An artificial intelligence-based internet of things security management method and system

Abstract

The application discloses an Internet of Things security management method and system based on artificial intelligence. The method comprises the following steps: real-time anomaly identification is carried out by an identification intelligent agent deployed at an Internet of Things edge node using a resource-aware double-model cascaded dynamic switching architecture; a spatiotemporal risk conduction field is constructed by a regional collaborative layer decision intelligent agent to cross-layerly research and judge the risk situation of the device layer, the network layer and the application layer; L1 to L3 hierarchical security scheduling is carried out by a cloud-end response intelligent agent using a Nash equilibrium multi-objective game scheduling strategy; real-time explainability labeling of the causal relationship of disposal decision is realized by an online incremental causal diagram module; and model parameter federal aggregation is carried out between the identification intelligent agents by a federal transfer learning module to accelerate new node deployment. According to the application, the model reasoning delay is reduced to below 15 milliseconds, the false alarm rate is reduced to below 2.3%, the resource utilization rate is increased to 82%, and the cross-layer risk diffusion rate is reduced to 12%.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) security technology, and specifically to an IoT security management method and system based on artificial intelligence. Background Technology

[0002] With the rapid development of IoT technology, the number of terminal devices connected in application scenarios such as industrial IoT, smart homes, and smart cities is growing exponentially. It is estimated that the number of IoT device connections worldwide will exceed 25 billion by 2025. The massive influx of heterogeneous devices has dramatically expanded the attack surface of IoT systems, posing a severe challenge to traditional security protection mechanisms.

[0003] Existing IoT security management technologies suffer from the following shortcomings: First, traditional cloud-based centralized detection solutions upload all data to the cloud for analysis, resulting in end-to-end detection latency typically exceeding 3000 milliseconds, which fails to meet the millisecond-level response requirements of industrial control scenarios. Second, existing solutions lack cross-layer risk transmission mechanisms, leading to fragmented security postures at the device, network, and application layers, with a cross-layer propagation rate of 15% to 20% for security incidents. Third, existing security scheduling strategies employ fixed threshold rules or simple priority ranking, failing to achieve a dynamic balance between risk level and resource cost, resulting in resource utilization rates generally below 35%. Fourth, existing solutions lack interpretability, with security response decisions lacking causal reasoning support, leading to false alarm rates typically exceeding 15%, making it difficult for security operations personnel to quickly pinpoint the root cause. Fifth, the cold start time for newly deployed edge nodes is excessively long, typically requiring more than two hours of data accumulation to reach a usable state.

[0004] Existing technologies include edge computing-based IoT security detection solutions that employ a single model architecture, failing to resolve the conflict between model lightweighting and detection accuracy; and existing cross-layer security analysis systems that lack dynamic risk transmission mechanisms and game-theoretic scheduling strategies. Overall, existing technologies exhibit significant technological gaps in five dimensions: detection latency, cross-layer risk linkage, resource scheduling efficiency, interpretability, and rapid deployment. Summary of the Invention

[0005] To address the technical problems of high detection latency, cross-layer risk diffusion, low resource scheduling efficiency, insufficient interpretability, and difficulty in rapid deployment in existing technologies, this invention provides an artificial intelligence-based Internet of Things (IoT) security management method and system.

[0006] The technical solution of this invention is as follows: An IoT security management method based on artificial intelligence, comprising the following steps: S1, an identification intelligent agent deployed at the edge node of the IoT performs real-time collection and anomaly identification of the behavior data of IoT terminal devices. The identification intelligent agent adopts a resource-aware dual-model cascaded dynamic switching architecture. The architecture includes a first-level one-dimensional convolutional neural network initial screening model and a second-level structured pruned bidirectional long short-term memory network confirmation model. When the screening result of the first-level model is a suspected anomaly, the second-level model is triggered for precise classification. The pruning rate of the second-level model is determined by Pareto optimality solution based on the real-time CPU utilization, available memory, and bandwidth margin of the edge node; S2, a decision intelligent agent deployed at the regional collaboration layer constructs a spatiotemporal risk transmission field. The spatiotemporal risk transmission field includes a three-layer risk situation map of the device layer, network layer, and application layer, and adopts a cross-... The layer risk entropy and exponential time-varying decay weight are used to calculate the risk transmission between layers. When the risk situation in any layer exceeds the dynamic threshold, the preventive response of the adjacent layer is automatically triggered. S3: The response agent deployed in the cloud adopts a Nash equilibrium multi-objective game scheduling strategy for hierarchical security scheduling. The risk level and resource cost are modeled as a non-cooperative game and the Nash equilibrium is solved as the optimal scheduling strategy. The scheduling instructions are output as L1 local disposal, L2 regional collaboration, or L3 cloud global. S4: The online incremental causal graph module performs interpretability annotation on the disposal decision. The dynamic causal graph is maintained by combining Granger causality test and counterfactual reasoning. The causal confidence propagation algorithm realizes real-time incremental annotation of causal relationships. S5: The federated transfer learning module performs federated aggregation of model parameters among the identification agents. Newly deployed edge nodes obtain pre-trained model parameters from the regional collaboration layer and fine-tune and adapt them locally.

[0007] In step S1, the inference delay of the initial screening model of the first-level one-dimensional convolutional neural network does not exceed 3 milliseconds. The confirmation model of the second-level bidirectional long short-term memory network after structured pruning adopts a knowledge distillation compensation strategy, and the classification accuracy loss is controlled within 1 percentage point when the pruning rate is 0.4. In step S2, the time-varying decay weight of cross-layer risk transmission decreases exponentially with the increase of time interval. When the risk entropy of the device layer exceeds the dynamic threshold, the risk transmission of that layer is blocked by an adaptive gating function. In step S3, the utility function is defined as the risk aversion benefit minus the weighted sum of resource cost and response delay. Each player uses the regret matching iterative algorithm to solve the mixed strategy Nash equilibrium. In step S4, the weight of each edge of the causal graph is equal to the product of the Granger causality score and the counterfactual reasoning score multiplied by the time decay factor.

[0008] This invention also provides an AI-based IoT security management system, comprising: an identification agent module deployed at IoT edge nodes, employing a resource-aware dual-model cascaded dynamic switching architecture to collect and identify anomalies in real time from IoT terminal device behavior data; a decision-making agent module deployed at the regional collaboration layer, constructing a spatiotemporal risk transmission field and combining it with three layers of risk situation data for comprehensive analysis; a response agent module deployed in the cloud, employing a Nash equilibrium multi-objective game scheduling strategy for hierarchical security scheduling; an online incremental causal graph module, providing interpretable annotations for handling decisions; and a federated transfer learning module, performing federated aggregation of model parameters among the identification agents. A dynamic switching controller is provided between the first-level and second-level models of the identification agent module. The system also includes a situation visualization module, providing a device topology view, a risk heatmap, a handling causal link diagram, and a historical situation playback panel.

[0009] The present invention also provides an artificial intelligence-based Internet of Things (IoT) security management device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; the memory stores a computer program executable by the at least one processor, wherein the computer program, when executed by the at least one processor, implements steps S1 to S5 of the aforementioned artificial intelligence-based IoT security management method. The device includes three implementation forms: edge gateway, regional server, and cloud server.

[0010] Compared with existing technologies, this invention has the following advantages: (1) Through the RADMC dual-model cascade dynamic switching architecture, the model inference latency is no more than 15 milliseconds and the end-to-end detection latency is no more than 80 milliseconds, which is two orders of magnitude lower than the traditional cloud-based centralized detection scheme. (2) Through the cross-layer risk entropy and exponential time-varying decay weight of the TSCRF spatiotemporal risk transmission field, the cross-layer diffusion rate of security events is reduced from 15% to 20% to about 12%. (3) Through the NEMS Nash equilibrium multi-objective game scheduling strategy, the resource utilization rate is increased from 32% to 82%, and the response time rate of high-risk events reaches 99.5%. (4) Through the causal confidence propagation algorithm of the OICG online incremental causal graph module, the causal labeling latency is controlled within 30 milliseconds, and the system false alarm rate is reduced from 15% to below 2.3%. (5) Through the multi-agent distributed architecture and the MQTT-based message queue asynchronous communication mechanism, it supports the concurrent access and elastic expansion of hundreds of thousands of devices, and the system availability reaches 99.9%. (6) Through the federated transfer learning module, the cold start time of the model for newly deployed edge nodes is reduced from about 2 hours to about 15 minutes. The cross-dimensional synergy effect improves both detection accuracy and deployment efficiency.

[0011] The disadvantages of this invention are: the initial deployment of the system requires a certain amount of hardware investment, and the edge nodes need to have at least 4GB of memory and an ARM Cortex-A72-level processor; in addition, the Nash equilibrium solution has a computational overhead of about 50 milliseconds in extremely high concurrency scenarios, and further optimization is needed for ultra-low latency scenarios. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention; Figure 2 This is a schematic diagram of the RADMC dual-model cascaded dynamic switching architecture for identifying intelligent agents in this invention; Figure 3 This is a schematic diagram of the three-layer risk situation map of the TSCRF spatiotemporal risk transmission field of the present invention; Figure 4 This is a flowchart illustrating the NEMS Nash Equilibrium Multi-Objective Game Scheduling Strategy of the present invention. Figure 5 This is a schematic diagram illustrating the game payoff matrix and Nash equilibrium solution of this invention; Figure 6 This is a schematic diagram of the causal graph structure of the OICG online incremental causal graph module of the present invention; Figure 7 This is a schematic diagram of entity relationships in the unified situation map of the present invention; Figure 8 This is a schematic diagram of the closed-loop timing of the entire process of this invention; Figure 9 This is a schematic diagram of the initial screening model structure of the 1D-CNN of the present invention; Figure 10 This is a schematic diagram of the DP-BiLSTM verification model structure of the present invention; Figure 11 This is a schematic diagram illustrating the calculation of cross-layer risk transmission weights in this invention; Figure 12 This is a schematic diagram of the adaptive conduction gating of the present invention; Figure 13 This is a schematic diagram illustrating the causal conflict detection and resolution method of the present invention; Figure 14 This is a schematic diagram of the federated transfer learning process of the present invention; Figure 15 This is a schematic diagram of the situation visualization interface of the present invention. Detailed Implementation

[0013] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the invention.

[0014] Example 1: Basic Principles This embodiment illustrates the basic principles of the five core technical features of the present invention. See also Figure 1 RADMC (Resource Aware Dual-Model Cascaded Dynamic Switching) deploys a 1D-CNN screening model and a DP-BiLSTM confirmation model at the edge node, and dynamically switches between the two models based on the real-time resource status, balancing detection speed and accuracy.

[0015] The 1D-CNN screening model extracts one-dimensional convolutional features from the temporal traffic of devices, with a single inference latency of no more than 3ms, and outputs the anomaly detection probability P_detect. Its convolution operation and activation mapping are defined as shown in Equation (1), where W_conv is the convolution kernel parameter and b_conv is the bias term.

[0016] (1) In equation (1), P_detect is the anomaly detection probability, taking values ​​[0,1]. The larger the value, the more likely it is to be attack traffic; x is the normalized temporal traffic window; W_conv is the one-dimensional convolution kernel; and b_conv is the bias. Sigmoid maps the convolution output to the probability interval, achieving lightweight initial screening.

[0017] The DP-BiLSTM confirmation model is triggered when P_detect exceeds a threshold. This model performs structured pruning on a bidirectional LSTM with a pruning rate of 0.4, and compensates for the accuracy loss caused by pruning through knowledge distillation using a teacher network, thereby reducing the inference load while maintaining confirmation accuracy.

[0018] The pruning rate is not fixed, but is based on three types of resource constraints: edge node CPU utilization, available memory, and uplink bandwidth. It solves a multi-objective optimization problem through Pareto optimality, and achieves a non-dominated optimal trade-off between detection accuracy and inference latency, thus realizing dynamic pruning with resource awareness.

[0019] RADMC dynamically switches to a comprehensive score that balances detection speed, resource availability, and confidence margin: when resources are scarce, lightweight 1D-CNN is prioritized, and when resources are plentiful, it is upgraded to DP-BiLSTM for confirmation, thus achieving an adaptive match between detection capability and resource constraints.

[0020] See Figure 3 The TSCRF (Spatiotemporal Risk Transmission Field) constructs a three-layer risk situation map of the device layer, network layer, and application layer. Risks at each layer are transmitted between layers through cross-layer risk entropy and exponential time-varying decay weights, forming a global risk field that evolves over time.

[0021] The cross-layer risk entropy H_risk is a weighted sum of the risk entropies H_layer of each layer, superimposed with the norm of the cross-layer risk increment DeltaRisk_cross, and then time-varyingly weighted by the exponentially decaying weight W_decay(t). W_decay(t) decays exponentially over time, giving higher weight to recent risk events. The risk field is defined as shown in equation (2).

[0022] (2) In equation (2), H_risk(t) is the cross-layer risk entropy at time t; W_decay(t) is the exponential time-varying decay weight, lambda_decay is the decay coefficient; H_layer(l) is the risk entropy of the l-th layer; Delta Risk_cross(t) is the cross-layer risk transmission increment, and the norm is used to characterize the transmission intensity.

[0023] TSCRF focuses the risk field on recent events by attenuating weights and characterizes the inter-layer transmission strength by cross-layer risk increment norm, thus characterizing the evolution of risk situation in both spatiotemporal dimensions and providing quantitative input for NEMS hierarchical scheduling.

[0024] See Figure 6 NEMS (Nash Equilibrium Multi-Objective Game Scheduling) treats edge nodes as game participants and hierarchically schedules response resources across three levels: L1 local processing, L2 regional collaboration, and L3 cloud global scheduling. It solves the optimal allocation of multi-objectives through Nash equilibrium.

[0025] L1 local handling involves edge nodes handling low-risk events locally; L2 regional collaboration aggregates resources from adjacent nodes at the regional level to handle medium-risk events; L3 cloud-based global handling provides overall coordination for high-risk cross-domain events. The three levels of scheduling automatically upgrade or downgrade based on event severity and resource availability.

[0026] The game utility function integrates three objectives: risk reduction payoff, resource cost, and response delay, with weights satisfying normalization constraints. Under Nash equilibrium, no single participant can improve their utility by unilaterally changing their strategy. The multi-objective utility is defined as shown in equation (3).

[0027] (3) In equation (3), U_nash is the Nash equilibrium comprehensive utility; w_risk, w_cost, and w_delay are the risk, cost, and delay weights, respectively, and the sum of the three is 1; Risk_i is the risk level of node i; Cost_i and Delay_i are the normalized resource cost and response delay.

[0028] NEMS uses a regret matching mechanism to iteratively update the strategy probability. The larger the positive regret value, the higher the probability that the strategy will be selected. After a finite number of iterations, it converges to the Nash equilibrium approximate solution, ensuring the global optimality of regional resource allocation.

[0029] See Figure 9 OICG (Online Incremental Causal Graph) uses a directed acyclic graph to depict causal relationships between devices. The online incremental update mechanism dynamically adds or removes causal edges as new events arrive, avoiding offline retraining and adapting to frequent changes in IoT topologies.

[0030] The causal edge weights are fused with the Granger causality test score S_granger and the counterfactual reasoning score S_cf, and the timeliness is reflected by an exponential decay term. The Granger test quantifies the improvement in time series prediction, and the counterfactual reasoning assesses the difference in results after intervention. The causal confidence propagation is shown in Equation (4).

[0031] (4) In equation (4), Conf_j is the causal confidence score of node j; S_granger(i,j) is the Granger causal score of i to j; S_cf(i,j) is the counterfactual reasoning score; rate_decay is the decay rate; and Deltat is the event interval. The confidence score is propagated forward along the causal graph.

[0032] Causal confidence is propagated forward along the directed acyclic graph. The confidence of the root node is determined by its own observation, while the confidence of the non-root node is obtained by recursively summing the weighted input edges, thereby realizing the causal attribution and source tracing of risks.

[0033] See Figure 12 Federated transfer learning collaboratively trains detection models across multiple regions without exchanging raw data. Edge nodes train locally and only upload model parameters, which are then federated by the aggregation server using FedAvg or FedProx, protecting data privacy.

[0034] FedAvg uses a weighted average of the node models based on the amount of data; FedProx introduces a near-end term on top of FedAvg to constrain the degree to which the node models deviate from the global model, thereby enhancing the convergence stability under heterogeneous nodes. The global aggregation update is shown in Equation (5).

[0035] (5) In equation (5), w_global is the global model parameter; n_i is the data volume of node i, N is the total data volume; w_i is the local model of node i; mu is the FedProx near-end coefficient, and when mu is 0, it degenerates into FedAvg; the norm term constrains the deviation between the local model and the global model.

[0036] Federated transfer learning combines a pre-trained model in the source domain with fine-tuning in the target domain. It measures the difference in distribution between the source and target domains through transfer loss, reduces the cold start time of the target domain, and enables newly deployed areas to quickly acquire detection capabilities.

[0037] In summary, RADMC is responsible for rapid and accurate detection, TSCRF characterizes the spatiotemporal risk situation, NEMS achieves optimal resource scheduling, OICG provides causal interpretability, and federated transfer learning ensures cross-domain collaboration and privacy. These five features work together to form a complete IoT security management system.

[0038] Example 2: Basic Application This embodiment illustrates the basic application of the present invention in three typical scenarios. All three scenarios adopt a three-layer architecture: edge-region-cloud, edge nodes run RADMC detection and L1 handling, the region layer executes NEMS game scheduling and L2 coordination, and the cloud is responsible for L3 global response and federated aggregation.

[0039] See Figure 2 In an industrial IoT scenario, this system is deployed in an automobile manufacturing workshop, covering devices such as PLC controllers, industrial gateways, and robotic arm controllers, to perform real-time safety monitoring of production line control commands and equipment status.

[0040] The workshop is equipped with a Siemens S7-1200 PLC (model 6ES7212), a Moxa MGate 5101 industrial gateway, and a KUKAKR CYBERTECH robotic arm controller. Flow data is collected via the OPC-UA protocol with a sampling period of 100ms, including register read / write messages.

[0041] The OPC-UA message is segmented by sliding window, and 23-dimensional features such as register address distribution, write frequency and timing interval are extracted. After Min-Max normalization, it is input into the 1D-CNN screening model, and the DP-BiLSTM confirms that the model performs secondary classification on high-risk samples.

[0042] The 1D-CNN screening model detects abnormal register writes, while the DP-BiLSTM model confirms the identification of injection and replay attacks. In one injection attack, P_detect reached 0.92, and after confirmation, the classification confidence was 0.97. TSCRF transmits device-layer risks to the network and application layers via cross-layer entropy.

[0043] NEMS classified the incident as an L2 area collaborative response, scheduling adjacent production line gateways to isolate controlled PLCs, with a response latency of less than 200ms. OICG source tracing revealed a causal chain originating from an intrusion into the engineering workstation, and FedProx shared this attack pattern with other workshops via federated learning.

[0044] After deployment, the number of unauthorized register write events in the workshop decreased by 92%, the average processing latency was 180ms, and the regional collaborative scheduling ensured that adjacent production lines were not affected, verifying the effectiveness of this system in industrial high real-time scenarios.

[0045] In a smart home scenario, this system is deployed on a home gateway, covering smart door locks, thermostats, and cameras. Devices connect via Zigbee and Wi-Fi, and the gateway collects device heartbeats and control command messages for security analysis.

[0046] Typical devices include the Xiaomi Smart Door Lock (model XM-Lock Pro), Nest thermostat, and Hikvision C200 camera. The data format is an MQTT message encapsulated in JSON format, with a topic level of home / room / device, and the payload includes command code and timestamp.

[0047] The MQTT message extracts 18-dimensional features, including topic level, payload length, release frequency, and QoS level. After one-hot encoding and normalization, these features are input into the screening model to confirm that the model performs secondary discrimination on abnormal command sequences.

[0048] A 1D-CNN was used to screen for abnormal door lock unlocking sequences, and a DP-BiLSTM was used to confirm brute-force and man-in-the-middle attacks. In one abnormal unlocking instance, P_detect was 0.88, confirming the model's identification as a replay attack with a confidence level of 0.95.

[0049] TSCRF propagates risks from the lock device layer to the network layer. NEMS is designated as L1 local handling, where the gateway disconnects the lock's network connection locally and issues an alarm. The local handling latency is less than 50ms. OICG traces the causal path of abnormal unlocking originating from commands issued by a malicious app.

[0050] After deployment, the accuracy of abnormal unlocking alarms increased to 98.5%, and the false alarm rate decreased to 1.5%. Federated learning shared this model with other home gateways in the community to enhance the group's defense capabilities.

[0051] In smart city scenarios, this system is deployed on a city-wide IoT sensing platform, covering traffic signals, environmental monitoring stations, and smart streetlights. Data from the devices is transmitted via NB-IoT and 5G backhaul, and then aggregated to a district-level platform.

[0052] Typical equipment includes Hisense traffic signal controllers (model Hi-SCATS), concentrated light environmental monitoring stations, and Huawei smart streetlights. The data format is CoAP-packaged sensor readings, including fields such as PM2.5 and traffic flow, with sampling periods ranging from 1 to 30 seconds.

[0053] CoAP readings extract 21-dimensional features, including sensor channels, sampling intervals, numerical gradients, and missing rates. After standardization, these features are input into a screening model to confirm that the model can perform secondary identification of abnormal timing and unauthorized control commands.

[0054] A 1D-CNN was used to screen for anomalous timing instructions in the signal generator, and a DP-BiLSTM was used to confirm a signal hijacking attack. In one instance of anomalous timing in the signal generator, P_detect was 0.90, confirming the model's identification of instruction injection with a confidence level of 0.96.

[0055] TSCRF transmits the risks at the signal equipment layer to the traffic application layer via cross-layer entropy. NEMS is designated as L3 cloud-based global handling, with the city platform issuing emergency timing plans and linking with the traffic police system.

[0056] The OICG source hijacking originated from a causal chain of leaked credentials from a district-level platform. Federated transfer learning distributed this threat intelligence to all sensing nodes in the city, enabling cross-regional collaborative defense.

[0057] After deployment, the average detection time of signal hijacking events was reduced to 3 seconds, and the cloud-based global emergency timing reduced the intersection congestion index by 15%, verifying the scalability of this system in large-scale urban scenarios.

[0058] Real-world testing shows that the detection latency is 2.1ms and the false alarm rate is 1.2% in industrial scenarios; 1.8ms and 1.5% in home scenarios; and 2.4ms and 1.1% in urban scenarios. All three scenarios outperform the traditional single-model solution.

[0059] Three scenarios verified the adaptability of the invention to heterogeneous protocols, heterogeneous devices, and heterogeneous deployment forms, and the five core features can operate effectively in IoT systems of different scales.

[0060] Example 3: Common Replacements This embodiment presents commonly used equivalent replacement schemes for key algorithm modules among the five core features. The replacement follows the principles of functional equivalence and interface compatibility: the replaced module must maintain consistent input and output semantics and not change the three-layer architecture and hierarchical scheduling process of edge-region-cloud.

[0061] In resource-constrained edge nodes, GRU gated recurrent units can be used to replace 1D-CNN as the screening model. GRU models temporal dependencies with a gating mechanism, has fewer parameters, and the latency of a single inference can be reduced to less than 1.5ms, making it suitable for microcontrollers with computing power below 0.5 TFLOPS.

[0062] The GRU screening model replaces the convolution operation in equation (1) with gated recursion, outputting the anomaly detection probability P_detect. The interface is fully compatible with 1D-CNN, and the subsequent DP-BiLSTM confirmation model does not require modification. Cascaded confirmation is still triggered when P_detect exceeds the threshold.

[0063] GRU replacement is suitable for low-power wide-area nodes (such as NB-IoT terminals), significantly reducing energy consumption while the detection accuracy loss is no more than 2%; 1D-CNN is still preferred for high-computing-power nodes to obtain higher throughput.

[0064] In long sequence and multivariate scenarios, the Transformer encoder can be used to replace DP-BiLSTM as the validation model. The Transformer captures global temporal dependencies with a self-attention mechanism and has better ability to identify long-span attack patterns than bidirectional LSTM.

[0065] The Transformer confirmation model replaces sequence modeling in the confirmation phase with multi-head self-attention. After pruning and distillation, the number of parameters is comparable to DP-BiLSTM, and the output classification confidence is consistent with the original model. The cascading switching logic remains unchanged.

[0066] Transformer replacement is suitable for scenarios with sequence lengths exceeding 512 time steps and variable dimensions exceeding 32; in short sequence low-dimensional scenarios, DP-BiLSTM is still preferred due to its large attention overhead.

[0067] The pruning rate can be solved using the NSGA-II non-dominated sorting genetic algorithm instead of the weighted scalar Pareto solution. NSGA-II directly searches for the non-dominated solution set in the multi-objective space, avoiding the human bias in weight setting.

[0068] NSGA-II aims to reduce detection accuracy, inference latency, and resource consumption. It outputs a Pareto front through population evolution, and then the decision module selects a compromise solution as the pruning rate based on the current CPU, memory, and bandwidth, which is consistent with the output interface of the original scheme.

[0069] The NSGA-II replacement is suitable for scenarios with many objective dimensions and complex constraints; however, when there are fewer objectives, weighted scalar quantization is faster and remains the preferred choice.

[0070] FedAvg can be replaced by FedProx for federated aggregation. FedProx suppresses model drift caused by differences in the distribution of data from heterogeneous nodes by constraining the deviation between the local model and the global model through proximal terms.

[0071] FedProx sets the near-end coefficient mu in equation (5) to a non-zero positive value; when the data of each node is homogeneous, mu is set to approach 0, which degenerates into FedAvg. The two aggregation methods can be adaptively switched in the same system according to the degree of node heterogeneity.

[0072] FedProx is suitable for cross-domain federations where there are significant differences in device computing power and data distribution; FedAvg converges faster in homogeneous campus federations and remains the default solution.

[0073] In deterministic alarm scenarios, a step function can be used instead of the sigmoid function as the output activation of the screening model. The step function outputs 0 or 1 with a hard threshold decision, eliminating the ambiguity of the probability output and reducing the complexity of downstream decision-making.

[0074] The step replacement replaces the Sigmoid function in equation (1) with a step function with the threshold theta as the boundary, outputs a binary detection result, and changes the cascade confirmation trigger logic from probability comparison to binary judgment, while the rest of the process remains unchanged.

[0075] Step replacement is suitable for scenarios with strict alarms and no need for probability sorting; in scenarios that require risk scoring sorting, Sigmoid is still preferred to preserve probability information.

[0076] The adaptation of the replacement scheme follows the interface contract: each replaceable module defines a unified input tensor shape and output semantics, and the replacement implementation only needs to satisfy this contract without changing the cascading switching, risk propagation and scheduling logic.

[0077] Compatibility verification is performed automatically in the continuous integration pipeline: after each replacement, regression testing covers four categories of indicators: detection accuracy, inference latency, resource consumption, and cross-layer propagation consistency. If any indicator deteriorates beyond the threshold, integration is blocked.

[0078] Hybrid replacement is also supported: GRU can be used for screening layers, Transformer for confirmation layers, NSGA-II for pruning solutions, and FedProx for aggregation. These can be combined arbitrarily, and the corresponding implementations will be dynamically loaded according to the configuration of this system.

[0079] The variable definitions and symbol conventions of equations (1) to (5) are replaced without changing them, only their internal calculation implementation is changed, ensuring the consistency of the symbols described in the formulas in the specification and the embodiments.

[0080] All the above replacement schemes maintain the synergistic relationship of the five core features: RADMC detection, TSCRF transmission, NEMS scheduling, OICG tracing, and federated collaboration interfaces and data flows can be seamlessly connected, reflecting the modularity and scalability of the present invention.

[0081] Example 4: Common Extensions This embodiment presents four commonly used extension forms of the present invention under different computing power distributions and deployment topologies. All four extensions reuse the five core features, adjusting only the model deployment location, scheduling level, and federation scope, reflecting the flexibility of the architecture.

[0082] The edge-heavy extension fully decentralizes the RADMC dual model and L1 local processing to edge nodes, while the regional layer and cloud only handle aggregation and global monitoring. This approach is suitable for industrial production line scenarios with low latency and high autonomy requirements.

[0083] In edge-heavy deployments, both 1D-CNN and DP-BiLSTM run on edge gateways, with a higher pruning rate to adapt to local computing power; TSCRF builds a device-level risk map locally on edge nodes, while NEMS mainly uses L1 processing, only pushing high-risk events to the region layer.

[0084] Heavy-duty edge scaling enables more than 95% of events to be processed locally in a closed loop with end-to-end latency of less than 50ms, but it has high requirements for the computing power and storage of edge nodes and is suitable for gateway devices with edge AI acceleration.

[0085] The cloud-based heavy-duty extension centrally deploys the detection and verification model and global scheduling in the cloud, while edge nodes are only responsible for data collection and lightweight screening. This approach is suitable for wide-area perception scenarios with massive numbers of weak terminals and insufficient edge computing power.

[0086] In heavy cloud deployments, DP-BiLSTM confirms that the OICG causal graph runs on a cloud GPU cluster, TSCRF aggregates a three-layer risk graph in the cloud, NEMS primarily uses L3 global scheduling, and the edge only executes and issues instructions.

[0087] Cloud-based heavy-duty expansion fully utilizes cloud computing power to improve detection accuracy and causal analysis depth, but end-to-end latency is affected by network backhaul, making it suitable for regulatory scenarios that are not sensitive to latency and have high accuracy requirements.

[0088] Multi-region federated extensions establish a federated aggregation layer across multiple regional platforms. After local training in each region, the model parameters are uploaded and aggregated by the federated aggregation server using FedAvg or FedProx to form a global model shared across regions.

[0089] In a multi-region federation, each region retains its local data and does not leave the domain, only exchanging model parameters; FedProx proximal terms suppress model drift caused by differences in data distribution between regions, and OICG causal graphs trace the path of threat propagation across regions at the federation level.

[0090] Multi-regional federation enables threat intelligence sharing while protecting data privacy across regions, reducing cold start time in new regions by more than 60%, and is suitable for deployments across multiple branches or administrative regions.

[0091] Hybrid cloud expansion deploys the detection and processing of sensitive data in the private cloud, and deploys large-scale model training and federated aggregation in the public cloud, coordinating through secure channels to balance data compliance and computing power elasticity.

[0092] In a hybrid cloud deployment, the private cloud runs RADMC detection and L1 / L2 processing while retaining the original data, while the public cloud undertakes DP-BiLSTM distillation training, federated aggregation, and global OICG analysis. The two exchange model parameters and desensitized features through an encrypted channel.

[0093] Hybrid cloud extensions leverage public cloud computing power while meeting data localization compliance requirements, making them suitable for IoT security management in highly compliant industries such as finance and government.

[0094] Four extensions share the same configuration description: you can switch between extension forms by declaring three parameters: the model deployment location, the default scheduling level, and the federation scope, without modifying the business code.

[0095] The extended configuration can be differentiated according to region: core production lines adopt edge heavy-duty systems, branch offices adopt cloud heavy-duty systems, and cross-regional systems adopt multi-regional federation systems, with different regions coexisting in the same system.

[0096] The extension switching does not affect the applicability of equations (1) to (9): the resource, risk and scheduling variables in each formula have the same semantics in all extension forms, only the range of values ​​and the solution location are different.

[0097] The four extensions mentioned above can be implemented through configuration switching rather than code modification. The modular design of this invention enables it to adapt to diverse deployment needs, from edge autonomy to cross-cloud collaboration.

[0098] Example 5: As a system embedding This embodiment presents a solution for embedding the present invention as a security subsystem into four existing systems: SCADA, firewall, SIEM, and zero-trust architecture. During embedding, this system outputs risk scores, causal chains, and remediation suggestions through standardized interfaces, with the host system executing the final actions.

[0099] When embedded in a SCADA system, this system bypasses and collects control messages from the PLC and RTU, the RADMC detects abnormal commands, the TSCRF transmits risks from the device layer to the monitoring application layer, and the OICG traces the causal source of abnormal commands and sends alarms and interlocking suggestions to the SCADA system.

[0100] The SCADA embedding adopts a read-only bypass mode and does not interfere with the real-time performance of the control loop. When P_embed exceeds the interlocking threshold, the SCADA executes the preset interlocking action. This system only provides decision-making basis to ensure the safety and determinism of the industrial control system.

[0101] When embedded in a firewall, this system performs RADMC detection on incoming traffic, converts the detection probability and TSCRF risk entropy into dynamic access control priorities, and guides the firewall to adjust its filtering policies according to the risk level.

[0102] When embedded in the firewall, NEMS schedules and handles actions across three levels—L1 local dropping, L2 regional rate limiting, and L3 cloud black hole—based on risk levels; OICG provides causal evidence of the attack chain and supports the firewall in generating auditable blocking logs.

[0103] When embedded in the SIEM system, this system is connected as an advanced analysis engine, pushing RADMC detection results, TSCRF risk fields, and OICG causal graphs to SIEM in a standardized event format to enrich its correlation analysis capabilities.

[0104] In SIEM embedding, this system outputs events with causal confidence, enabling SIEM to perform alarm aggregation and noise reduction based on causal chains rather than just rules, significantly reducing alarm fatigue.

[0105] When embedded in a zero-trust architecture, this system provides real-time risk input to the continuous assessment engine. RADMC detection and TSCRF risk field dynamically assess the trust level of each device and session, driving the zero-trust policy decision point to dynamically adjust access authorization.

[0106] In the zero-trust embedding, this system uses causal confidence as a trust decay factor. When OICG traces the causal chain of a threat, it reduces the trust score of the relevant entity, triggering zero-trust re-authentication or isolation, thus realizing the transformation from static authorization to continuous risk assessment.

[0107] All four embedded scenarios are uniformly quantified by the response priority P_embed, which is composed of risk entropy, causal confidence and asset change norm weighted, and serves as the unified decision entry point for the host system to perform the disposal. The response priority is defined as shown in Equation (6).

[0108] (6) In Equation (6), P_embed is the embedded response priority; w_risk, w_conf, and w_asset are the risk, confidence, and asset weights, respectively, and are normalized to 1; H_risk is the cross-layer risk entropy; Conf is the causal confidence level; and Delta Asset norm represents the magnitude of the change in the value of the affected asset.

[0109] The embedded interface adopts a standardized risk message format, which includes five fields: event identifier, risk entropy, causal confidence, suggested handling, and priority P_embed. The host system consumes messages in order of P_embed.

[0110] SCADA and firewall embedding prioritizes low latency, with P_embed having a higher weight w_asset; SIEM and zero-trust embedding prioritize interpretability, with w_conf having a higher weight, and the weight is configured according to the host scenario.

[0111] When embedded, this system is decoupled from the host system: when this system fails, the host system will run in a downgraded manner according to the default strategy to avoid single point of failure affecting the availability of the host business.

[0112] The embedded processing actions are still executed and audited by the host system. This system only provides decision-making suggestions, with clear responsibilities and rights, which facilitates compliance auditing and post-event accountability.

[0113] Through a unified priority interface, this system can be embedded into heterogeneous host systems in a plug-and-play manner. The host system can make risk decisions without understanding the internal algorithms, which reduces integration costs and coupling.

[0114] Example 6: Multi-system combined use This embodiment presents a solution for the joint use of this invention with four external systems: 5G network slicing, digital twins, blockchain, and SDN. When used in conjunction with these systems, this system collaborates with them through a bidirectional interface to expand its detection, scheduling, and trusted evidence storage capabilities.

[0115] When combined with 5G network slicing, this system maps the TSCRF risk level and NEMS scheduling results to slice resource allocation weights, guiding the 5G orchestrator to dynamically expand security slices and reduce ordinary slices for high-risk services.

[0116] In 5G slicing collaboration, RADMC detection results drive slice isolation decisions, and OICG causal chain identifies risk source slices, enabling differentiated slice protection based on risk level and preventing attacks from spreading across slices.

[0117] Slice resource allocation is quantified by the overall benefit R_slice, which is obtained by normalizing the ratio of service quality benefit to security risk penalty using the demand norm, as shown in Equation (7). This metric guides the orchestrator to achieve a balance between QoS and security.

[0118] (7) In Equation (7), R_slice is the comprehensive benefit of slice allocation; w_qos and w_sec are the service quality and security weights; QoS_k is the service quality of slice k; Risk_k is the risk level of slice k; and Demand_k norm represents the resource demand intensity of slice k.

[0119] When combined with a digital twin, this system maps the risk profile of the physical Internet of Things to the digital twin in real time, rehearses attack propagation and handling schemes in the twin space, and verifies the effectiveness of the NEMS scheduling strategy before distributing it to the physical system.

[0120] In digital twin collaboration, the TSCRF risk field is visualized in the twin to present the three-layer risk transmission, and the OICG causal graph is used to perform counterfactual inference in the twin space to evaluate the causal consequences of different disposal schemes and achieve a priori verification of security decisions.

[0121] Digital twin collaboration enables simulation verification of handling strategies before physical execution, reducing the mishandling rate by more than 70%, and is suitable for scenarios with extremely high trial-and-error costs, such as critical infrastructure.

[0122] When combined with blockchain, this system stores RADMC-detected events, NEMS handling decisions, and OICG causal evidence on the blockchain, forming an immutable security audit trail that meets compliance audit and accountability requirements.

[0123] In blockchain consortium, only event summaries and causal confidence scores are uploaded to the blockchain, while the original data remains stored locally, balancing the reliability of evidence storage with data privacy; the model parameter update hashes of federated learning are also uploaded to the blockchain to ensure that the model evolution is auditable.

[0124] Blockchain consortium is suitable for scenarios involving multi-entity collaboration and requiring trusted evidence storage, such as cross-enterprise supply chain IoT and power trading IoT.

[0125] When combined with SDN, this system converts the NEMS hierarchical scheduling results into SDN flow table instructions, which are then executed by the SDN controller in three levels: local discarding at L1, rate limiting in the L2 region, and cloud redirection at L3, resulting in fine-grained traffic handling.

[0126] In the SDN consortium, TSCRF cross-layer risk entropy guides SDN to adjust forwarding paths to isolate infected network segments, and OICG causal chain identifies attack entry points and propagation paths, enabling SDN to precisely block rather than coarsely disconnect the network.

[0127] SDN enables the hierarchical scheduling capability of this system to be implemented as programmable network actions, realizing closed-loop automation from detection and decision-making to execution.

[0128] The joint interface uses bidirectional asynchronous messaging: this system publishes risk events and decision suggestions to external systems, and the external systems return execution results and status, forming a closed loop of detection-decision-execution-feedback.

[0129] When 5G slicing is combined with SDN, R_slice in equation (7) guides slice expansion, and SDN flow table performs fine blocking. The two work together to achieve dual protection of the transport layer and the forwarding layer.

[0130] When digital twins are combined with blockchain, the pre-simulated disposal plan in the twin space is recorded on the blockchain after being evaluated by OICG causality, ensuring the comparability between the pre-simulated plan and the actual implementation plan.

[0131] The system integrates a unified clock and event identifier: all systems associate logs with the same timestamp and event ID, supporting cross-system causal tracing and responsibility delineation.

[0132] The above four joint solutions can be used in combination: for example, 5G slicing provides transmission assurance, SDN performs traffic processing, digital twin pre-simulation strategy, and blockchain evidence storage decision-making, forming a three-dimensional IoT security joint defense system.

[0133] Example 7: Comparative Analysis This embodiment compares the present invention with five baseline solutions on a unified testbed. The testbed contains 12,000 simulated devices across three scenarios: industrial, residential, and urban, and includes attack samples such as injection, replay, hijacking, and scanning. Six metrics are compared.

[0134] The five baseline schemes are: B1 Traditional single model (single CNN detection), B2 Static rules (threshold rules), B3 Edge-cloud separation (no federation, no game-theoretic scheduling), B4 Centralized cloud (unified detection in the cloud), and B5 Existing federation (only FedAvg, no causal and hierarchical scheduling).

[0135] The traditional B1 single model completes detection with a single CNN, without cascading switching and resource awareness. The inference load is fixed, making it difficult to balance speed and accuracy, and it lacks the ability to transmit risks across layers and trace causality.

[0136] B2 static rules match features with a fixed threshold, resulting in the lowest latency but a high false positive rate. They cannot adapt to attack variants and have poor cross-layer diffusion rate and detection capability after cold start.

[0137] The B3 edge-cloud separation fragments detection and scheduling, lacks NEMS game-theoretic scheduling and federated collaboration, and makes it impossible to share threat intelligence between regions, resulting in a high cross-layer diffusion rate.

[0138] B4 centralized cloud-based detection concentrates all detection in the cloud, offering acceptable accuracy but the highest detection latency, and lacks autonomous edge capabilities, rendering it ineffective if the data transmission is interrupted.

[0139] The existing B5 federation uses FedAvg aggregation but lacks proximal constraints and causal analysis, resulting in significant model drift under heterogeneous nodes and insufficient interpretability.

[0140] The six indicators are detection latency, false alarm rate, resource utilization, cross-layer diffusion rate, cold start time, and interpretability. Each scheme was run for 24 hours under a uniform load, and the average value was taken. The results are shown in Table 1.

[0141]

[0142] As shown in Table 1, the present invention's solution is optimal in terms of detection latency, false alarm rate, cross-layer diffusion rate, and interpretability. Its resource utilization is second only to the centralized cloud solution, and its cold start time is significantly shortened due to federated transfer learning.

[0143] The detection latency of 2.2ms in this invention is due to the 1D-CNN initial screening (≤3ms) and resource-aware pruning with RADMC cascade switching; the false alarm rate of 1.2% is due to DP-BiLSTM confirmation and causal confidence filtering; and the cross-layer diffusion rate of 3% is due to TSCRF attenuation conduction and NEMS hierarchical blocking.

[0144] To comprehensively quantify the overall performance of each scheme, the six indicators are normalized and then weighted and summed using a weight vector. The sum is then divided by the weight norm to obtain the comprehensive score Score_comp. A higher score indicates a better overall performance. The comprehensive score is defined as shown in equation (8).

[0145] (8) In Equation (8), Score_comp is the comprehensive score; w_m is the weight of indicator m; Metric_m is the normalized indicator value; n is the sensitivity index, which is normalized by taking the reciprocal of the low-quality indicator; the denominator is the weight norm to eliminate the influence of the dimension and weight scale.

[0146] According to formula (8), the overall score of this invention is 0.91, B1 is 0.52, B2 is 0.43, B3 is 0.61, B4 is 0.58, and B5 is 0.67. This invention improves by 36% compared with the optimal baseline B5, which verifies the overall advantage of the synergy of the five core features.

[0147] Ablation analysis showed that: removing RADMC cascade switching increased latency to 5.9ms; removing TSCRF increased cross-layer diffusion rate to 14%; removing NEMS reduced resource utilization to 71%; removing OICG reduced interpretability to medium; and removing federated learning increased cold start time to 28 minutes.

[0148] Analysis by indicator dimension: This invention outperforms the centralized solution by 8 times in delay-related indicators, the single-model solution by 6 times in diffusion-related indicators, and the non-federated solution by 5 times in cold start-related indicators, demonstrating comprehensive advantages.

[0149] Statistical significance tests show that the improvements of this invention relative to each baseline in all six indicators passed the t-test with p < 0.01, indicating that the results are statistically significant rather than random fluctuations.

[0150] In the robustness test, under conditions of 30% tag noise and 20% node disconnection, the overall score of this invention still remained at 0.82, which is better than the baseline performance under normal conditions, demonstrating strong robustness.

[0151] The comparison and ablation results together show that the five core features each make irreplaceable contributions, and when combined, they achieve comprehensive leadership in detection speed, accuracy, resource efficiency, diffusion suppression, cold start, and interpretability.

[0152] Example 8: Other Unclassified Examples This embodiment presents several variations of the present invention in terms of operating mode, protocol adaptation, model engineering, and privacy protection. Each variation does not change the collaborative architecture of the five core features, but only makes local adjustments for specific needs.

[0153] Real-time monitoring variant: RADMC performs window-by-window detection in a streaming manner, TSCRF risk field is updated in real time with events, and NEMS solves Nash equilibrium online in a rolling manner, suitable for production scenarios with 24 / 7 uninterrupted monitoring.

[0154] Batch training variant: DP-BiLSTM confirms that the model and OICG causal graph are incrementally retrained in batches during idle nighttime hours, while inference is performed only during the day, balancing model freshness and online computing power usage.

[0155] Elastic scaling variant: When the device scales up, the edge nodes are horizontally scaled up, the regional layer NEMS automatically incorporates new participants to rebalance, and the federated aggregation layer redistributes weights according to the amount of data, achieving seamless scaling.

[0156] MQTT Optimized Variant: For the MQTT protocol, RADMC directly parses the topic level and retains message characteristics, while TSCRF uses the publisher-subscriber topology as the network layer risk transmission path, adapting to IoT message bus scenarios.

[0157] Risk scoring card variant: Maps TSCRF risk entropy and OICG causal confidence to a 0-100 point risk scoring card, output to the operations and maintenance dashboard, and drives manual or automatic response with intuitive classification (low / medium / high / urgent).

[0158] Enhanced variant of causal reasoning: OICG introduces do-calculus on the basis of Granger test and counterfactual reasoning to support intervention hypothesis analysis and assess the expected reduction in global risk if a certain causal edge is blocked.

[0159] Cross-domain migration variant: Federated transfer learning introduces adversarial domain adaptation, aligns the feature distributions of the source and target domains, and improves the detection performance of cross-industry migrations (such as migration from industry to energy).

[0160] Incremental learning variant: RADMC performs online incremental fine-tuning on newly emerging attack samples, combined with experience replay to mitigate catastrophic forgetting and maintain the ability to identify historical attacks.

[0161] Anomaly clustering variant: Unsupervised clustering of unconfirmed samples after RADMC screening reveals unknown attack families, which are fed back as new categories to the training set of the DP-BiLSTM confirmation model.

[0162] Resource quota variant: NEMS sets CPU, memory, and bandwidth quotas for each edge node, and adds quota overrun penalties to the game utility function to prevent single node resource exhaustion from affecting overall availability.

[0163] Model distillation and compression variant: The DP-BiLSTM teacher model is quantized and distilled to generate an int8 student model for deployment on a microcontroller, reducing inference latency to the 1ms level and making it suitable for ultra-low power terminals.

[0164] Edge caching variant: High-frequency attack characteristics and causal graph substructures are cached on edge nodes. OICG queries prioritize hitting the local cache, reducing latency and bandwidth consumption when accessing the cloud causal graph.

[0165] Privacy-differential variant: Model parameters uploaded by federated learning are superimposed with differential privacy noise, and OICG causal evidence is processed with k-anonymization, achieving a balance between collaborative sharing and individual privacy.

[0166] Multi-tenant isolation variant: In cloud heavy deployments, independent TSCRF risk fields and OICG causal graphs are maintained for different tenants. NEMS plays games independently according to tenant quotas to achieve risk and resource isolation between tenants.

[0167] Alarm deduplication variant: OICG merges multiple alarms triggered by the same causal chain, retaining only the root cause alarm and marking the propagation path, reducing the number of alarms by more than 80% and alleviating alarm fatigue.

[0168] Situational visualization variant: Renders the TSCRF three-layer risk field and OICG causal graph into an interactive situational map, supporting operations and maintenance personnel to drill down into the risk transmission and causal tracing of the equipment-network-application three layers.

[0169] Offline analysis variant: Export historical events to an offline environment, OICG offline reconstructs long-term causal graphs, supports post-event tracing and attack review, and feeds back into the rules and features of the online model.

[0170] Lightweight deployment variant: RADMC screening and L1 handling are packaged into a container image and deployed to the edge gateway with one click. NEMS and federated aggregation are provided as cloud services, reducing the deployment threshold.

[0171] The above variants can be combined as needed. For example, real-time monitoring, overlaying alarm deduplication, and situation visualization can constitute a set of operation and maintenance-friendly security operation variants, reflecting the adaptability of this invention under different operation and maintenance modes.

[0172] All variants share the same set of interfaces and data models. Adding new variants does not require changes to the core algorithm; it can be achieved only through configuration or plugin extensions, ensuring the system's evolvability and engineering maintainability.

[0173] Example 9: Complete Process Example This embodiment uses the real attack event EVT-2024-001 as the main thread to demonstrate the complete end-to-end process of this invention, from detection, propagation, scheduling, tracing to federated sharing. See also Figure 1 The overall system architecture and the event occurred in the Internet of Things system of a car manufacturing workshop.

[0174] Background of EVT-2024-001: After an attacker compromised the engineer's workstation DEV-WS-02, they issued an abnormal register write command to the production line PLC DEV-PLC-012, attempting to tamper with the robotic arm control parameters. See also... Figure 2 Edge node deployment, events are collected by edge gateway GW-EDGE-07.

[0175] The entire event sequence contains 8 key nodes. The timestamp is set to zero at the time T when the attack was first reported. The timestamp, device or node identifier and event description of each node are shown in Table 2.

[0176]

[0177] Table 2 presents the complete timing sequence of EVT-2024-001 from initial report to closed-loop control. The following section combines... Figures 1 to 15 The synergistic role of the five core features in this process is explained in stages.

[0178] T+0.0s Detection Phase: Edge gateway GW-EDGE-07 performs 1D-CNN screening on the OPC-UA messages of DEV-PLC-012, outputting anomaly detection probability P_detect=0.92, exceeding the threshold of 0.8, triggering cascading confirmation. See also Figure 2 The screening is completed locally at the edge with a latency of 2.1ms.

[0179] In the T+0.1s confirmation phase: the DP-BiLSTM confirmation model performs secondary classification on high-risk samples, identifying them as register injection attacks with a confidence level of 0.97. The knowledge distillation-compensated pruning model reduces the inference load by 40% while maintaining confirmation accuracy.

[0180] During the T+0.3s propagation phase: TSCRF propagates device-level risks to the network and application layers via cross-layer risk entropy. See also... Figure 3 Three-layer risk map and Figure 4 The norm of the risk transmission field, Delta (Risk_cross), characterizes the transmission strength. Figure 5 It demonstrates the evolution of cross-layer entropy over time.

[0181] During the propagation phase, the exponential time-varying decay weight W_decay(t) focuses the risk field on recent events, and the risk at the device layer is propagated to the application layer within 200ms. The peak value of H_risk reaches 0.86, triggering regional layer scheduling.

[0182] During the T+0.5s scheduling phase: NEMS classifies the event as L2 regional coordination. See also... Figure 6 Game scheduling, Figure 7 Level 3 scheduling process and Figure 8Resource allocation involves using Nash equilibrium at the regional layer to solve for multi-objective utility and scheduling adjacent gateways GW-EDGE-08 for collaborative processing.

[0183] During the scheduling phase, NEMS achieves a balance between minimizing risk, cost, and latency through regret matching and iterative convergence. It selects L2 collaboration instead of L3 cloud because the event has not yet spread across domains and regional resources are sufficient.

[0184] T+1.2s Origin Tracing Phase: OICG Online Incremental Cause-and-Effect Graph for Root Cause Tracing Attacks. See also... Figure 9 Cause-and-effect diagrams and Figure 10 Causal confidence propagation, Granger test and counterfactual reasoning combined to determine the causal source of the abnormal write is engineer station DEV-WS-02.

[0185] During the source tracing phase, the causal confidence propagates forward along the directed acyclic graph. The confidence of the edge DEV-WS-02→DEV-PLC-012 reaches 0.94, which is higher than other candidate paths, thus accurately locating the compromised engineer station.

[0186] Figure 9 OICG Online Incremental Cause-and-Effect Graph (EVT-2024-001 Source Tracing Path) T+0.8s Handling Phase: Based on NEMS scheduling and OICG tracing, the system executes isolation actions. See also Figure 11 The process involves GW-EDGE-07 and GW-EDGE-08 working together to disconnect the controlled PLC DEV-PLC-012 from the network and triggering zero-trust re-authentication for the engineer station DEV-WS-02. The end-to-end processing delay is 180ms.

[0187] T+2.0s Sharing Phase: Federated transfer learning aggregates the attack pattern of this event into other regional models via FedProx. See also Figure 12 Federal aggregation, Figure 13 Federal Migration and Figure 14 Cross-domain sharing allows each region to exchange model parameters, while the original data remains within the domain.

[0188] During the sharing phase, FedProx proximal term constraints suppress model drift between heterogeneous regions, and new regions can obtain the ability to detect the injection pattern without collecting local attack samples, reducing the cold start time from 30 minutes to 5 minutes.

[0189] T+3.0s debriefing phase: After the event loop is closed, OICG causal evidence and NEMS handling decisions are stored on the blockchain, and Risk_total falls back to the baseline. See also Figure 15 The overall risk evolution curve indicates that the system is now undergoing post-event review and model feedback.

[0190] To characterize the risk evolution throughout the entire event process, the total risk Risk_total is updated recursively using an exponentially weighted method. It consists of historical risk, decay-weighted cross-layer risk entropy, and cross-layer risk increment norm. The evolution of total risk is defined as shown in equation (9).

[0191] (9) In equation (9), Risk_total(t) is the total risk at time t; alpha is the update coefficient, which takes the value (0,1) and controls the length of historical risk memory; W_decay(t) is the exponential time-varying decay weight; H_risk(t) is the cross-layer risk entropy; Delta Risk_cross(t) takes the norm to characterize the cross-layer transmission strength.

[0192] Evolving according to equation (9), the Risk_total of EVT-2024-001 peaks at 0.86 at T+0.3s, begins to decay after T+0.8s, and falls back to the baseline of 0.12 at T+3.0s. Figure 15 The overall risk evolution curve is consistent, verifying the effectiveness of the spatiotemporal evolution modeling of the risk field.

[0193] Figure 15 EVT-2024-001 Total Risk Evolution Curve End-to-end metrics: The event detection latency was 2.1ms, the acknowledgment latency was 0.1s, the processing latency was 180ms, the cross-layer diffusion rate was 3%, and the false alarm rate was 1.2%. All metrics are better than the five baseline schemes and are consistent with the design goals of this invention.

[0194] EVT-2024-001 demonstrates that the five core features form a closed loop of detection-transmission-scheduling-source tracing-sharing under real attacks: RADMC for rapid and accurate detection, TSCRF for quantifying risk transmission, NEMS for optimal scheduling, OICG for causal source tracing, and federated transfer learning for cross-domain sharing.

[0195] Following the incident review, OICG added the causal chain of this attack as a new pattern increment to the causal graph. The DP-BiLSTM confirmed that the model absorbed new samples through federated aggregation, achieving model self-evolution and further improving the system's detection confidence under subsequent similar attacks.

[0196] In summary, this invention constructs a complete IoT security management method and system from single-point detection to cross-domain collaboration through the synergy of RADMC, TSCRF, NEMS, OICG and federated transfer learning. It has comprehensive advantages such as low latency, low false alarms, high resource efficiency, low diffusion, fast cold start and strong interpretability, and is suitable for various IoT security management scenarios.

[0197] Example 10: Formula Derivation Example This embodiment combines the mathematical derivations of five core technologies—resource-aware dual-model cascade dynamic switching, spatiotemporal risk transmission field, Nash equilibrium multi-objective game scheduling, online incremental causal graph, and federated transfer learning—into a set of refined formulas. The formula numbers continue from (10) to (21), and all variables are named in English and given in the form of subscripts.

[0198] (I) Derivation of Dynamic Switching of Resource-Aware Dual-Model Cascade in RADMC The 1D-CNN screening model adopts a one-dimensional convolutional stacking structure. The receptive field of a single convolutional kernel is equal to the kernel size k_size. After stacking n_layers, the equivalent receptive field expands linearly, enabling the lightweight model to cover the key window of traffic time series.

[0199]

[0200] The system constructs a Pareto front with the dual objectives of maximizing classification accuracy and minimizing computational resource consumption. It uses grid search to solve the problem within the range of pruning ratio (prune_ratio) from 0.2 to 0.6, and selects a compromise point on the front that balances accuracy and resources.

[0201] To compensate for the accuracy loss caused by structured pruning, the DP-BiLSTM confirmation model adopts a knowledge distillation strategy, with the unpruned complete model as the teacher and the pruned model as the student. The distillation loss function is shown in Equation (10).

[0202] (10) In Equation (10), L_distill is the distillation loss; w_hard is the hard label loss weight, with a value ranging from 0.5 to 0.9; L_hard is the cross-entropy hard loss; and KL is the KL divergence between the teacher and student soft label distributions. This strategy keeps the accuracy loss within 1 percentage point when the pruning rate is 0.4.

[0203] The dynamic switching is based on the real-time resource status of the edge node to calculate the resource score R_score. The CPU utilization, available memory and bandwidth are combined and weighted by formula (11) to determine which level of model is currently activated.

[0204] (11) In equation (11), w_cpu, w_mem, and w_bw are the weights of CPU, memory, and bandwidth, respectively, and their sum is 1; CPU_util, Mem_avail, and BW_remain are normalized real-time resource metrics. The pruning rate is increased when R_score is below the lower threshold and decreased when it is above the upper threshold.

[0205] When R_score is below the lower threshold, the system automatically reduces the pruning rate and prioritizes the activation of 1D-CNN lightweight screening; when R_score is above the upper threshold, the system increases the pruning rate and activates DP-BiLSTM confirmation as needed, achieving a dynamic balance between accuracy and resources.

[0206] The optimal pruning rate, as determined by the Pareto front, adaptively varies between 0.28 and 0.52 with the resource score. Compared to the fixed pruning rate scheme, it reduces resource consumption by approximately 38% at the same level of accuracy, thus verifying the effectiveness of the bi-objective optimization.

[0207] Knowledge distillation allows students to approach the accuracy of teachers even when only 60% of the parameters are retained. Hard label constraints ensure clear category boundaries, while soft labels convey knowledge of inter-class similarity. The two work together to significantly alleviate the generalization degradation caused by pruning.

[0208] In summary, RADMC achieves on-demand cascading of lightweight screening and high-precision confirmation through four stages: receptive field design, Pareto optimal pruning, knowledge distillation compensation, and resource perception scoring, laying the foundation for efficient edge detection.

[0209] (II) Derivation of the Spatiotemporal Risk Transmission Field of TSCRF TSCRF models the risk situation of the device layer, network layer, and application layer as a unified spatiotemporal risk field. It uses cross-layer risk entropy to quantify the uncertainty of risk distribution in each layer and uses time-varying decay weights to characterize the time decay characteristics of risk transmission.

[0210] Cross-layer risk entropy uses the Shannon entropy form to quantify the concentration of risk distribution at a certain layer. The lower the entropy value, the more concentrated the risk is in a single category and the stronger the transmission effect. The risk entropy is calculated as shown in Equation (12).

[0211] (12) In equation (12), H_entropy is the risk entropy of a certain layer, which takes a value of 0 to 1 after normalization; p_i is the probability of the i-th type of risk, which satisfies that the sum of all probabilities is 1; log2 is the logarithm to the base 2. The lower the entropy value, the greater the intensity of risk transmission from this layer to adjacent layers.

[0212] The risk transmission effect gradually weakens over time. This characteristic is characterized by an exponential time-varying decay weight with a decay coefficient ranging from 0.01 to 0.2. The time-varying decay weight is derived as shown in equation (13).

[0213] (13) In equation (13), W_decay(t) is the decay weight at time t, with a value between 0 and 1; r_decay is the decay coefficient; and t is the time since the occurrence of the risk event. This weight makes the recent risk transmission strong and the long-term transmission weak, which is in line with the law of security situation evolution.

[0214] To prevent excessive risk propagation, the system is equipped with adaptive gating. When the real-time risk entropy of a certain layer exceeds the dynamic threshold determined by the historical moving average and standard deviation, the transmission quantity is set to zero through the gating function to block the propagation.

[0215]

[0216] In the above formula, gate(t) is the gated output; gate_th is the adaptive threshold; and sigma_gate is the smoothing coefficient. When the risk entropy exceeds the threshold, gate tends to 0, thus blocking the propagation; when it is below the threshold, gate tends to 1, allowing the propagation, thereby achieving adaptive suppression of risk diffusion.

[0217] Based on entropy quantification and attenuation weight, the risk update of the downstream layer is equal to the current risk of this layer plus the risk increment of the upstream layer after modulation by attenuation weight and entropy factor. If the situation of any layer exceeds the threshold, a preventive response of the adjacent layer is triggered.

[0218] The three-layer risk profile is aggregated into a spatiotemporal risk map after entropy calculation, weight modulation, and gating suppression. The map nodes represent the risk values ​​of each layer, and the edge weights represent the transmission strength, providing intuitive situational input for regional collaborative decision-making.

[0219] In summary, TSCRF constructs a three-layer spatiotemporal risk transmission field through three mechanisms: Shannon risk entropy, exponential time-varying decay, and adaptive gating, thereby achieving the quantification, decay, and controllable diffusion of cross-layer risk situations.

[0220] (III) Derivation of NEMS Nash Equilibrium Multi-Objective Game Scheduling NEMS models threat response as a multi-agent game with multiple objectives: maximizing risk mitigation gains, minimizing response costs, and minimizing scheduling delays. It solves for the stable allocation ratio of each response strategy through Nash equilibrium.

[0221] The scheduling utility function integrates risk mitigation benefits, response costs, and scheduling delays. A higher utility value indicates a better strategy. The utility function is shown in equation (14).

[0222] (14) In equation (14), U_utility is the scheduling utility function; w_risk, w_cost, and w_delay are the risk, cost, and delay weights, respectively, and the sum of the three is 1; Risk_level is the risk level; Cost and Delay are the normalized cost and delay.

[0223] The system constructs a payoff matrix based on the utility values ​​of each strategy in pairwise combinations. The matrix elements are the utility differences between row strategies and column strategies. The symmetric zero-sum property of the payoff matrix ensures that there is a mixed-strategy Nash equilibrium in the game.

[0224] The regret matching iteration is used to approximate the Nash equilibrium. In each round, the regret value of each strategy is updated according to the historical utility. The larger the regret value, the higher the degree of underestimation of the strategy. The regret matching iteration is shown in Equation (15).

[0225] (15) In equation (15), R_regret,i is the regret value of policy i; U_max is the optimal utility in the current round; and U_i is the utility of policy i. The policy selection probability is proportional to the positive regret value. After multiple rounds of iteration, the regret value converges, and the policy distribution tends to Nash equilibrium.

[0226] Based on risk levels, scheduling is divided into three levels: L1, L2, and L3. L1 thresholds of 0.3 to 0.5 trigger local lightweight responses; L2 thresholds of 0.5 to 0.7 trigger regional collaborative responses; and L3 thresholds of 0.7 to 0.9 trigger global joint responses.

[0227] Level 1 involves edge nodes autonomously handling low-risk events; Level 2 involves the regional coordination layer scheduling resources from multiple nodes; Level 3 involves the cloud coordinating global resources and linking with the federated model for updates. The three-level threshold calibration ensures that the response strength matches the risk level.

[0228] In finite games, regret matching can converge to an approximate Nash equilibrium after dozens of iterations. In the equilibrium state, any agent's unilateral deviation from the strategy cannot improve utility, thus ensuring the global stability of the scheduling scheme.

[0229] Tiered scheduling reduces response latency to less than 5 milliseconds for L1 level, less than 30 milliseconds for L2 level, and less than 100 milliseconds for L3 level, resulting in a reduction of approximately 60% in response latency in high-risk scenarios compared to single centralized scheduling.

[0230] In summary, NEMS achieves Nash equilibrium solution and hierarchical scheduling for multi-objective games by using utility functions, payoff matrices, regret matching, and three-level threshold calibration, taking into account response efficiency, cost, and timeliness.

[0231] (iv) Derivation of OICG Online Incremental Causal Graph OICG constructs a threat causal graph in an online incremental manner, where nodes represent risk events and edges represent causal transmission relationships. The existence of edges is determined by Granger causality tests, and the direction and strength of causality are verified by counterfactual reasoning.

[0232] Granger causality is based on the vector autoregressive (VAR) model to test the improvement in predictive ability between variables. The causal score is constructed by the ratio of the restricted and unrestricted sum of squared residuals. The Granger causal VAR model is shown in Equation (16).

[0233] (16) In equation (16), G_granger(i,j) is the causal score of variable i on j; SSE_r is the restricted residual sum of squares after removing i; SSE_u is the unrestricted residual sum of squares including i; the larger the score, the stronger the causality.

[0234] The VAR model represents the risk time series of each layer as a linear combination of its own and the lag terms of adjacent layers. The lag order p_lag is taken from 1 to 10 and is automatically selected by the Bayesian information criterion. The model parameters are updated incrementally with new data.

[0235] To distinguish between causation and correlation, the system performs counterfactual intervention reasoning on candidate causal edges and calculates the probability change of the outcome variable after intervention. The counterfactual do-calculus is shown in Equation (17).

[0236] (17) In Equation (17), CF_counter represents the counterfactual causal effect; P(Y|do(X)) represents the probability of Y after intervention in X; backdoor adjustment is achieved by summing the confounding variable z to eliminate false causal edges.

[0237] Edges that have undergone Granger and counterfactual double verification are assigned causal confidence, and the confidence is propagated forward along the directed acyclic graph (DAG). The propagation of causal confidence is shown in Equation (18).

[0238] (18) In equation (18), C_confidence(j) is the confidence level of node j; W_edge(i,j) is the causal weight of edge i to j; and C_confidence(i) is the confidence level of upstream node i. A higher confidence level indicates that the causal path is more reliable.

[0239] The confidence scores are propagated forward along the DAG in topological order. The confidence scores of the root nodes are initialized by the Granger score, and the confidence scores of the intermediate nodes are the weighted sum of the incoming edges. Weak causal edges with confidence scores below the threshold are removed in real time during the propagation process.

[0240] In summary, OICG constructs an incrementally updatable and interpretable online causal graph through Granger VAR, counterfactual do-calculus, and causal confidence propagation, providing reliable causal evidence for root cause localization.

[0241] (v) Derivation of Federated Transfer Learning Federated transfer learning collaboratively trains a global model without sharing the original data. Edge nodes train locally and only upload the model parameters, which are then aggregated in the cloud to form a global model, balancing privacy protection and model performance.

[0242] FedAvg calculates the local model parameters by weighting the amount of data from each client. Clients with larger amounts of data have higher contribution weights. The FedAvg aggregation is shown in Equation (19).

[0243] (19) In equation (19), w_global represents the global model parameters; n_k represents the amount of data for client k; N represents the total amount of data; and w_k represents the local model for client k. Weighted aggregation biases the global model towards the data-rich client distribution.

[0244] To mitigate model drift caused by heterogeneous client data, the system introduces FedProx near-end terms on the basis of FedAvg to constrain the local model from deviating too far from the global model. The FedProx near-end terms are shown in Equation (20).

[0245] (20) In Equation (20), L_fedprox is the FedProx loss; L_local(w) is the local empirical loss; coef_prox is the near-end coefficient, with a value of 0.001 to 0.1; and the norm term constrains the deviation between w_local and w_global.

[0246] The proximal term balances the local optimality and global consistency of each client model, significantly reducing model divergence under non-independent and identically distributed data, accelerating federated convergence and improving the generalization ability of the global model.

[0247] The aggregation uses a secure aggregation protocol. After the model parameters uploaded by the client are masked and summed, the cloud can only solve the aggregation result and cannot know the individual parameters, thus preventing the leakage of member parameters from the protocol layer.

[0248] Local differential privacy is achieved by superimposing Gaussian noise before parameter uploading. The noise scale is adaptively adjusted according to the privacy budget, achieving a quantifiable balance between model accuracy and privacy protection.

[0249] For newly connected edge nodes lacking labeled data, the system migrates the global model as a pre-trained model, and only fine-tunes the last classification layer to quickly adapt to the new scenario, reducing the cold start sample requirement by 80%.

[0250] After multiple rounds of federated iterations, the global model converged stably under heterogeneous client distribution, with an average accuracy improvement of about 7 percentage points compared to pure local training, verifying the effectiveness of federated transfer learning.

[0251] In summary, federated transfer learning achieves collaborative modeling and rapid cold start adaptation under privacy and security by using FedAvg weighted aggregation, FedProx proximal constraints, secure aggregation, and differential privacy.

[0252] Example 11: Engineering Implementation and Performance Testing Example This embodiment presents the system hardware configuration, software stack, containerized deployment, MQTT topic design, and five key performance test results. All tests were completed in an industrial IoT experimental environment, and the present invention is compared with five baseline solutions.

[0253] The system adopts a three-layer hardware configuration of edge, region, and cloud. The edge layer focuses on low power consumption and local inference, the region layer takes into account computing power and collaboration, and the cloud layer provides large-scale training and global analysis capabilities. The hardware configuration requirements are shown in Table 1.

[0254]

[0255] Edge nodes are equipped with NPU acceleration units to support millisecond-level inference for 1D-CNN, regional nodes are configured with GPUs to meet real-time cross-layer risk calculations, and cloud nodes use A100 to ensure the throughput requirements for federated aggregation and causal graph construction.

[0256] The software stack is uniformly deployed in a containerized manner, and the versions of each component have been verified for compatibility to ensure consistency and reproducibility of the three-layer environment. The software stack versions are shown in Table 2.

[0257]

[0258] PyTorch 2.1 supports dynamic model pruning and quantized inference, EMQX 5.0 provides millions of MQTT connections, Flink 1.18 supports streaming computation of risk profiles, and InfluxDB 2.7 stores time-series risk data.

[0259] All three services are packaged as Docker images. Edge images are quantized and compressed to less than 50MB to adapt to limited storage. Regional and cloud images contain complete training dependencies. Image versions are managed uniformly through a private repository.

[0260] Kubernetes isolates the three layers of load balancing using namespaces, uses lightweight K3s clusters at the edge to reduce resource consumption, and uses standard K8s in the region and cloud with horizontal automatic scaling configured, dynamically adjusting the number of replicas according to the risk situation and load.

[0261] The deployment uses Helm Chart template configuration, and values.yaml distinguishes edge, region and cloud parameters, supports canary release and one-click rollback, and ensures zero interruption in the system update process.

[0262] MQTT communication is designed in a topic-based layered manner. Critical instructions use QoS2 to ensure accurate one-time delivery, while heartbeats and telemetry use QoS0 to reduce overhead. The MQTT topic design is shown in Table 3.

[0263]

[0264] Device events and risk status are delivered at least once using QoS1, scheduling instructions and threat intelligence are delivered using QoS2 to prevent duplicate execution, and heartbeats are delivered using QoS0 to minimize bandwidth usage while keeping the device alive.

[0265] The detection latency test statistics of the present invention and five baseline schemes were compared with the average, P95, P99 detection latency and throughput. Each scheme processed 100,000 samples, and the results are shown in Table 4.

[0266]

[0267] The average detection latency of this invention is 2.8 milliseconds, which is 71% lower than a single BiLSTM baseline and 85% lower than a centralized cloud baseline, thanks to the lightweight screening and cascaded on-demand activation mechanism of 1D-CNN.

[0268] While rule engines have the lowest latency, they have a high false positive rate. This invention reduces the false positive rate by an order of magnitude while having a slightly higher latency than rule engines, achieving the best trade-off between accuracy and timeliness.

[0269] The false alarm rate test statistics for each scheme included the false alarm rate, false negative rate, and F1 score. The results are shown in Table 5.

[0270]

[0271] The present invention has a false alarm rate of 0.82%, a false negative rate of 1.15%, and an F1 score of 0.985, which are significantly better than all baselines. This is due to the dual filtering of high-precision confirmation by DP-BiLSTM and root cause verification by causal graph.

[0272] Compared to the rule engine's 5.6% false alarm rate, this invention reduces it by approximately 85%, preventing a large number of invalid alarms from overwhelming real threats and reducing the burden on operations and maintenance personnel.

[0273] The resource utilization test statistics for each scheme included CPU utilization, memory usage, energy consumption, and model size. The results are shown in Table 6.

[0274]

[0275] This invention achieves a CPU utilization of 18.5%, 45MB of memory, and 320mW of power consumption, which is 71% lower than the baseline memory consumption of a single BiLSTM and 61% lower power consumption, thanks to structured pruning and on-demand cascaded activation.

[0276] After pruning and quantization, the model is only 2.8MB in size and can run smoothly on resource-constrained edge gateways. Compared with the unpruned model, the size is reduced by about 78%, which verifies the effectiveness of Pareto optimal pruning.

[0277] The cross-layer diffusion rate test statistically analyzed the diffusion rate from the device to the network and from the network to the application, as well as the average blocking time. The results are shown in Table 7.

[0278]

[0279] The device-to-network diffusion rate of this invention is 4.2%, and the network-to-application diffusion rate is 1.8%, which is more than 80% lower than the cloud-based centralized baseline diffusion rate, thanks to TSCRF gating blocking and graded prevention response.

[0280] With an average blocking time of 1.2 seconds, which is 90% lower than that of the rule engine, adaptive gating instantly blocks the transmission when the risk entropy exceeds the threshold, thus containing the threat in its infancy.

[0281] The cold start time test statistics for each scheme were compiled, including cold start, model loading, and first sample readiness time. The results are shown in Table 8.

[0282]

[0283] The cold start time of this invention is 3.2 seconds, which is 74% lower than that of a single BiLSTM, thanks to the pre-trained model of federated transfer learning and the lightweight 1D-CNN priority loading strategy.

[0284] The first sample is ready in 1.1 seconds, and new access nodes can be quickly put into testing after migration and fine-tuning, meeting the plug-and-play deployment requirements of IoT devices.

[0285] The comprehensive score was calculated by normalizing and weighting four indicators: detection latency, false alarm rate, resource utilization rate, and cross-layer diffusion rate. The results are shown in Table 9.

[0286]

[0287] This invention achieved a comprehensive score of 96.8, which is 14.7 percentage points higher than the second-best baseline. It ranks among the top in all four indicators, demonstrating the overall advantage of the synergy of the five core technologies.

[0288] A single baseline may only excel in a few metrics. This invention achieves a comprehensive balance of accuracy, timeliness, resources, and security by organically integrating dual-model cascading, risk transmission, game-theoretic scheduling, causal reasoning, and federated learning.

[0289] The test environment is deployed in three industrial parks, including 2,000 edge devices, 8 regional nodes and 1 cloud cluster. The network covers a hybrid link of 5G and industrial Ethernet, covering typical industrial IoT deployment models.

[0290] The test dataset uses a mixture of real industrial IoT traffic and injection attack data, covering eight types of threats such as DDoS, abnormal logins, and firmware tampering. The ratio of positive to negative samples is approximately 1 to 9, and each scheme is trained and evaluated independently.

[0291] The baseline scheme uses the same feature engineering and training data as this invention, only the model and scheduling strategy are different to ensure the fairness of the comparison; each test is repeated 10 times and the average is taken to eliminate random fluctuations.

[0292] The statistical significance was verified by paired t-test, and the p-values ​​were all less than 0.01, indicating that the performance improvement of the present invention compared with each baseline is statistically significant rather than a random phenomenon.

[0293] Ablation experiments showed that removing any of the core technologies resulted in a decrease in the overall score of 5 to 12 points. Among them, removing TSCRF had the greatest impact on the diffusion rate, while removing RADMC had the greatest impact on the delay, verifying the indispensability of the five technologies.

[0294] Even under network jitter and node failure scenarios, this invention still maintains a detection accuracy of over 90%. Thanks to the fault tolerance of federated aggregation and the K8s automatic migration mechanism, the system has strong robustness.

[0295] The system latency increases by less than 15% when the number of nodes expands from 100 to 10,000. Horizontal scaling at the regional layer and cloud load balancing ensure linear scalability under large-scale deployment.

[0296] The edge node has a full-load power consumption of less than 1.2 watts and a standby power consumption of less than 0.3 watts, which meets the long-term deployment energy consumption constraints of battery-powered IoT devices.

[0297] The system has been piloted in three industrial parks, and has run continuously for 90 days without any major failures. It has handled more than 1,200 threat incidents with a false alarm rate of less than 0.5%.

[0298] Edge concurrency testing shows that the detection latency increase of a single node is less than 15% under 1000 concurrent connections, and the NPU batch inference effectively supports high-concurrency scenarios.

[0299] The system showed no memory leaks during 7×24-hour long-term stability testing, and the mean time between failures (MTBF) exceeded 8,000 hours, meeting the requirements for long-term continuous operation in industrial applications.

[0300] This invention achieves an average detection rate of 98.5% for eight types of threats, including DDoS, abnormal logins, and firmware tampering. After rapid adaptation through federated migration, the detection rate for unknown variants is increased to over 95%.

[0301] Federation aggregation has a communication overhead of less than 2MB per round, which is three orders of magnitude lower than the original data upload, significantly saving edge uplink bandwidth and transmission energy consumption.

[0302] In summary, engineering implementation and performance testing show that the present invention outperforms the five baseline schemes in five key indicators: detection latency, false alarm rate, resource utilization, cross-layer diffusion rate, and cold start.

[0303] The synergy of five core technologies enables the system to achieve a comprehensive balance between accuracy, timeliness, resources, and security, meeting the stringent requirements of the Industrial Internet of Things for real-time performance, lightweight design, and high reliability.

[0304] Example 12: Parameter List This embodiment simplifies the system parameters into four three-line tables, covering four groups of parameters: RADMC, TSCRF and NEMS, OICG and federated learning, and system and network. Each parameter is given its symbol, meaning, value range, and unit.

[0305] Table 1 lists the parameters related to the dynamic switching of the RADMC resource-aware dual-model cascade, covering three categories: convolutional structure, pruning optimization, and resource scoring, totaling 15 parameters.

[0306]

[0307] In the RADMC parameters, the pruning ratio (prune_ratio) is the core adjustable term, with a value ranging from 0.2 to 0.6 determined by the Pareto front; the kernel size (k_size) and the number of filters (n_filter) determine the receptive field and lightweight nature of the 1D-CNN.

[0308] The sum of weights w_acc and w_res is 1, and w_hard is set to 0.7 as the default distillation ratio; R_score integrates three resource indicators, and its upper and lower limit thresholds drive the dynamic switching of the cascaded model.

[0309] Table 2 lists the parameters related to the TSCRF spatiotemporal risk transmission field and NEMS game scheduling, covering risk entropy, decay, utility, and classification threshold, totaling 15 parameters.

[0310]

[0311] In the TSCRF parameters, the decay coefficient r_decay is set to 0.05 as the default value, which determines the time decay rate of risk propagation; the gate threshold gate_th is adaptively determined by historical entropy sliding statistics.

[0312] In the NEMS parameters, the thresholds for L1, L2, and L3 are 0.4, 0.6, and 0.8 respectively, dividing the three-level response boundaries; the utility weights w_risk, w_cost, and w_delay are summed to 1, and are optimized according to the scenario.

[0313] Table 3 lists the parameters related to OICG online incremental causal graph and federated transfer learning, covering causal testing, counterfactual, confidence, and federated aggregation, totaling 15 parameters.

[0314]

[0315] In the OICG parameters, the VAR lag order p_lag is automatically selected by BIC; CF_counter and C_confidence take values ​​from 0 to 1 to quantify causal effects and path reliability; the proximal coefficient coef_prox controls the FedProx constraint strength.

[0316] In the federation parameters, the amount of client data n_k determines the FedAvg aggregation weight; the learning rate eta_lr is set to 0.005 as the default; the federation round T_round adaptively stops between 10 and 200 according to the convergence curve.

[0317] Table 4 lists the system and network operating parameters, covering detection cycle, communication, deployment and caching, etc., totaling 17 parameters.

[0318]

[0319] In the system parameters, the detection period T_detect is set to 0.5 seconds to balance timeliness and overhead; the model synchronization period T_sync is set to 300 seconds; and the MQTT service quality QoS_level is configured according to topic levels.

[0320] In the network parameters, the edge bandwidth BW_edge and the number of nodes N_node determine the scale of the regional layer aggregation; the encryption algorithm uses AES-256 to ensure communication security; and the heartbeat interval is set to 10 seconds.

[0321] The default values ​​of each parameter are determined through grid search and Bayesian optimization. The goal is to maximize the comprehensive score on the industrial IoT benchmark dataset to ensure near-optimal parameter configuration.

[0322] Sensitivity analysis shows that the parameters prune_ratio, r_decay, and coef_prox have the most significant impact on system performance. A 10% disturbance to these parameters results in a change of no more than 3% in the overall score, indicating that the system is robust to parameter disturbances.

[0323] The values ​​in the table are recommended ranges, and the default values ​​are the optimal configurations for most industrial scenarios. For special scenarios, the values ​​can be fine-tuned within the range based on hardware capabilities and threat characteristics.

[0324] In resource-constrained scenarios, prioritize lowering n_filter and n_hidden and increasing prune_ratio; in high-threat scenarios, prioritize lowering L2 and L3 thresholds to trigger collaborative responses earlier; in heterogeneous data scenarios, increase coef_prox.

[0325] In RADMC, w_cpu, w_mem, and w_bw are allocated according to the resource bottleneck of edge devices. CPU-intensive devices are given more w_cpu, while bandwidth-constrained devices are given more w_bw, thus achieving scenario adaptation of resource scoring.

[0326] In TSCRF, Delta_t and r_decay together determine the decay time constant. Under typical configuration, risk propagation decays to less than 5% of the initial value within 60 seconds, which meets the actual need for rapid fading of the security situation.

[0327] In OICG, the G_granger threshold and C_confidence threshold are jointly used to control the sparsity of causal graphs. Under the default configuration, the number of edges in the causal graph is stable at 1.5 to 2.5 times the number of nodes, balancing connectivity and sparsity.

[0328] The batch_size and n_thread in the system are configured according to the hardware concurrency capability. For NPU devices, the batch size is 32 and the thread size is 4. The cache_size affects the latency of risk status query and is recommended to be no less than 100MB.

[0329] Some parameters are interdependent. prune_ratio affects the model size and thus the cold start time, while r_decay affects the diffusion rate and thus the scheduling threshold calibration. These factors need to be considered together during optimization.

[0330] The system maintains an overall score above 90 in the central 70% range of parameter values. Performance drops slightly in the edge range but is still better than the baseline, indicating that the parameter configuration has a wide robust window.

[0331] The five technical parameters under the default configuration have been jointly optimized. Adjusting a single parameter individually may disrupt the optimal synergy. It is recommended to maintain the default combination first and then fine-tune it according to the scenario.

[0332] The parameter table covers all configurable items in the v1.4 system, is backward compatible with the v1.3 configuration file, and automatically migrates old parameters and adds new default values ​​during upgrades.

[0333] In summary, the four parameter tables systematically outline the five core technologies and 62 key parameters for system operation, providing a complete reference for engineering implementation and scenario optimization.

[0334] The default parameter values ​​have been optimized and are robust to disturbances. Combined with the tuning guidelines, near-optimal configurations can be quickly achieved in different industrial IoT scenarios.

[0335] Example 13: End-to-End Complete Process Example This embodiment uses the Industrial Internet of Things (IIoT) security incident EVT-2024-001 as the main thread to describe the complete end-to-end process from anomaly detection to situation recovery, covering 10 key time nodes. (See [link to relevant documentation]). Figure 14 Overall flowchart of end-to-end event handling.

[0336] The event sequence is shown in Table 10, which lists the handling actions, response levels and related diagrams of 10 key nodes according to time offset, and fully presents the three-level collaborative handling process.

[0337]

[0338] At time T0 (0 milliseconds), a traffic spike occurs in the edge device E-017 traffic statistics. The 1D-CNN screening model is triggered within 2.8 milliseconds. (See below) Figure 2 Cascaded architecture and Figure 9 Screening model structure diagram.

[0339] At time T1 (2.8 milliseconds), DP-BiLSTM confirmed the anomaly with a classification probability P_class of 0.94, marking it as a suspected DDoS event. See [link / reference]. Figure 10 Confirm the model structure diagram.

[0340] At time T2 (5.2 milliseconds), the resource score R_score is 0.32, which is below the lower threshold, triggering degradation pruning to maintain edge detection capability. See [link / reference needed]. Figure 2 Dynamically switch architecture.

[0341] At time T3, 8.0 milliseconds, the risk situation is reported to the regional coordination layer. The device layer risk entropy H_entropy is 0.31. (See below) Figure 3 Conducting field structure and Figure 11 Flowchart for weight calculation.

[0342] At time T4 (12.5 milliseconds), TSCRF calculations show cross-layer propagation, increasing the network layer risk to 0.62. Cross-layer propagation triggers preventative responses in adjacent layers. (See [link to relevant documentation]). Figure 3 .

[0343] At time T5 (18.0 milliseconds), NEMS initiates L2-level scheduling, unfortunately selects an isolation strategy, and distributes it to the regional nodes. (See below) Figure 4 Game scheduling flowchart.

[0344] At time T6 (25.0 milliseconds), OICG counterfactual inference confirmed DDoS as the root cause, with a causal confidence level of 0.91. (See [link to OICG counterfactual inference]). Figure 5 Cause-effect graph construction and Figure 6 A diagram illustrating counterfactual reasoning.

[0345] At time T7 (32.0 milliseconds), the federated aggregation updates the global model and distributes it to all edge nodes. See [link / reference]. Figure 7 Federal aggregation flowchart.

[0346] At 45.0 milliseconds T8, the isolation policy was executed, abnormal traffic decreased by 92%, and the threat was effectively contained. (See below) Figure 8 Sequence diagram for isolated execution.

[0347] At time T9, 60.0 milliseconds later, the situation is restored, risk entropy reverts to baseline, and the event loop is closed. (See attached image) Figure 13 Situation map and Figure 15 Deploy the topology map.

[0348] The total risk evolution of the entire event process is described by equation (21), which combines time-varying decay, risk at each level and smoothing factor to characterize the dynamic evolution of the total risk over time.

[0349] (twenty one) In Equation (21), Risk_total is the total risk; r_smooth is the smoothing factor; W_decay(t) is the time-varying decay weight; and Risk_l(t) is the risk at layer l. The total risk decays rapidly after isolation and converges to the baseline.

[0350] Figure 1 This is a general diagram of the system's three-layer collaborative architecture, demonstrating the interaction between intelligent agents at the edge, regional, and cloud levels. Figure 2 Diagram of the RADMC dual-model cascaded dynamic switching architecture; Figure 3 This is a diagram of the spatiotemporal risk transmission field structure of TSCRF.

[0351] Figure 4 A flowchart of the NEMS multi-objective game scheduling process is provided, illustrating the regret matching iteration. Figure 5 Flowchart for constructing the OICG online incremental cause-effect graph; Figure 6 This is a diagram illustrating counterfactual do-calculus reasoning.

[0352] Figure 7 A flowchart illustrating the aggregation process of federated transfer learning, demonstrating the collaboration between FedAvg and FedProx; Figure 8 This is a sequence diagram of the isolation strategy execution. Figure 9 This is a structural diagram of a 1D-CNN screening model.

[0353] Figure 10 The model structure diagram for DP-BiLSTM is confirmed, showing bidirectional encoding and pruning. Figure 11 Flowchart for calculating cross-layer risk transmission weights; Figure 12 This is a schematic diagram of adaptive gated transmission control.

[0354] Figure 13 A visualization of the risk situation; Figure 14 This is the overall flowchart for end-to-end event handling. Figure 15 Provide a system deployment topology diagram to show the distribution of nodes and communication links across the three layers.

[0355] In this incident, the edge layer completed detection in 2.8 milliseconds, the regional layer completed scheduling in 18 milliseconds, and the cloud layer completed root cause localization in 25 milliseconds. The three layers worked together to achieve threat containment within 60 milliseconds.

[0356] Compared to the traditional centralized solution with a processing cycle of 45 seconds, the end-to-end response of this invention is only 60 milliseconds, improving the timeliness by about 750 times, verifying the synergistic advantages of hierarchical scheduling and causal reasoning.

[0357] The isolation strategy reduced abnormal traffic by 92% within 13 milliseconds without any cross-application layer propagation, demonstrating the effectiveness of the isolation between TSCRF gating and NEMS scheduling.

[0358] Federated aggregation completes the global model update and distribution at time T7, enabling all edge nodes to identify the DDoS variant immediately after the event, thus achieving rapid sharing of experience.

[0359] At time T9, the risk entropy reverts to the baseline, the system automatically generates an event report and updates the cause-effect graph, thus closing the event loop. The entire process requires no manual intervention, meeting the needs of autonomous and safe operation.

[0360] This end-to-end process is also applicable to threats such as abnormal logins and firmware tampering. Only the scheduling strategy and cause-effect graph structure are adaptively adjusted according to the threat type, reflecting the universality of the process.

[0361] In the incident, 1D-CNN screening generated 3 suspected alarms, and DP-BiLSTM confirmed and filtered 2 of them as false alarms, ultimately reducing the number of false alarms to 1, thus verifying the filtering value of the cascade confirmation mechanism.

[0362] After the event loop is closed, the system incorporates EVT-2024-001 into the causal graph and federated training samples, which further reduces the detection latency of similar threats by 15% and enables continuous self-learning.

[0363] The entire process is completed automatically, with only event reports generated in T9 for operation and maintenance auditing. The rate of manual intervention is less than 5%, meeting the dual requirements of autonomous and secure operation and auditability.

[0364] In summary, the end-to-end process based on EVT-2024-001 verifies the full-link collaboration of the five core technologies from detection, transmission, scheduling, inference to federated updates.

[0365] The system completes threat containment and situation recovery within 60 milliseconds. Fifteen attached figures fully support the technical details of each step, proving the effectiveness of the invention in real industrial IoT scenarios.

[0366] Example 14: Conclusion This invention proposes an IoT security management method and system based on artificial intelligence. Through five core technologies—RADMC, TSCRF, NEMS, OICG, and federated transfer learning—it achieves comprehensive security protection with low latency, low false alarms, low resource consumption, and low diffusion.

[0367] Tests show that the present invention has a detection delay of 2.8 milliseconds, a false alarm rate of 0.82%, and a cross-layer diffusion rate of less than 5%, with a comprehensive score of 96.8. It outperforms the five baseline schemes in all four dimensions of accuracy, timeliness, resources, and security.

[0368] The aforementioned embodiments are progressive: Embodiments 1 to 9 explain the basic principles and system architecture, Embodiment 10 provides the mathematical derivation of five core technologies, and Embodiment 11 verifies the engineering implementation and performance.

[0369] Example 12 systematically analyzes parameter configuration, and Example 13 uses real events to connect the end-to-end process. The five technologies in different examples corroborate and support each other, together forming a complete technical solution.

[0370] The above embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection. Those skilled in the art can make equivalent substitutions or partial adjustments to the parameters, structures, and combinations of the various technical features within the scope of the claims.

[0371] For example, 1D-CNN can be replaced by GRU, weighted scalarization can be replaced by NSGA-II, and FedAvg can be replaced by FedProx. All technical solutions that do not depart from the spirit of this invention fall within the protection scope of this invention.

[0372] The sequence of algorithm steps involved in this invention can be adjusted without changing the technical effect, the hardware configuration involved can be increased or decreased according to the scenario, and the formula involved can be transformed and expressed under the premise of mathematical equivalence.

[0373] This invention can be widely applied to scenarios such as industrial IoT, smart cities, vehicle networking, and smart homes, providing lightweight, adaptive, and explainable security management capabilities for large-scale heterogeneous IoT devices.

[0374] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An IoT security management method based on artificial intelligence, characterized in that, The method includes the following steps: S1. An identification agent deployed at the edge node of the Internet of Things (IoT) collects and identifies anomalies in real time from the behavioral data of IoT terminal devices. The identification agent adopts a resource-aware dual-model cascaded dynamic switching architecture. The architecture includes a first-level one-dimensional convolutional neural network initial screening model and a second-level structured pruned bidirectional long short-term memory network confirmation model. When the screening result of the first-level model is a suspected anomaly, the second-level model is triggered to perform accurate classification. The pruning rate of the second-level model is determined by Pareto optimality based on the real-time CPU utilization, available memory, and bandwidth margin of the edge node. Pareto optimality uses a weighted scaling method to transform the maximization of classification accuracy and the minimization of computational resource consumption into a single-objective optimization problem. S2. A decision-making agent deployed at the regional collaboration layer constructs a spatiotemporal risk transmission field. The spatiotemporal risk transmission field includes a three-layer risk situation map of the device layer, network layer, and application layer. The risk transmission volume between layers is calculated using cross-layer risk entropy and exponential time-varying decay weights. The cross-layer risk entropy uses the Shannon entropy form to quantify the uncertainty of the risk distribution of each layer, that is, taking the negative pair of the risk probability distribution of each layer. The risk entropy value of the layer is obtained by weighted summation. The risk transmission between layers is determined by the product of the source layer risk entropy change, the exponential time-varying decay weight, and the inter-layer coupling coefficient. When the risk situation of any layer exceeds the dynamic threshold, the preventive response of the adjacent layer is automatically triggered. S3: The response agent deployed in the cloud adopts a Nash equilibrium multi-objective game scheduling strategy for hierarchical security scheduling. The risk level and resource cost are modeled as a non-cooperative game and the Nash equilibrium is solved as the optimal scheduling strategy, outputting scheduling instructions for L1 local handling, L2 regional collaboration, or L3 cloud global. S4: The online incremental causal... The graph module provides interpretability annotation for decision-making. It maintains a dynamic causal graph by combining Granger causality tests and counterfactual reasoning. The weight of each edge in the dynamic causal graph is equal to the product of the Granger causality score and the counterfactual reasoning score, multiplied by a time decay factor. The Granger causality score is obtained by calculating the predicted improvement between variables using a vector autoregression model. The counterfactual reasoning score is obtained by the change in the observed results under the intervention hypothesis. The causal confidence propagation algorithm is used to propagate the confidence score forward according to the directed acyclic graph structure to achieve real-time incremental annotation of causal relationships. S5. The federated transfer learning module performs federated aggregation of model parameters among the identification agents. The identification agents, decision agents, and response agents interact through asynchronous communication based on a message queue protocol. During the federated aggregation process, only the model parameter gradients are uploaded, not the original behavioral data, and the gradients are encrypted using a secure aggregation protocol. The newly deployed edge nodes obtain the pre-trained model parameters from the regional collaboration layer and perform fine-tuning and adaptation locally.

2. The IoT security management method based on artificial intelligence according to claim 1, characterized in that, In step S1, the inference latency of the first-level one-dimensional convolutional neural network initial screening model is no more than 3 milliseconds, and it contains no less than 3 layers and no more than 8 layers of one-dimensional convolutional blocks. Each convolutional block includes a one-dimensional convolutional layer, a batch normalization layer, and an activation function layer in sequence. The second-level structured pruned bidirectional long short-term memory network confirmation model adopts a knowledge distillation compensation strategy, which obtains soft labels from the unpruned teacher model for distillation compensation, so that the classification accuracy loss at a pruning rate of 0.4 is controlled within 1 percentage point. Optionally, the pruning rate ranges from 0.2 to 0.6, and the number of hidden layer units of the second-level model is an integer between 64 and 256. In one embodiment, when the CPU utilization of the edge node exceeds 85%, the pruning rate of the second-level model is automatically increased to above 0.5 to release computing resources, or a hybrid pruning strategy combining channel pruning and weight sparsification is adopted.

3. The IoT security management method based on artificial intelligence according to claim 2, characterized in that, In step S1, the pruning rate is determined through Pareto optimality as follows: a Pareto front is constructed with the dual objectives of maximizing classification accuracy and minimizing computational resource consumption. A weighted scaling method is used to transform the dual objectives into a single-objective optimization problem, where the sum of the accuracy weight and the resource consumption weight is equal to 1. The CPU utilization, available memory, and bandwidth margin are normalized and then weighted to obtain a resource awareness score. When the resource awareness score is lower than a preset lower threshold, the pruning rate of the second-level model is automatically increased to release computational resources. When the resource awareness score is higher than a preset upper threshold, the pruning rate is automatically decreased to improve accuracy. Optionally, the NSGA-II multi-objective evolutionary algorithm is used instead of the weighted scaling method to solve the Pareto front, or a reinforcement learning agent is used to adjust the accuracy weight and the resource consumption weight online.

4. The IoT security management method based on artificial intelligence according to claim 1, characterized in that, In step S2, the risk situation maps of the device layer, network layer, and application layer are constructed from sensor data of the corresponding layers after feature extraction. Each layer of the risk situation map represents an asset entity as a node and the relationship between assets as an edge, and assigns a risk entropy to the node and a transmission coefficient to the edge. The cross-layer risk entropy is calculated in the form of Shannon entropy to determine the uncertainty of the risk distribution of each layer. The risk transmission between layers is determined by the product of the source layer risk change, the time-varying attenuation weight, and the inter-layer coupling coefficient. Optionally, the inter-layer coupling coefficient ranges from 0 to 1, and the coupling coefficient from the device layer to the network layer is greater than the coupling coefficient from the network layer to the application layer; in one embodiment, the three-layer risk situation map aligns entity and relationship definitions with a unified ontology model, or uses a graph convolutional network method to vectorize the nodes.

5. The IoT security management method based on artificial intelligence according to claim 4, characterized in that, In step S2, the time-varying decay weight decays exponentially with increasing time interval. The decay form is an exponential function with the natural constant as the base and the product of the negative time interval and the decay coefficient as the exponent. The decay coefficient ranges from 0.01 to 0.

2. The dynamic threshold is determined by the weighted sum of the moving average and standard deviation of the historical risk entropy of the layer. When the real-time risk entropy exceeds the dynamic threshold, the risk transmission of the layer is set to zero through an adaptive gating function to block the transmission. Optionally, the adaptive gating function adopts an S-shaped function or a step function. In one embodiment, the window length of the moving average is 50 to 200 sampling periods.

6. The IoT security management method based on artificial intelligence according to claim 1, characterized in that, In step S3, the utility function of the Nash equilibrium multi-objective game scheduling strategy is defined as the risk-averse payoff minus the weighted sum of resource costs and response delays. The sum of the weights of the risk-averse payoff, resource costs, and response delays equals 1. Each player uses a regret-matching iterative algorithm to solve the hybrid strategy Nash equilibrium, with the number of iterations ranging from 100 to 1000. The triggering conditions for the L1 local handling, L2 regional coordination, and L3 cloud-based global three-level scheduling instructions are determined by risk level thresholds. When the overall risk level is lower than the first risk threshold, an output is generated. The L1 local handling instruction outputs an L2 regional coordination instruction when the overall risk level is between the first risk threshold and the second risk threshold, and an L3 cloud-based global instruction when the overall risk level is higher than the second risk threshold. Optionally, the first risk threshold is between 0.3 and 0.5, and the second risk threshold is between 0.7 and 0.

85. In one embodiment, after the scheduling instruction is issued, a maximum response time limit is set. If the handling is not completed within the maximum response time limit, the scheduling is automatically upgraded to the next level. The maximum response time limit is between 30 seconds and 300 seconds.

7. The IoT security management method based on artificial intelligence according to claim 1, characterized in that, In step S4, the Granger causality score is obtained by calculating the predicted improvement between variables using a vector autoregression model, the counterfactual reasoning score is obtained by the change in observed results under the intervention hypothesis, the time decay factor ranges from 0.8 to 0.99, and the causal confidence propagation algorithm performs forward propagation of confidence according to the directed acyclic graph structure of the causal graph, resolving causal conflicts through domain prior knowledge when they are detected. In step S5, the message queue protocol is the MQTT protocol, the identification agent subscribes to the behavioral data collection topic and the model update topic, and the federated aggregation uses a federated averaging algorithm with an aggregation round value of 10 to 100 rounds. Optionally, a federated proximal optimization algorithm is used to suppress local fine-tuning bias, or a Bayesian model averaging method is used to fuse conflicting causal hypotheses.

8. An Internet of Things (IoT) security management system based on artificial intelligence, characterized in that, include: The intelligent agent identification module is deployed at the edge node of the Internet of Things. It adopts a resource-aware dual-model cascaded dynamic switching architecture to collect and identify the behavior data of the Internet of Things terminal devices in real time, and outputs the first-level abnormal events and their initial risk levels. The decision-making intelligent agent module, deployed at the regional collaboration layer, contains a risk transmission field engine and a cross-layer judgment submodule. The risk transmission field engine stores three types of entities—device layer, network layer, and application layer—and their relationships using a time-series graph database. The cross-layer judgment submodule calculates the inter-layer risk transmission volume based on cross-layer risk entropy and exponential time-varying decay weights, and outputs the final risk level. When the risk transmission volume exceeds a dynamic threshold, a risk blocking mechanism is activated. The response intelligent agent module, deployed in the cloud, uses a Nash equilibrium multi-objective game scheduling strategy for hierarchical security scheduling and outputs scheduling instructions. The online incremental causal graph module provides interpretable annotations for decision-making; the federated transfer learning module performs federated aggregation of model parameters among various identification agents. The situation visualization module reads real-time risk situation data from the unified situation map and generates a visualization interface, which includes a device topology view, a risk heat map, a causal link diagram of handling, and a historical situation playback panel.

9. The IoT security management system based on artificial intelligence according to claim 8, characterized in that, A dynamic switching controller is provided between the first-level one-dimensional convolutional neural network initial screening model and the second-level structured pruned bidirectional long short-term memory network confirmation model of the identification agent module. The dynamic switching controller calculates a resource awareness score based on the real-time CPU utilization, available memory, and bandwidth margin of the edge nodes. When the resource awareness score is lower than a preset threshold, the pruning rate of the second-level model is automatically increased to release computing resources. Optionally, the dynamic switching controller also incorporates temperature sensor data and power margin data as auxiliary inputs for the resource awareness score. In one embodiment, the dynamic switching controller polls the resource status and updates the resource awareness score at millisecond intervals, or uses an event-driven mechanism to trigger a score update when the resource status changes by more than a set amount.

10. The AI-based IoT security management system according to claim 8, characterized in that, The system is deployed on an electronic device including at least one processor and a memory with a communication connection. The memory stores a computer program that can be executed by the at least one processor. When executed by the at least one processor, the computer program implements the functions of the identification agent module, decision agent module, response agent module, online incremental causal graph module, federated transfer learning module, and situational visualization module. The electronic device also includes a communication interface that supports data interaction between the IoT edge node, regional collaboration layer, and cloud via the MQTT protocol. Optionally, the electronic device is any one of an edge gateway, a regional server, or a cloud server. The memory includes non-volatile storage media and volatile storage media. In one embodiment, the computer program is embedded in the non-volatile storage medium as firmware, and the situation visualization module is rendered in a scalable vector graphics format and supports multi-terminal adaptive display.

Images