Multi-terminal oriented federated learning security intelligent agent cooperative defense method and system

Abstract

This invention discloses a collaborative defense method and system for multi-terminal federated learning security agents. The method includes: initializing and configuring security agents for different types of terminal nodes and registering federated learning participation permissions for each terminal node; each terminal node training a local threat detection sub-model through a local security agent and uploading model parameters to a federated learning center node; the federated learning center node aggregating and generating a global collaborative defense model based on the model parameters of each terminal node using a weighted average algorithm; and each terminal node using the local threat detection sub-model and the global collaborative defense model to perform dual verification of local risks. This invention adopts a federated learning model, scheduling multiple terminal agents to share threat features through the federated learning center node, achieving a two-layer defense of "local detection + global collaboration," improving the identification rate of advanced threats propagating across terminals, and eliminating the defense blind spots of traditional solutions.

Description

Technical Field

[0001] This invention relates to the field of network security technology, and in particular to a collaborative defense method and system for federated learning security agents for multiple terminals. Background Technology

[0002] With the deepening of digital transformation, the large-scale deployment of various types of terminals, such as edge terminals, industrial control terminals, cloud servers, and IoT devices, has created a complex multi-terminal collaborative working environment. However, existing multi-terminal security defense solutions have the following shortcomings:

[0003] The risk of data privacy leakage is prominent: Traditional centralized security defense solutions require uploading the operation logs, behavioral data, and business data of each terminal to a central server for model training and threat detection. This leads to the risk of leakage and tampering of sensitive terminal data (such as industrial control instructions and user privacy information) during transmission and storage, which does not meet the requirements of the Cybersecurity Law and the Data Security Law for data privacy protection.

[0004] Lack of cross-terminal collaborative defense capabilities: In existing solutions, the security protection of each terminal operates independently, lacking effective threat information sharing and collaborative response mechanisms between terminals. When facing attacks such as advanced persistent threats (APTs) and ransomware that spread across multiple terminals, a single-terminal defense system struggles to identify the attack chain, easily creating blind spots and leading to an expansion of the attack scope.

[0005] Poor model generalization ability and adaptability to heterogeneous terminals: The detection models of traditional security agents rely on limited local data for training, resulting in weak model generalization ability. They are difficult to adapt to heterogeneous hardware environments, operating systems (such as Windows, Linux, or embedded systems) and business scenarios of different terminals (such as resource-constrained edge terminals and high-performance cloud terminals). At the same time, model updates require manual intervention and cannot achieve dynamic iterative optimization, making it difficult to cope with rapidly evolving network threats.

[0006] Insufficient threat detection accuracy and response efficiency: Existing solutions mostly use single feature matching or static rule detection methods, which have limited ability to identify unknown and mutated threats; and lack a hierarchical defense mechanism, using a uniform defense strategy for different levels of threats, resulting in wasted resources or insufficient defense, and high response latency (usually more than 5 seconds), which cannot meet the real-time protection needs of critical business scenarios. Summary of the Invention

[0007] The purpose of this invention is to provide a collaborative defense method and system for federated learning security agents for multiple terminals, so as to solve at least some of the problems in the prior art.

[0008] To achieve the above objectives, one aspect of the present invention provides a multi-terminal federated learning security agent cooperative defense method, comprising:

[0009] Initialize and configure the security agent for different types of terminal nodes, and complete the registration of federated learning participation permissions for each terminal node;

[0010] Each terminal node trains a local threat detection sub-model through a local security agent and uploads the model parameters to the federal learning center node.

[0011] The Federated Learning Center nodes aggregate and generate a global collaborative defense model based on the model parameters of each terminal node using a weighted average algorithm.

[0012] Each terminal node uses a local threat detection sub-model and a global collaborative defense model to perform dual verification of local risks. This includes: each terminal node first uses the local threat detection sub-model to perform local security detection on local data. If the detection result indicates that the risk level meets preset conditions, it sends a collaborative verification request carrying threat characteristics to the federated learning center node. The federated learning center node matches the associated terminals with threat characteristics based on global threat intelligence, sends the threat characteristics to the associated terminals, and receives the verification results of the associated terminals on the threat characteristics. The federated learning center node performs threat processing based on the local security detection results, the verification results of the associated terminals on the threat characteristics, and the global threat intelligence.

[0013] Preferably, the verification results of the associated terminal on the threat characteristics include the degree of correlation between the threat characteristics and the known risks of the associated terminal, and the global threat intelligence includes the set of threat information pre-received by the federated learning center node from each terminal node;

[0014] The Federated Learning Center node performs threat handling based on local security detection results, the verification results of threat signatures by associated terminals, and global threat intelligence, including:

[0015] Threats are classified and handled as follows: For low-risk threats, logs are recorded and the global threat intelligence database is updated; for high-risk threats, at least one of the following actions is taken: triggering local blocking on the terminal and synchronizing threat information to associated terminals; performing cross-terminal isolation, synchronizing global threat intelligence, and sending alarm information to the administrator; or triggering the manual emergency response channel.

[0016] Preferably, different types of terminal nodes include edge terminals, cloud terminals, or industrial control terminals;

[0017] The initial configuration of the security agent for each terminal node includes configuring the security agent to integrate a threat detection module, a local model training module, an encrypted communication module, a local defense execution module, and a permission management module.

[0018] Preferably, before uploading model parameters to the federated learning center node, the process also includes:

[0019] Each terminal node encrypts the communication between itself and the federated learning center node through an encrypted communication module, including: first adding dynamic noise to the parameters using a differential privacy algorithm, and then completing the transmission encryption using Paillier homomorphic encryption.

[0020] Preferably, completing the federated learning participation permission registration for each terminal node includes: managing the terminal node's access registration permission, parameter upload permission during the local model training phase, feature sharing permission during the collaborative detection phase, and cross-terminal operation permission during the defense response phase through the permission management module;

[0021] Before each terminal node trains the local threat detection sub-model through the local security agent, it also includes: dynamically configuring the local model training resource ratio based on the preset resource ratio threshold, business load status, and federated learning task priority through the permission management module.

[0022] Preferably, the federated learning center nodes aggregate and generate a global collaborative defense model based on the model parameters of each terminal node using a weighted average algorithm, including:

[0023] The Federated Learning Center node receives model parameters from each terminal node and verifies whether the terminal node uploading the model parameters has completed the registration for Federated Learning participation permissions, whether it has the qualification to upload parameters, and whether the communication link is encrypted.

[0024] If the terminal has completed the registration of federated learning participation permissions, has the qualification to upload parameters, and the communication link is encrypted, then the model parameters are verified, and a global collaborative defense model is generated based on the verified model parameters.

[0025] Preferably, each terminal node first performs local security checks on local data using a local threat detection sub-model, including:

[0026] Static threats are identified by matching terminal behavior characteristics based on a preset security baseline; among them, static threats include one or more of the following: abnormal process startup, illegal port access, and data transmission traffic exceeding the threshold.

[0027] Preferably, the preset safety baseline includes:

[0028] Non-core baseline rules are used to identify low-risk threats;

[0029] The core baseline rules are used to determine whether a threat is of medium or high risk.

[0030] Preferably, sending a collaborative verification request carrying threat characteristics to the federated learning center node includes:

[0031] Threat signatures are anonymized and encrypted;

[0032] Send a collaborative verification request carrying the threat signature after de-identification and encryption to the federated learning center node.

[0033] Another aspect of the present invention provides a multi-terminal federated learning security agent collaborative defense system for implementing the multi-terminal federated learning security agent collaborative defense method provided by the above aspect of the present invention. The system includes different types of terminal nodes and federated learning center nodes, wherein...

[0034] The terminal node is used to initialize and configure the security agent and complete the registration of federated learning participation permissions; train the local threat detection sub-model through the local security agent and upload the model parameters to the federated learning center node; use the local threat detection sub-model to perform local security detection on local data, and if the detection result shows that the risk level meets the preset conditions, send a collaborative verification request carrying threat characteristics to the federated learning center node.

[0035] The Federated Learning Center node is used to aggregate and generate a global collaborative defense model based on the model parameters of each terminal node using a weighted average algorithm; it receives collaborative verification requests carrying threat features from terminal nodes, matches related terminals with threat features based on global threat intelligence, sends threat features to related terminals, and receives the verification results of the threat features from related terminals; and it performs threat processing based on the local security detection results of terminal nodes, the verification results of the threat features from related terminals, and global threat intelligence.

[0036] Compared with the prior art, the present invention has at least the following beneficial effects:

[0037] By adopting a federated learning model, terminal data is trained locally, and only encrypted model parameters are uploaded, blocking the data leakage path from both the transmission and training stages, which fully complies with the relevant regulations on data privacy protection. By scheduling multiple terminal agents to share threat characteristics through the federated learning center node, a two-layer defense of "local detection + global collaboration" is achieved, which improves the identification rate of advanced threats spreading across terminals and eliminates the defense blind spots of traditional solutions. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a multi-terminal federated learning security agent collaborative defense method provided in an embodiment of the present invention.

[0039] Figure 2 This is a flowchart illustrating a multi-terminal federated learning security agent collaborative defense method provided in Embodiment 2 of the present invention.

[0040] Figure 3This is a flowchart illustrating a multi-terminal federated learning security agent collaborative defense method provided in Embodiment 3 of the present invention. Detailed Implementation

[0041] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0042] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be used interchangeably where appropriate to understand the embodiments of the invention described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a product or device comprising a series of units is not necessarily limited to those explicitly listed, but may include other units not explicitly listed or inherent to such product or device.

[0043] In this invention, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing the invention and its embodiments, and are not intended to limit the indicated devices, elements, or components to having a specific orientation, or to be constructed and operated in a specific orientation.

[0044] Furthermore, in addition to indicating direction or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in certain situations to indicate a dependency or connection. Those skilled in the art can understand the specific meaning of these terms in this invention based on the specific circumstances.

[0045] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linking," and "socketing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral structure; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of these terms in this invention based on the specific circumstances.

[0046] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0047] Example 1

[0048] This invention provides a federated learning-based collaborative defense method for secure intelligent agents across multiple terminals. Figure 1 A flowchart illustrating this method is shown, as follows: Figure 1 As shown, the method includes:

[0049] Step 101: Initialize and configure the security agent for different types of terminal nodes, and complete the registration of federated learning participation permissions for each terminal node.

[0050] Step 102: Each terminal node trains a local threat detection sub-model through a local security agent and uploads the model parameters to the federated learning center node.

[0051] Step 103: The federated learning center nodes aggregate and generate a global collaborative defense model based on the model parameters of each terminal node using a weighted average algorithm.

[0052] Step 104: Each terminal node uses a local threat detection sub-model and a global collaborative defense model to perform dual verification of local risks. This includes: Each terminal node first uses the local threat detection sub-model to perform local security detection on local data. If the detection result indicates that the risk level meets preset conditions, it sends a collaborative verification request carrying threat characteristics to the federated learning center node. The federated learning center node matches the associated terminals with threat characteristics based on global threat intelligence, sends the threat characteristics to the associated terminals, and receives the verification results of the threat characteristics from the associated terminals. The federated learning center node performs threat processing based on the local security detection results, the verification results of the threat characteristics from the associated terminals, and the global threat intelligence.

[0053] Among them, the verification results of the associated terminals on the threat characteristics include the degree of correlation between the threat characteristics and the known risks of the associated terminals, and the global threat intelligence includes the set of threat information pre-received by the federated learning center nodes from each terminal node;

[0054] The Federated Learning Center node performs threat handling based on local security detection results, the verification results of threat signatures by associated terminals, and global threat intelligence, including:

[0055] Threats are classified and handled as follows: For low-risk threats, logs are recorded and the global threat intelligence database is updated; for high-risk threats, at least one of the following actions is taken: triggering local blocking on the terminal and synchronizing threat information to associated terminals; performing cross-terminal isolation, synchronizing global threat intelligence, and sending alarm information to the administrator; or triggering the manual emergency response channel.

[0056] Different types of terminal nodes can include edge terminals, cloud terminals, or industrial control terminals, among other types.

[0057] The initial configuration of the security agent for each terminal node includes configuring the security agent to integrate a threat detection module, a local model training module, an encrypted communication module, a local defense execution module, and a permission management module.

[0058] In the preferred implementation scheme, the process of uploading model parameters to the federated learning center node further includes:

[0059] Each terminal node encrypts the communication between itself and the federated learning center node through an encrypted communication module, including: first adding dynamic noise to the parameters using a differential privacy algorithm, and then completing the transmission encryption using Paillier homomorphic encryption.

[0060] In a preferred implementation scheme, completing the federated learning participation permission registration for each terminal node includes: managing the terminal node's access registration permission, parameter upload permission during the local model training phase, feature sharing permission during the collaborative detection phase, and cross-terminal operation permission during the defense response phase through the permission management module.

[0061] Before each terminal node trains the local threat detection sub-model through the local security agent, it also includes: dynamically configuring the local model training resource ratio based on the preset resource ratio threshold, business load status, and federated learning task priority through the permission management module.

[0062] In a preferred implementation scheme, the federated learning center nodes aggregate and generate a global collaborative defense model based on the model parameters of each terminal node using a weighted average algorithm, including:

[0063] The Federated Learning Center node receives model parameters from each terminal node and verifies whether the terminal node uploading the model parameters has completed the registration for Federated Learning participation permissions, whether it has the qualification to upload parameters, and whether the communication link is encrypted.

[0064] If the terminal has completed the registration of federated learning participation permissions, has the qualification to upload parameters, and the communication link is encrypted, then the model parameters are verified, and a global collaborative defense model is generated based on the verified model parameters.

[0065] In a preferred implementation scheme, each terminal node first performs local security checks on local data using a local threat detection sub-model, including:

[0066] Static threats are identified by matching terminal behavior characteristics based on a preset security baseline; among them, static threats include one or more of the following: abnormal process startup, illegal port access, and data transmission traffic exceeding the threshold.

[0067] In a preferred embodiment, the preset safety baseline includes:

[0068] Non-core baseline rules are used to identify low-risk threats;

[0069] The core baseline rules are used to determine whether a threat is of medium or high risk.

[0070] In a preferred implementation scheme, sending a collaborative verification request carrying threat characteristics to the federated learning center node includes:

[0071] Threat signatures are anonymized and encrypted;

[0072] Send a collaborative verification request carrying the threat signature after de-identification and encryption to the federated learning center node.

[0073] The method provided in this embodiment of the invention adopts a federated learning model, in which terminal data is trained locally and only encrypted model parameters are uploaded, blocking the data leakage path from both the transmission and training stages, which fully complies with the relevant regulations on data privacy protection; by scheduling multiple terminal agents to share threat features through the federated learning center node, a two-layer defense of "local detection + global collaboration" is achieved, which improves the identification rate of advanced threats spreading across terminals and eliminates the defense blind spots of traditional solutions.

[0074] Example 2

[0075] Based on the same technical concept as in Embodiment 1, this embodiment of the invention provides a federated learning-based collaborative defense method for secure intelligent agents across multiple terminals. Figure 2 A flowchart illustrating this method is shown, as follows: Figure 2 As shown, the method includes:

[0076] Step 201: Initialize and configure multiple terminal nodes.

[0077] Deploy independent security agents in target terminals such as edge terminals, cloud terminals, and industrial control terminals. These security agents integrate threat detection modules, local model training modules, encrypted communication modules, local defense execution modules, and access control modules.

[0078] ① Threat Detection Module: It integrates static matching of security baselines with dynamic inference capabilities of federated dual models. The static baseline can automatically generate a customized rule base based on the terminal type (edge / cloud / industrial control). The dynamic inference synchronously calls the local sub-model and the global collaborative model for dual verification. The model inference process is linked with the local hardware resource status in real time (such as automatically switching to lightweight inference mode when edge terminal resources are insufficient). Unlike existing modules with single detection logic, it has the original ability of adaptive and dual-engine detection.

[0079] ② Local model training module: Equipped with a heterogeneous terminal adaptive training scheduler, it can automatically adjust training parameters according to the terminal hardware specifications (CPU computing power, memory capacity, network bandwidth). For example, edge terminals automatically reduce the batch size to 16 and shorten the iteration rounds to 50 rounds, while cloud terminals enable full parameter training. At the same time, it has a built-in "defense log-training sample" automatic conversion interface, which can convert local defense execution data into labeled samples in real time for model iteration, solving the problem that traditional training modules cannot adapt to heterogeneous terminal technology.

[0080] ③ Encrypted Communication Module: It pioneers a dual-track protection channel of differential privacy + homomorphic encryption. When transmitting model parameters, dynamic noise is first added to the parameters through differential privacy algorithm (the noise intensity is adaptively adjusted according to the sensitivity of terminal data), and then the transmission is encrypted through Paillier homomorphic encryption. The module also has a built-in "encrypted parameter - terminal permission" binding mechanism, which only allows terminal nodes that have been certified by the permission management module to decrypt and receive the data. Its dual-layer encryption and permission binding collaborative protection logic technology.

[0081] ④ Local Defense Execution Module: Constructs an adaptive mapping engine for threat level, endpoint resources, and defense strategy. It can automatically match differentiated defense actions based on threat level (e.g., low, medium, high, critical) and endpoint resource status (e.g., when an edge endpoint encounters a medium-risk threat, it only performs process blocking, while cloud endpoints simultaneously start traffic scrubbing and log tracing). It also supports federated-level synchronization of defense strategies (i.e., local defense actions can be synchronized to the global threat intelligence database as features), breaking through the limitations of traditional defense modules that are "fixed in strategy and lack global linkage".

[0082] ⑤ Permission Management Module: The innovative design of the full lifecycle permission control system for federated learning not only has access and registration permissions for terminal nodes, but also allows for fine-grained control over parameter upload permissions during the model training phase, feature sharing permissions during the collaborative detection phase, and cross-terminal operation permissions during the defense and response phase. Moreover, the permission level is dynamically linked to the terminal model accuracy and data quality (e.g., terminals with model accuracy below 85% are automatically restricted from voting on global model parameters), forming a unique dynamic permission control mechanism in federated scenarios.

[0083] In this step, the security agent of the terminal node can be responsible for unified resource scheduling, data bus management and module communication through the microkernel. Each functional module (threat detection, local model training, etc.) is connected to the microkernel in the form of plug-in, realizing flexible integration of "plug and play".

[0084] The microkernel has a built-in cross-module data bus (based on a lightweight MQTT protocol or shared memory mechanism) that supports real-time data flow between modules. For example, terminal operation data collected by the threat detection module can be synchronously pushed to the local model training module as training samples, and the transmission status of the encrypted communication module can be fed back to the permission management module for permission verification.

[0085] In one implementation, an independent resource quota is allocated to each module (e.g., the CPU usage of the local model training module on the edge terminal is ≤30%), which is dynamically controlled by the microkernel's resource scheduling unit to avoid overloading of a single module and affecting the overall operation of the agent.

[0086] In this embodiment, each terminal completes the registration of federated learning participation permissions through the permission management module, presets a security baseline (including terminal process whitelist, port access rules, data transmission thresholds, etc.), and dynamically configures the proportion of local model training resources based on terminal hardware resources (CPU computing power, memory capacity, network bandwidth) (edge ​​terminal resource proportion ≤30%, cloud terminal resource proportion ≤60%).

[0087] ① Hardware resource information collection and profile construction

[0088] After the security agent is deployed on the terminal, the permission management module will trigger the built-in hardware monitoring unit to conduct a 5-minute baseline collection of the terminal's core hardware resources and generate a terminal hardware resource profile, including collection dimensions, data processing, and terminal type determination.

[0089] ② Initial allocation of resource proportion threshold

[0090] The permission management module, based on terminal type, calls a preset basic resource allocation strategy library to complete the initial resource allocation, including: edge terminals, which have limited hardware resources and need to ensure core business; cloud terminals, which have sufficient hardware resources and lower real-time requirements for business, ensuring model training efficiency; and industrial control terminals, which prioritize ensuring the hard real-time performance of industrial control command execution.

[0091] ③ Dynamic adjustment based on business load

[0092] The resource scheduling unit of the security intelligent agent monitors the terminal business load in real time. When the business load fluctuates, it automatically adjusts the proportion of model training resources to ensure that the training task and the core business do not interfere with each other. When the terminal core business load exceeds the preset threshold (such as the CPU usage of industrial control terminal ≥70%, the memory usage of edge terminal ≥80%, and the bandwidth usage of cloud terminal ≥80%), the resource proportion adjustment is triggered.

[0093] Adjustment logic: If the business load increases (e.g., the CPU usage of the edge terminal rises to 85% due to the connection of more sensors), the resource scheduling unit will reduce the proportion of model training resources in a 10% gradient, for example, from 20% to 10%, and at the same time suspend non-urgent model training tasks (such as regular iterative training), and only retain incremental training of threat samples.

[0094] ④ Linked adaptation of federated learning task priorities

[0095] The resource allocation process is also linked to the priority of federated learning tasks to ensure the supply of resources for critical tasks:

[0096] Task Priority Classification: Federated learning tasks are divided into three priorities: emergency iteration (triggered when global model accuracy is <95%), regular iteration (once every 24 hours), and incremental training (triggered when new threat samples are added);

[0097] Resource allocation strategy: If the terminal undertakes an urgent iteration task, the resource allocation can be temporarily increased (up to 35% for edge terminals and 65% for cloud terminals) without affecting core business, and an extension of the training window can be requested from the federation center through the encrypted communication module; if it is a regular iteration task, the initial resource allocation threshold should be strictly followed to avoid excessive resource consumption.

[0098] Step 202: Construct and train the federated learning collaborative model.

[0099] Each terminal security agent collects local terminal operation data (including, for example, log data, traffic data, behavior data, and hardware status data) through the local model training module, and trains a local threat detection sub-model based on a lightweight CNN network (for example, with 128 input feature dimensions, including CPU utilization, memory usage, network connection count, process feature vectors, port access records, etc.). The training parameters are set to, for example, 100 iterations, a learning rate of 0.001, and a batch size of 32.

[0100] Training the local threat detection sub-model is the "source" of the threat detection module's dynamic detection capability: the terminal first completes the training of the sub-model through the local model training module, and then deploys the trained sub-model to the dynamic detection unit of the threat detection module. Only then can the dynamic detection unit call the sub-model to infer the real-time collected terminal operation data and realize dynamic threat identification.

[0101] In this invention, the local threat detection sub-model is the core inference model deployed within the security intelligence body of each terminal for real-time threat identification on the terminal side. Its design adapts to heterogeneous environments across multiple terminals, and features lightweight, high accuracy, and self-iterative characteristics. This sub-model adopts a lightweight CNN (Convolutional Neural Network) architecture, with an overall network structure of a simplified "input layer - convolutional layer - pooling layer - fully connected layer - output layer," adapting to the limited hardware resources of the terminal. Specific layer parameters are as follows:

[0102] Input layer: The feature dimension is, for example, 128 dimensions, covering the core state and behavioral features of the terminal operation, including 7 major categories of features such as CPU utilization, memory usage, number of network connections, process feature vector, port access records, hardware temperature, and data transmission rate. All features are normalized and uniformly mapped to the [0,1] interval to ensure the consistency of model input.

[0103] Convolutional layers: For example, set two convolutional kernels. The first convolutional kernel has a size of 3×3 and a quantity of 16, and the second convolutional kernel has a size of 3×3 and a quantity of 32. It captures abnormal patterns of terminal behavior (such as abnormal process startup frequency and abnormal port access timing) through local feature extraction.

[0104] Pooling layer: Max pooling is used, with a pooling kernel size of, for example, 2×2, to reduce feature dimensionality and computational cost while retaining key feature information;

[0105] Fully connected layer: Set up, for example, 1 hidden layer (64 neurons) and 1 output layer. The hidden layer uses the ReLU activation function to achieve non-linear feature mapping, and the output layer uses the Sigmoid activation function to output the threat confidence in the range of 0-1 (confidence ≥85% is judged as a clear threat, <85% is judged as a suspected unknown threat).

[0106] The model's overall parameter count is kept below 1 million, with inference time on edge terminals ≤10ms and inference time on cloud terminals ≤5ms, balancing detection efficiency with terminal resource consumption.

[0107] The logic for model training in this step includes:

[0108] The training of the sub-model is completed independently by the local model training module of the terminal security agent, without the need to upload the original terminal data. The specific training process is as follows:

[0109] Training data sources: collect local terminal operation data, such as the data from the past 3 days, including log data (10GB / terminal), traffic data (5GB / terminal), behavioral data (process start / stop records, instruction execution records), and hardware status data (real-time CPU / memory / bandwidth usage). The data is labeled based on a security baseline (e.g., “non-whitelist process startup” is labeled as a threat sample, and “compliant port communication” is labeled as a normal sample).

[0110] Training parameter configuration: fixed number of iterations, e.g., 100 rounds; learning rate, e.g., 0.001; batch size, e.g., 32; cross-entropy loss function is used to optimize model parameters; for edge terminals, gradient clipping will be enabled during training (gradient threshold set to 1.0) to avoid gradient explosion during training, while limiting the proportion of training resources to ≤, e.g., 30%, so as not to affect the operation of the terminal's core business;

[0111] Privacy enhancement processing: After training, noise perturbation (noise intensity, for example, ε=0.1) will be added to the model parameters based on differential privacy technology. This avoids terminal data leakage caused by parameter reverse inference and ensures the detection accuracy of the model.

[0112] An improved FedAvg algorithm based on differential privacy is adopted to perform noise perturbation on the local sub-model parameters. The encrypted model parameter update values ​​are then uploaded to the federated learning center node through an encrypted communication module (e.g., using the Paillier homomorphic encryption algorithm with a key length of 2048 bits).

[0113] The model aggregation unit of the federated learning center node receives the updated model parameters from each terminal, aggregates them based on a weighted average algorithm (the weight coefficients are positively correlated with the terminal data quality and model accuracy) to generate a global collaborative defense model, and feeds it back to each terminal after encryption to complete the first round of model training.

[0114] Before aggregating and generating a global collaborative defense model, the model aggregation unit needs to complete two pre-verification checks to ensure the legality and validity of the aggregated data: ① Terminal identity and permission verification: The model aggregation unit connects to the terminal identity database of the permission management module to verify whether the terminal uploading parameters has completed federated learning permission registration and whether it has the qualification to upload parameters; at the same time, it verifies whether the terminal's communication link is encrypted. If the identity or link is invalid, the parameter update value of that terminal is directly discarded. ② Parameter validity verification: The model parameter update values ​​uploaded by each terminal are verified for format and rationality: Format verification: Confirm that the parameter dimension is consistent with the parameter dimension of the preset lightweight CNN model (e.g., the convolutional layer weight dimension is 3×3×16, and the fully connected layer bias dimension is 64-dimensional) to avoid parameter dimension errors caused by abnormal local training of the terminal; Rationality verification: Through gradient range verification (e.g., setting the gradient threshold to [-1,1]), abnormal parameters with gradient values ​​exceeding the normal range (e.g., gradient explosion parameters caused by terminal hardware failure) are removed to ensure that the parameters participating in the aggregation are valid. The model aggregation unit generates the global model based on the process of "parameter preprocessing → weight allocation → weighted aggregation → model verification".

[0115] Step 203, Multi-agent collaborative threat detection.

[0116] Each terminal security agent collects terminal operation data in real time through the threat detection module, and inputs it into both the local threat detection sub-model and the global collaborative defense model for dual verification.

[0117] Static detection: Based on the security baseline, it matches the terminal behavior characteristics to identify static threats such as abnormal process startup, illegal port access, and data transmission traffic exceeding the threshold;

[0118] Dynamic detection: Identifies dynamic threats such as malicious code injection, data theft, and remote control through real-time model reasoning;

[0119] The technical implementation of static detection is as follows.

[0120] Rule-based matching technology: A predefined security baseline rule base (including process whitelists, port access rules, etc.) identifies static threats through feature matching. For example, using the Snort firewall's rule base, configuring a rule to "allow only communication on ports 80 / 443" will trigger an alarm when unauthorized port access is detected, thus identifying the threat of illegal port access.

[0121] Static code analysis technology: This involves analyzing static text such as terminal configuration files and process startup scripts to detect unauthorized configurations. For example, using the Cppcheck tool to scan the PLC startup script of an industrial control terminal can identify "non-whitelisted process startup commands," thus identifying potential threats of abnormal process startup.

[0122] Data flow statistics threshold detection technology: This technology presets a data transmission traffic threshold and tracks the amount of data transmitted within a statistical period. Exceeding this threshold indicates a threat. For example, deploying a NetFlow analytics tool on an IoT gateway and setting a threshold of "data transmission ≤ 500MB per terminal within 5 minutes" will trigger an alarm when this limit is exceeded, thus identifying a data transmission traffic exceeding the threshold threat.

[0123] The technical implementation of dynamic detection is as follows.

[0124] Machine learning model inference techniques: Training dynamic threat detection models to infer endpoint behavior in real time. For example, using a detection model based on lightweight CNNs to analyze process instruction sequences in endpoint memory in real time, identifying characteristic patterns of malicious code injection, and correspondingly identifying malicious code injection threats.

[0125] Anomaly detection technology: Establish a baseline for normal terminal behavior and monitor dynamic behaviors that deviate from the baseline. For example, using UEBA (User and Entity Behavior Analysis) technology, learn the normal data access patterns of cloud servers. When "unauthorized IP addresses reading database data in batches" is detected, it is determined to be data theft, and the corresponding data theft threat is identified.

[0126] Real-time memory monitoring technology: Monitors dynamic operations such as process interactions and system calls in terminal memory. For example, the Volatility tool can be used to analyze terminal memory in real time to detect behaviors such as "remote processes injecting into memory and executing control commands," thereby identifying remote control threats.

[0127] In this step, when the local detection result is low risk, only logs are recorded and the local model training samples are updated; when the detection result is medium risk, a local alarm is triggered; when the detection result is high risk or suspected unknown threat (model confidence < 85%), a collaborative verification request is generated, which includes encrypted and de-identified threat features (such as attack traffic signature codes and abnormal behavior sequences), and is sent to the federated learning center node through the encrypted transmission module.

[0128] in:

[0129] ① Local testing process:

[0130] The threat detection module collects real-time terminal operation data, covering 128-dimensional feature data such as process start / stop records, port access logs, data transmission traffic, CPU / memory usage, and network connection topology. The data is then standardized (e.g., mapping CPU usage to the [0,1] range and performing feature encoding on port access records), while filtering out invalid noise data (such as instantaneous memory fluctuations) to generate a detection input dataset. The preprocessed data is then matched against a locally preset security baseline library.

[0131] Low-risk assessment rules:

[0132] Only non-core baseline rules are triggered (such as occasional unauthorized port probes on cloud terminals, but no communication connection is established; data transmission traffic exceeds the threshold by less than 5% and is a momentary fluctuation), without affecting the core business of the terminal;

[0133] • Rules for determining medium risk:

[0134] Triggering 1-2 core baseline rules (such as industrial control terminals starting non-whitelist industrial control processes, data transmission traffic exceeding the threshold by 20%-50%) has already affected some terminal services;

[0135] • High-risk assessment rules:

[0136] Triggering three or more core baseline rules (such as simultaneous abnormal process startup, illegal port communication, or large-volume data outflow), or directly triggering the "core business blocking" rule (such as PLC control process being tampered with or database service port being maliciously occupied).

[0137] • Risk assessment linkage verification mechanism

[0138] To avoid false positives, the local detection system also incorporates dual verification logic:

[0139] Cross-module cross-validation: If static detection determines it to be high risk, but the confidence level of dynamic detection is less than 0.5, the threat detection module will call the historical defense logs of the local defense execution module for comparison to eliminate false alarms (such as baseline triggering caused by compliance operation).

[0140] Confidence fluctuation correction: If the dynamic detection confidence is at the level threshold (e.g., 0.48-0.52), the detection window period will be extended by 10 seconds to collect more data for re-inference and ensure the accuracy of the level determination.

[0141] ② Innovative attributes and generation methods of collaborative verification requests

[0142] The generation logic and triggering mechanism of collaborative verification requests are the key implementation carriers of the innovation of the "static + dynamic dual collaborative detection mechanism" in this invention. It is a unique technical solution under the integrated architecture of federated learning and secure intelligent agents, which is significantly different from the independent detection and alarm mode of traditional terminals.

[0143] When a traditional terminal detects a suspected threat, it only triggers a local alarm and lacks cross-terminal collaborative verification logic. However, the collaborative verification request of this invention is generated based on a composite judgment result of "local dual-model inference (local sub-model + global model) + security baseline matching", and the request content is encrypted and de-identified threat features. This not only protects data privacy but also realizes cross-terminal threat collaborative verification, making it an important component of the overall collaborative defense architecture.

[0144] The collaborative verification request is generated by the threat detection module of the endpoint security agent in conjunction with the encrypted communication module:

[0145] The threat detection module extracts core threat characteristics from terminal runtime data, categorizing them into two types:

[0146] Attack traffic signature: includes the source / destination IP of the abnormal traffic (masked, only network segment information is retained), port number, packet size distribution, transmission protocol type, traffic timing characteristics, etc., generating a 128-bit traffic signature hash value;

[0147] Abnormal behavior sequences include process start-stop sequences (such as "starting a non-whitelisted process → reading a sensitive file → establishing an external network connection"), system call instruction sequences, abnormal hardware resource usage curves, etc., and the behavior sequences are transformed into standardized feature vectors (64-dimensional).

[0148] At the same time, the features are anonymized: sensitive fields that can identify the terminal (such as terminal MAC address and device serial number) are deleted, and only threat-related behaviors and traffic features are retained.

[0149] The encrypted communication module uses the Paillier homomorphic encryption algorithm (2048-bit key length) to encrypt the de-identified threat features. Then, it encapsulates the encrypted features along with basic terminal information (containing only terminal type identifiers, such as "Industrial Control Terminal-01," without specific identity information) into a collaborative verification request. The request format is JSON, as shown in the following example:

[0150] {

[0151] "terminal_type":"industrial_control",

[0152] "request_id":"FL-SR-20251208-001",

[0153] "encrypted_traffic_feature":"Paillier encrypted 128-bit traffic feature hash value",

[0154] "encrypted_behavior_sequence":"Paillier's encrypted 64-dimensional anomalous behavior feature vector",

[0155] "local_detection_confidence":0.82,

[0156] "trigger_rule":

[0157]

[0158] The "federation scenario adaptation + dual-protocol encrypted transmission" logic of the encrypted transmission module in this invention is a unique design, and its core encryption algorithm and transmission mechanism can be implemented through existing technologies.

[0159] Step 204, hierarchical and coordinated defense response.

[0160] After receiving the collaborative verification request, the collaborative scheduling unit of the federated learning center node matches associated terminals (such as terminals within the same local area network or terminals with data interaction) based on the global threat intelligence unit, and schedules the associated terminal security agents to share the encrypted threat characteristics.

[0161] The threat level assessment unit combines local detection results, feedback characteristics from associated terminals, and global threat intelligence to classify threats into four levels: low, medium, high, and critical.

[0162] • Low risk: Only logs are recorded and the global threat intelligence database is updated;

[0163] Medium risk: Triggers local blocking on the terminal (such as terminating abnormal processes or closing illegal ports) and synchronizes threat information to associated terminals;

[0164] • High risk: Perform cross-terminal isolation (e.g., disconnect the communication link between the threat terminal and associated terminals), synchronize global threat intelligence, and send alert information to the administrator;

[0165] • Fatal Risk: Based on high-risk defense measures, trigger the manual emergency response channel, such as the intervention of a 24 / 7 expert team;

[0166] After the defense is completed, each terminal security agent records the defense process data (including threat characteristics, defense measures, and execution results).

[0167] The triggering logic for the associated terminal feedback features includes:

[0168] The features fed back by the associated terminals are not passively fed back after receiving threat features sent by the global threat intelligence unit. Instead, after the collaborative scheduling unit of the federated learning center node receives the collaborative verification request, it first sends the encrypted and de-identified threat features of the requesting terminal to the associated terminals. The associated terminals complete local verification based on their own detection capabilities and then send the features corresponding to the verification results back to the central node. The global threat intelligence unit only provides a reference for the feature verification of the associated terminals, rather than directly sending features to trigger feedback.

[0169] The specific process and feedback content for the associated terminal feedback features are as follows:

[0170] • Receive feature verification instructions

[0171] The collaborative scheduling unit of the Federated Learning Center node sends encrypted feature verification instructions to associated terminals (terminals on the same local area network, those with data interaction, or those with similar business operations). The instructions contain two core parts: first, the de-identified threat features of the requesting terminal (such as attack traffic signature codes and abnormal behavior sequences); and second, the historical features of similar threats provided by the global threat intelligence unit (such as typical port access features of similar ransomware).

[0172] Local feature matching and verification

[0173] The threat detection module of the associated endpoint security agent compares the received features with its own local data in multiple dimensions:

[0174] Compare local traffic logs, process records, and port access data from the past 24 hours to confirm whether there are any threat characteristics consistent with the terminal that initiated the request.

[0175] Based on the historical characteristics of the global threat intelligence unit, determine whether the characteristic belongs to a variant of a known threat family;

[0176] If similar features exist locally, collect additional information about the feature in the local environment (such as the time when the feature appeared and the associated business processes).

[0177] Threat-related characteristics

[0178] Core content: Based on the historical feature database of the global threat intelligence unit, feedback is provided on the correlation between the feature and known threats (e.g., "the feature overlap with the industrial control ransomware variant in X month of 2025 reaches 85%"), and whether there are derivative threats of the feature locally (e.g., "data theft behavior accompanied by this feature was detected").

[0179] Function: To assist central nodes in accurately determining threat levels. For example, if the feedback characteristics show a connection to a high-risk APT attack, the threat level can be directly upgraded to "lethal risk".

[0180] Step 205: Dynamically iteratively optimize the model.

[0181] Each terminal security agent uses the defense process data as new training samples, repeats steps S2-S4, and completes a model iteration update every 24 hours; the federated learning center node regularly (weekly) evaluates the performance of the global collaborative defense model. When the model detection accuracy is <95%, an emergency iteration is triggered to ensure that the model adapts to the ever-changing threat environment and terminal scenarios.

[0182] Compared with existing technologies, the present invention has the following significant advantages:

[0183] • Significantly enhanced data privacy protection capabilities: Adopting a federated learning model, terminal data is only used for local training, and only encrypted model parameters are uploaded. Combining homomorphic encryption and differential privacy technology, data leakage paths are blocked from both the transmission and training stages, fully complying with relevant data privacy protection regulations.

[0184] • Cross-terminal collaborative defense with no blind spots: By scheduling multiple terminal agents to share threat characteristics through federated learning center nodes, a two-layer defense of "local detection + global collaboration" is achieved, which improves the recognition rate of advanced threats spreading across terminals to over 98% and eliminates the defense blind spots of traditional solutions.

[0185] • Strong adaptability to heterogeneous terminals: The security intelligent agent adopts a modular design, supports multiple operating systems such as Windows, Linux, and embedded systems, and adapts to the hardware differences between edge terminals and cloud terminals by dynamically configuring resource allocation, improving deployment flexibility by 60%;

[0186] • Optimized detection accuracy and response efficiency: Dual verification mechanism and dynamic iterative model achieve threat detection accuracy ≥95% and unknown threat identification rate ≥80%; tiered defense system reduces response latency to ≤1 second, an 80% reduction compared to traditional solutions.

[0187] • Reduced operation and maintenance costs: Automated model iteration and collaborative defense mechanisms reduce manual intervention, improving operation and maintenance efficiency by 70%, and are suitable for security management of large-scale multi-terminal clusters.

[0188] Example 3

[0189] Based on the same technical concept as in Embodiment 1, this embodiment of the invention provides a federated learning-based collaborative defense method for secure intelligent agents across multiple terminals. Figure 3 A flowchart illustrating this method is shown, as follows: Figure 3 As shown, the method includes:

[0190] Step 301, Initialize deployment.

[0191] Deploy secure agents one by one in, for example, 100 terminals, and complete federated learning participation permission registration through the permission management module.

[0192] Set a security baseline for edge terminals: the process whitelist includes core industrial control processes (such as PLC control processes), port access rules restrict communication to only ports 80, 443, and 502, and local model training resources account for 20%.

[0193] Set a security baseline for cloud terminals: the process whitelist includes web service processes and database processes, the port access rules open ports 80, 443, and 3306, and local model training resources account for 50%.

[0194] Step 302: Perform model training.

[0195] Each terminal security agent collects local operating data (10GB log data / terminal, 5GB traffic data / terminal) for up to 3 days, and trains a local sub-model based on a lightweight CNN network. The training process takes about 2 hours.

[0196] The model parameters are perturbed with noise using the FedAvg algorithm based on differential privacy, and then encrypted using the Paillier homomorphic encryption algorithm before being uploaded to the central node.

[0197] The central node receives the updated model parameters from 100 terminals and aggregates them into a global collaborative defense model based on a weighted average algorithm (the weight coefficient of the top 20% of terminals in the data quality score is 0.015, and the rest is 0.009). The aggregation takes about 30 minutes, the model size is 500MB, and it is encrypted and then fed back to each terminal.

[0198] Step 303: Perform collaborative detection.

[0199] Simulated attack scenario: Malicious code (disguised as PLC control instructions) is injected into an industrial control terminal. The terminal's security agent collects data in real time on abnormal process startup (non-whitelisted process) and data transmission traffic exceeding the threshold (1GB of data transmitted outward within 5 minutes).

[0200] The local sub-model detects a confidence level of 82% (<85%), which is determined to be a suspected unknown threat. A collaborative verification request is generated, which includes encrypted malicious code signatures and traffic characteristics.

[0201] After receiving the request, the central node schedules 10 associated terminals within the same industrial control network to share threat characteristics, and the associated terminals report "no similar characteristics detected".

[0202] Step 304: Execute the defense response.

[0203] The central node, combining local detection results, feedback from associated terminals, and global threat intelligence (matching the malicious code as an industrial control ransomware variant), determined the threat level to be "critical risk".

[0204] Triggering defense actions: The industrial control terminal locally blocks the malicious process and cuts off the communication link with the external network; 10 associated terminals perform isolation operations and close the 502 port communication with the terminal; Send SMS and email alarms to the administrator and simultaneously trigger the expert emergency response channel;

[0205] The defense response took a total of 0.8 seconds, successfully preventing the spread of malicious code to other terminals.

[0206] Step 305: Iterative optimization of the model.

[0207] Each terminal records the threat characteristics, defense measures, and execution effects of this attack, which are then used as new training samples.

[0208] The model iteration is automatically triggered after 24 hours. The central node aggregates the updated model parameters and generates a new version of the global model. This version of the model has a detection confidence of 96% for this type of ransomware variant.

[0209] Model performance is evaluated weekly, and an emergency iteration is triggered when the detection accuracy drops to 94.5% to ensure model performance stability.

[0210] The technical effects of the present invention are verified through the above implementation process:

[0211] Data privacy protection: No local data was uploaded to the terminal, and model parameters were encrypted throughout the transmission process, so no data leakage occurred;

[0212] Threat detection capability: 97.2% accuracy in detecting known threats and 83.5% accuracy in identifying unknown threats, a 35% improvement over traditional solutions;

[0213] Response efficiency: Average response latency is 0.7 seconds, a reduction of 83.3% compared to the traditional solution (4.2 seconds);

[0214] Compatibility: All 100 heterogeneous terminals were successfully deployed and are running normally, with no resource overload (peak CPU utilization of edge terminals was 28%, and peak memory usage was 1.2GB).

[0215] Example 4

[0216] Based on the same technical concept as embodiments 1-3 above, this embodiment of the invention provides a multi-terminal federated learning security agent collaborative defense system for implementing the methods described in embodiments 1-3 and any of their implementations. The system includes different types of terminal nodes and federated learning center nodes, wherein...

[0217] The terminal node is used to initialize and configure the security agent and complete the registration of federated learning participation permissions; train the local threat detection sub-model through the local security agent and upload the model parameters to the federated learning center node; use the local threat detection sub-model to perform local security detection on local data, and if the detection result shows that the risk level meets the preset conditions, send a collaborative verification request carrying threat characteristics to the federated learning center node.

[0218] The Federated Learning Center node is used to aggregate and generate a global collaborative defense model based on the model parameters of each terminal node using a weighted average algorithm; it receives collaborative verification requests carrying threat features from terminal nodes, matches related terminals with threat features based on global threat intelligence, sends threat features to related terminals, and receives the verification results of the threat features from related terminals; and it performs threat processing based on the local security detection results of terminal nodes, the verification results of the threat features from related terminals, and global threat intelligence.

[0219] In one implementation:

[0220] Deploy a secure intelligent agent module on each terminal node. The secure intelligent agent module includes:

[0221] Threat detection unit: Used to collect terminal operation data, perform static and dynamic detection, and output detection results;

[0222] Local model training unit: Trains and updates the local threat detection sub-model based on local data, and outputs updated model parameter values;

[0223] Encrypted communication unit: Homomorphic encryption + differential privacy technology is used to achieve encrypted transmission of data and model parameters;

[0224] Local defense execution unit: Performs local blocking, logging, and other defense operations based on the threat level;

[0225] Access Control Unit: Responsible for the registration and management of access rights for federated learning.

[0226] Federated learning center nodes can be deployed on cloud servers, including:

[0227] Model aggregation unit: Receives updated parameter values ​​from each terminal model and aggregates them using an improved FedAvg algorithm to generate a global collaborative defense model;

[0228] The collaborative scheduling unit receives collaborative verification requests and schedules associated terminal security agents to perform feature sharing and collaborative detection.

[0229] Threat Level Determination Unit: Combines multi-source data to determine the threat level and outputs defense strategy instructions;

[0230] Global Threat Intelligence Unit: Stores and updates the threat signature database to support detection and assessment.

[0231] The encrypted communication unit connects the multi-terminal secure intelligent agent module and the federated learning center node module. It uses Paillier homomorphic encryption algorithm and differential privacy technology to ensure privacy and security during data transmission, with a transmission delay of ≤1 second.

[0232] It can also set up hierarchical defense execution units that communicate with multi-terminal security intelligence modules and federated learning center node modules to perform defense operations such as logging, local blocking, cross-terminal isolation, and manual alerts according to the threat level, and supports custom configuration of defense strategies.

[0233] In one implementation, the verification result of the associated terminal on the threat feature includes the degree of correlation between the threat feature and the known risks of the associated terminal, and the global threat intelligence includes a set of threat information pre-received by the federated learning center node from each terminal node.

[0234] The Federated Learning Center node includes a threat level determination unit, which is used to classify and handle threats: for low-risk threats, logs are recorded and the global threat intelligence database is updated; for high-risk threats, at least one of the following actions is performed: triggering local terminal blocking and synchronizing threat information to associated terminals; performing cross-terminal isolation, synchronizing global threat intelligence, and sending alarm information to the administrator; or triggering the manual emergency response channel.

[0235] In one implementation, different types of terminal nodes include edge terminals, cloud terminals, or industrial control terminals;

[0236] Each terminal node may include: a threat detection module, a local model training module, an encrypted communication module, a local defense execution module, and a permission management module.

[0237] The encrypted communication module can be used to encrypt communication between terminal nodes and federated learning center nodes, including: first adding dynamic noise to parameters using a differential privacy algorithm, and then completing transmission encryption using Paillier homomorphic encryption.

[0238] The permission management module can be used to: manage the access and registration permissions of terminal nodes, the parameter upload permissions during the local model training phase, the feature sharing permissions during the collaborative detection phase, and the cross-terminal operation permissions during the defense and response phase.

[0239] The permission management module can also be used to dynamically configure the proportion of local model training resources based on preset resource proportion thresholds, business load status, and federated learning task priorities.

[0240] In one implementation, the federated learning center node is used for:

[0241] Receive model parameters from each terminal node, and verify whether the terminal node that uploaded the model parameters has completed the registration of federated learning participation rights, whether it has the qualification to upload parameters, and whether the communication link is encrypted;

[0242] If the terminal has completed the registration of federated learning participation permissions, has the qualification to upload parameters, and the communication link is encrypted, then the model parameters are verified, and a global collaborative defense model is generated based on the verified model parameters.

[0243] In one implementation, the terminal's threat detection module is used to: identify static threats by matching terminal behavior characteristics based on a preset security baseline; wherein, static threats include one or more of the following: abnormal process startup, illegal port access, and data transmission traffic exceeding a threshold.

[0244] The preset security baseline includes:

[0245] Non-core baseline rules are used to identify low-risk threats;

[0246] The core baseline rules are used to determine whether a threat is of medium or high risk.

[0247] The terminal's encrypted communication module is used to: de-identify and encrypt threat features; and send a collaborative verification request carrying the de-identified and encrypted threat features to the federated learning center node.

[0248] This invention employs a federated learning model, where terminal data is trained locally, and only encrypted model parameters are uploaded. This blocks data leakage paths at both the transmission and training stages, fully complying with relevant data privacy protection regulations. By scheduling multiple terminal agents to share threat characteristics through the federated learning center node, a two-layer defense of "local detection + global collaboration" is achieved, improving the identification rate of advanced threats spreading across terminals and eliminating the defense blind spots of traditional solutions.

[0249] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Those skilled in the art should understand that modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A federated learning-based collaborative defense method for secure intelligent agents across multiple terminals, characterized in that, include: Initialize and configure the security agent for different types of terminal nodes, and complete the registration of federated learning participation permissions for each terminal node; Each terminal node trains a local threat detection sub-model through a local security agent and uploads the model parameters to the federal learning center node. The Federated Learning Center nodes aggregate and generate a global collaborative defense model based on the model parameters of each terminal node using a weighted average algorithm. Each terminal node uses a local threat detection sub-model and a global collaborative defense model to perform dual verification of local risks; including: each terminal node first uses the local threat detection sub-model to perform local security detection on local data, and if the detection result shows that the risk level meets the preset conditions, it sends a collaborative verification request carrying threat characteristics to the federated learning center node. The Federated Learning Center node matches threat characteristics with associated terminals based on global threat intelligence, sends threat characteristics to associated terminals, and receives the verification results of the threat characteristics from associated terminals. The Federated Learning Center node then performs threat processing based on local security detection results, the verification results of threat characteristics from associated terminals, and global threat intelligence.

2. The multi-terminal federated learning security agent collaborative defense method according to claim 1, characterized in that, in, The verification results of the associated terminals on threat characteristics include the degree of correlation between the threat characteristics and the known risks of the associated terminals. Global threat intelligence includes a collection of threat information pre-received by the federated learning center nodes from each terminal node. The Federated Learning Center node performs threat handling based on local security detection results, the verification results of threat signatures by associated terminals, and global threat intelligence, including: Threats are categorized and handled accordingly: For low-risk threats, logs are recorded and the global threat intelligence database is updated. For high-risk threats, perform at least one of the following actions: trigger local blocking on the terminal and synchronize threat information to associated terminals; perform cross-terminal isolation, synchronize global threat intelligence, and send alarm information to the administrator; trigger the manual emergency response channel.

3. The multi-terminal federated learning security agent collaborative defense method according to claim 1, characterized in that, Different types of terminal nodes include edge terminals, cloud terminals, or industrial control terminals; The initial configuration of the security agent for each terminal node includes configuring the security agent to integrate a threat detection module, a local model training module, an encrypted communication module, a local defense execution module, and a permission management module.

4. The multi-terminal federated learning security agent collaborative defense method according to claim 3, characterized in that, Before uploading model parameters to the Federation Learning Center node, the following steps are also required: Each terminal node encrypts the communication between itself and the federated learning center node through an encrypted communication module, including: first adding dynamic noise to the parameters using a differential privacy algorithm, and then completing the transmission encryption using Paillier homomorphic encryption.

5. The multi-terminal federated learning security agent cooperative defense method according to claim 3, characterized in that, The registration of federated learning participation permissions for each terminal node includes: managing the terminal node's admission registration permissions through the permission management module, parameter upload permissions during the local model training phase, feature sharing permissions during the collaborative detection phase, and cross-terminal operation permissions during the defense and response phase. Before each terminal node trains the local threat detection sub-model through the local security agent, it also includes: dynamically configuring the local model training resource ratio based on the preset resource ratio threshold, business load status, and federated learning task priority through the permission management module.

6. The multi-terminal federated learning security agent collaborative defense method according to claim 1, characterized in that, The federated learning center nodes aggregate and generate a global collaborative defense model based on the model parameters of each terminal node using a weighted average algorithm, including: The Federated Learning Center node receives model parameters from each terminal node and verifies whether the terminal node uploading the model parameters has completed the registration for Federated Learning participation permissions, whether it has the qualification to upload parameters, and whether the communication link is encrypted. If the terminal has completed the registration of federated learning participation permissions, has the qualification to upload parameters, and the communication link is encrypted, then the model parameters are verified, and a global collaborative defense model is generated based on the verified model parameters.

7. The multi-terminal federated learning security agent cooperative defense method according to claim 1, characterized in that, Each terminal node first uses the local threat detection sub-model to perform local security checks on local data, including: Static threats are identified by matching terminal behavior characteristics based on a preset security baseline; among them, static threats include one or more of the following: abnormal process startup, illegal port access, and data transmission traffic exceeding the threshold.

8. The multi-terminal federated learning security agent cooperative defense method according to claim 7, characterized in that, Preset security baselines include: Non-core baseline rules are used to identify low-risk threats; The core baseline rules are used to determine whether a threat is of medium or high risk.

9. The multi-terminal federated learning security agent collaborative defense method according to claim 1, characterized in that, Sending a collaborative verification request carrying threat signatures to the federated learning center node includes: Threat signatures are anonymized and encrypted; Send a collaborative verification request carrying the threat signature after de-identification and encryption to the federated learning center node.

10. A multi-terminal federated learning security intelligent agent collaborative defense system, used to implement the method according to any one of claims 1-9, characterized in that, This includes different types of terminal nodes and federated learning center nodes, among which The terminal node is used to initialize and configure the security agent and complete the registration of federated learning participation permissions; train the local threat detection sub-model through the local security agent and upload the model parameters to the federated learning center node; use the local threat detection sub-model to perform local security detection on local data, and if the detection result shows that the risk level meets the preset conditions, send a collaborative verification request carrying threat characteristics to the federated learning center node. The Federated Learning Center node is used to aggregate and generate a global collaborative defense model based on the model parameters of each terminal node using a weighted average algorithm; it receives collaborative verification requests carrying threat features from terminal nodes, matches related terminals with threat features based on global threat intelligence, sends threat features to related terminals, and receives the verification results of the threat features from related terminals; and it performs threat processing based on the local security detection results of terminal nodes, the verification results of the threat features from related terminals, and global threat intelligence.

Images