosystem through open security capability APIs.

[0301] Figure 5 This is a structural block diagram illustrating a security defense system for an education and training platform according to other embodiments of the present disclosure. The security defense system includes a memory 510 and a processor 520.

[0302] in:

[0303] The memory 510 can be a disk, flash memory, or any other non-volatile storage medium. The memory is used for storing... Figure 2 to Figure 4 At least one of the instructions in the corresponding embodiment.

[0304] Processor 520 is coupled to memory 510 and can be implemented as one or more integrated circuits, such as a microprocessor or microcontroller. Processor 520 executes instructions stored in memory, facilitating the detection and identification of complex threat signatures.

[0305] In one embodiment, it can also be as follows: Figure 6 As shown, the security defense system 500 includes a memory 510 and a processor 520. The processor 520 is coupled to the memory 510 via a BUS bus 530. The security defense system 500 can also be connected to an external storage device 550 via a storage interface 540 to access external data, and can also be connected to a network or another computer system (not shown) via a network interface 560, which will not be described in detail here.

[0306] In this embodiment, data instructions are stored in a memory and then processed by a processor, which facilitates the detection and identification of complex threat characteristics.

[0307] In another embodiment, this disclosure also provides a computer-readable storage medium (e.g., a non-transitory computer-readable storage medium) having stored thereon computer program instructions that, when executed by a processor, are implemented. Figure 2 to Figure 4The disclosure includes at least one step of the method in a corresponding embodiment. Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, apparatus, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable non-transitory storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0308] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0309] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0310] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0311] In some embodiments of this disclosure, a computer program product is also provided, which includes a computer program or instructions that, when executed by a processor, implement the security defense method for an education and training platform as described above.

[0312] In some embodiments of this disclosure, a computer program is also provided, comprising: instructions that, when executed by a processor, cause the processor to perform the security defense method for an education and training platform as described above.

[0313] This concludes the detailed description of the present disclosure. To avoid obscuring the concept of the disclosure, some details known in the art have not been described. Those skilled in the art will fully understand how to implement the technical solutions disclosed herein based on the above description.

[0314] While specific embodiments of this disclosure have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of this disclosure. Those skilled in the art should understand that modifications can be made to the above embodiments without departing from the scope and spirit of this disclosure. The scope of this disclosure is defined by the appended claims.

Claims

1. A security defense system for an education and training platform, comprising: A traffic protection layer is used to process network traffic to obtain first feature data, wherein the first feature data is feature data related to the traffic data of the network traffic; A container security layer is used to process the network traffic to obtain second feature data, which is feature data related to container attack data of the network traffic; A privacy protection layer is used to process the network traffic to obtain third feature data, which is feature data related to the privacy data of the network traffic; A gated router is configured to receive the first feature data, the second feature data, and the third feature data; construct a feature vector based on the first feature data, the second feature data, and the third feature data; determine the weight of each of the n expert models according to the feature vector; arrange the n expert models in descending order of weight; and select the first k expert models to identify attack behavior information in the network traffic, where 1≤k≤n, and k and n are positive integers. and The intelligent decision-making layer is used to generate a defense strategy based on the attack behavior information, evaluate the effect of the defense strategy after its execution to obtain the effect information corresponding to each of the top k expert models, and feed the effect information back to the gated router.

2. The security defense system according to claim 1 further includes: The gated router is also used to update the weights of each expert model based on the current weights of each expert model and the effect information corresponding to each expert model.

3. The security defense system according to claim 1, wherein, The attack behavior information includes: a first threat feature vector obtained based on the first feature data, a second threat feature vector obtained based on the second feature data, and a third threat feature vector obtained based on the third feature data.

4. The security defense system according to claim 3, wherein, The n expert models include: A traffic analysis expert model is used to perform spatiotemporal feature analysis on the network traffic to detect network layer attack behaviors. A container security expert model is used to detect network traffic attacks on containers; A federal privacy-preserving expert model is used to detect network traffic attacks on privacy data and to implement cross-tenant knowledge sharing under privacy protection; and A decision expert model is used to generate an adaptive defense strategy based on multi-source threat intelligence obtained from the network traffic analysis, wherein the multi-source threat intelligence includes a first threat feature vector, a second threat feature vector, and a third threat feature vector.

5. The security defense system according to claim 4, wherein, The traffic analysis expert model is also used to generate defense strategies against network layer attack behaviors based on single-source threat intelligence of network layer attack behaviors. The container security expert model is also used to generate defense strategies against container attacks based on single-source threat intelligence of network traffic attack behavior on containers. The federal privacy protection expert model is also used to generate defense strategies against privacy data attacks based on single-source threat intelligence derived from network traffic attacks on privacy data.

6. The security defense system according to any one of claims 1 to 5, wherein, The first feature data includes: the entropy value and IP dispersion obtained by the traffic protection layer from processing the network traffic; The second feature data includes: the system call frequency and vulnerability score obtained by the container security layer from processing the network traffic; The third feature data includes: the gradient norm of the privacy data obtained by the privacy protection layer from processing the network traffic.

7. The security defense system according to claim 1, wherein, The gated router is further configured to obtain network traffic anomaly degree based on the first feature data, vulnerability exploitation severity based on the second feature data, data sensitivity based on the third feature data, calculate a threat assessment value based on the network traffic anomaly degree and its corresponding weight coefficient, the vulnerability exploitation severity and its corresponding weight coefficient, and the data sensitivity and its corresponding weight coefficient, determine the threat level of the network traffic attack behavior based on the threat assessment value, and determine the server resources required based on the threat level.

8. The security defense system according to claim 7, wherein, The threat level includes a low threat level, a medium threat level, and a high threat level, wherein the high threat level is higher than the medium threat level, and the medium threat level is higher than the low threat level; The gated router is used to determine that the network traffic attack behavior belongs to the low threat level when the threat assessment value is within the assessment value range corresponding to the low threat level, to determine that the network traffic attack behavior belongs to the medium threat level when the threat assessment value is within the assessment value range corresponding to the medium threat level, and to determine that the network traffic attack behavior belongs to the high threat level when the threat assessment value is within the assessment value range corresponding to the high threat level.

9. A security defense method for an education and training platform, comprising: The network traffic is processed using a traffic protection layer to obtain first feature data, which is feature data related to the traffic data of the network traffic. The network traffic is processed using a container security layer to obtain second feature data, which is feature data related to container attack data of the network traffic; The network traffic is processed using a privacy protection layer to obtain third feature data, which is feature data related to the privacy data of the network traffic; The system uses a gated router to receive the first feature data, the second feature data, and the third feature data. Based on the first feature data, the second feature data, and the third feature data, a feature vector is constructed. The weight of each of the n expert models is determined according to the feature vector. The n expert models are arranged in descending order of weight. The first k expert models are selected to identify the attack behavior information in the network traffic, where 1≤k≤n, and k and n are positive integers. and The intelligent decision layer generates a defense strategy based on the attack behavior information. After the defense strategy is executed, the effect of the defense strategy is evaluated to obtain the effect information corresponding to each of the top k expert models. The effect information is then fed back to the gated router.

10. The security defense method according to claim 9 further includes: The gated router updates the weights of each expert model based on the current weights of each expert model and the corresponding performance information.

11. The security defense method according to claim 9, wherein, The attack behavior information includes: a first threat feature vector obtained based on the first feature data, a second threat feature vector obtained based on the second feature data, and a third threat feature vector obtained based on the third feature data.

12. The security defense method according to claim 11, wherein, The n expert models include: A traffic analysis expert model is used to perform spatiotemporal feature analysis on the network traffic to detect network layer attack behaviors. A container security expert model is used to detect network traffic attacks on containers; A federal privacy-preserving expert model is used to detect network traffic attacks on privacy data and to implement cross-tenant knowledge sharing under privacy protection; and A decision expert model is used to generate an adaptive defense strategy based on multi-source threat intelligence obtained from the network traffic analysis, wherein the multi-source threat intelligence includes a first threat feature vector, a second threat feature vector, and a third threat feature vector.

13. The security defense method according to claim 12, wherein, The traffic analysis expert model is also used to generate defense strategies against network layer attack behaviors based on single-source threat intelligence of network layer attack behaviors. The container security expert model is also used to generate defense strategies against container attacks based on single-source threat intelligence of network traffic attack behavior on containers. The federal privacy protection expert model is also used to generate defense strategies against privacy data attacks based on single-source threat intelligence derived from network traffic attacks on privacy data.

14. The security defense method according to any one of claims 9 to 13, wherein, The first feature data includes: the entropy value and IP dispersion obtained by the traffic protection layer from processing the network traffic; The second feature data includes: the system call frequency and vulnerability score obtained by the container security layer from processing the network traffic; The third feature data includes: the gradient norm of the privacy data obtained by the privacy protection layer from processing the network traffic.

15. The security defense method according to claim 9 further includes: The gated router is used to obtain network traffic anomaly based on the first feature data, vulnerability exploitation severity based on the second feature data, and data sensitivity based on the third feature data. The threat assessment value is calculated based on the network traffic anomaly degree and the weight coefficient corresponding to the network traffic anomaly degree, the vulnerability exploitation severity and the weight coefficient corresponding to the vulnerability exploitation severity, and the data sensitivity and the weight coefficient corresponding to the data sensitivity. The threat level of the network traffic attack behavior is determined based on the threat assessment value. and The required server resources are determined based on the threat level.

16. The security defense method according to claim 15, wherein, The threat level includes a low threat level, a medium threat level, and a high threat level, wherein the high threat level is higher than the medium threat level, and the medium threat level is higher than the low threat level; The threat level of the network traffic attack behavior is determined based on the threat assessment value, including: If the threat assessment value is within the range corresponding to the low threat level, the attack behavior of the network traffic is determined to be of the low threat level. If the threat assessment value falls within the range corresponding to the medium threat level, the attack behavior of the network traffic is determined to belong to the medium threat level; and If the threat assessment value falls within the range corresponding to the high threat level, the attack behavior of the network traffic is determined to be of the high threat level.

17. A security defense system for an education and training platform, comprising: Memory; as well as A processor coupled to the memory, the processor being configured to execute the security defense method for an education and training platform as described in any one of claims 9 to 16 based on instructions stored in the memory.

18. A computer-readable storage medium having stored thereon computer instructions that, when executed by a processor, implement the security defense method for an education and training platform as described in any one of claims 9 to 16.

19. A computer program product comprising a computer program or instructions that, when executed by a processor, implement the security defense method for an education and training platform as described in any one of claims 9 to 16.

Images