An IoT secure data transmission and routing control system

This IoT secure data transmission system, which combines physical layer feature zero-knowledge proofs and spatiotemporal graph neural networks, solves the problem of routing metric parameter forgery in IoT dynamic routing control. It enables rapid identification and isolation of attacking nodes, reduces computational and communication overhead, and is adaptable to resource-constrained IoT nodes.

CN123074831APending Publication Date: 2026-11-10HUNAN WEIZHENGYI TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611457245.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-18
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing IoT dynamic routing control systems fail to effectively verify the authenticity of routing metric parameters, leading to nodes being attacked and having their routing parameters forged, resulting in traffic black holes or data leaks. Furthermore, existing security solutions have excessive computational and communication overhead, making them unsuitable for resource-constrained IoT nodes.

Method used

A physical layer feature zero-knowledge proof verification subsystem is adopted to bind physical layer features and routing metric parameters to generate authentication labels. Combined with a spatiotemporal graph neural network, anomaly detection and multipath microstate injection transmission are performed to achieve low-overhead routing parameter verification and secure data transmission.

Benefits of technology

It enables the rapid identification and isolation of nodes with spoofed routing parameters with low computational and communication overhead, preventing traffic black holes and data leakage, reducing the risk of side-channel analysis attacks, and adapting to resource-constrained IoT nodes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074831A_ABST
    Figure CN123074831A_ABST
Patent Text Reader

Abstract

This invention discloses a secure data transmission and routing control system for the Internet of Things (IoT), belonging to the field of secure communication technology for IoT. The system includes a physical layer feature zero-knowledge proof verification subsystem, a spatiotemporal graph neural network anti-poisoning routing control subsystem, and a multipath micro-state injection transmission subsystem. The former binds the physical layer channel state information features of the terminal node with the network layer routing metric parameters to generate an authentication label, which is verified at the edge controller to obtain the confidence level, realizing the authenticity verification without decrypting the original physical layer data. The middle subsystem fuses graph attention space aggregation and temporal coding to obtain spatiotemporal joint embedding, calculates the abnormal deviation score based on the confidence level, and performs reinforcement learning routing decision after immediately isolating nodes that exceed the threshold. The latter subsystem divides the data into fragments, transmits them concurrently along multiple non-overlapping paths at pseudo-random time intervals, and then reassembles them at the destination node.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of secure communication technology for the Internet of Things (IoT), and specifically to a secure data transmission and routing control system for the Internet of Things. Background Technology

[0002] With the large-scale deployment of industrial IoT, vehicle-to-everything (V2X) and smart grids, IoT networks are characterized by a large number of nodes, highly dynamic topologies, and limited communication and computing resources. To adapt to rapid topology changes, existing technologies generally introduce dynamic routing control schemes based on deep reinforcement learning. Edge controllers centrally collect routing metrics such as remaining battery power, link latency, and cache utilization reported periodically by each terminal node, encode them into network state feature vectors, and input them into a deep reinforcement learning agent. The agent then outputs the next-hop routing decision and converts it into a forwarding flow table, which is then distributed to each node for data forwarding. For transmission security, existing solutions typically perform end-to-end encryption of the entire data content before forwarding to protect data confidentiality.

[0003] The existing solutions described above have the following drawbacks: First, the state coding process does not verify the authenticity of the routing metric parameters reported by the terminal nodes. Once a terminal node is physically captured or logically controlled, attackers can forge false high-priority routing parameters to mislead routing decisions, creating traffic black holes or diverting business data to malicious nodes. Second, the routing strategy is optimized solely based on the global reward function, lacking cross-verification of the consistency of the historical behavior of individual nodes and its spatiotemporal evolution with the local network topology, making it difficult to identify covert poisoning attacks in a timely manner. Third, the data content itself is encrypted without processing the external traffic characteristics of the data packets. Attackers can infer the device's operating status through side-channel traffic analysis, resulting in privacy leaks.

[0004] Furthermore, while existing security measures such as blockchain consensus and heavy firewalls can improve system security to some extent, their computational and communication overhead far exceeds the actual capacity of IoT terminal nodes with limited resources and low-latency business scenarios, making it difficult to scale up for deployment in industrial IoT. Therefore, how to verify the authenticity of routing metric parameters, quickly identify and isolate hidden poisoning nodes, and effectively hide external traffic characteristics during transmission while ensuring low computational and communication overhead has become a pressing technical problem to be solved in the field of IoT secure routing. Summary of the Invention

[0005] The technical problem this invention aims to solve is that existing AI-based IoT dynamic routing control systems lack authenticity verification of node-reported routing metric parameters. This allows physically captured or logically controlled internal Byzantine nodes to forge minute but deceptive routing parameters, enabling covert data poisoning attacks on the routing controller. This can induce traffic black holes in the system or divert business data to malicious eavesdropping nodes. Existing routing decision models lack cross-validation of the consistency of node historical behavior and the spatiotemporal evolution of local topology, making it difficult to identify covert poisoning behavior in a timely manner. Existing secure transmission schemes only encrypt data content without processing external traffic characteristics of data packets. Attackers can infer the device's operating status through side-channel traffic analysis. Furthermore, existing heavy-duty defense methods have excessive computational and communication overhead, making them unsuitable for the actual constraints of limited IoT node resources.

[0006] To address the aforementioned technical problems, this invention provides an IoT secure data transmission and routing control system, comprising a physical layer feature zero-knowledge proof verification subsystem, a spatiotemporal graph neural network anti-poisoning routing control subsystem, and a multipath microstate injection transmission subsystem.

[0007] The physical layer feature zero-knowledge proof verification subsystem is deployed on IoT terminal nodes and edge controllers. It is used to collect local physical layer channel state information features on the terminal node side and perform quantization encoding. It combines the collected routing metric parameters and the current timestamp to generate a commitment value that binds the physical layer features and routing metric parameters. Based on the commitment value and a random number, it generates an authentication tag and sends a routing request message containing the routing metric parameters, the random number, the authentication tag and quantization auxiliary error correction data to the edge controller.

[0008] The edge controller collects locally measured channel state information reference features, uses the quantization-assisted error correction data to perform error correction decoding to obtain candidate quantization features, combines the candidate quantization features, the routing metric parameters and the current timestamp to generate candidate labels, and outputs a confidence level characterizing the authenticity of the routing metric parameters by comparing the candidate labels with the received authentication labels.

[0009] The spatiotemporal graph neural network anti-poisoning routing control subsystem is deployed on the edge controller. It is used to construct a dynamic graph based on the network topology connection relationship, the routing metric parameters reported by each terminal node, and the confidence level. It uses a graph attention network to aggregate neighbor node information layer by layer to obtain node spatial embedding. The node spatial embeddings of multiple consecutive historical moments are input into a gated recurrent unit to obtain the spatiotemporal joint embedding of the node. Based on the difference between the spatiotemporal joint embedding and the historical normal behavior baseline embedding of the node, and combined with the confidence level, the abnormal deviation score of the node is calculated. Nodes with scores exceeding a preset threshold are isolated from the candidate next-hop node set. For nodes that are not isolated, a reinforcement learning strategy network is used to output a safe routing decision.

[0010] The multipath microstate injection transmission subsystem is deployed at the sending node, relay node, and destination node. It is used to divide the data packet to be transmitted into multiple data fragments and perform random filling processing. It generates a transmission time slot corresponding to each data fragment according to a pseudo-random number generator. According to the pseudo-random mapping relationship, it allocates each data fragment to multiple non-intersecting secure paths output by the secure routing decision. It sends each data fragment concurrently according to the transmission time slot. At the destination node, it decrypts, sorts, aggregates, and reassembles the arriving fragments to restore the original data packet.

[0011] The beneficial effects of this invention are:

[0012] 1. By binding physical layer features and routing metric parameters to generate authentication tags and comparing and verifying them on the edge controller side, the authenticity of routing metric parameters can be verified without decrypting the original physical layer data of the terminal node. The verification process only involves hash operations and error correction decoding, with low computational overhead and adaptability to resource-constrained IoT terminal nodes.

[0013] 2. By fusing graph attention space aggregation features and temporal features to calculate anomaly deviation scores and combining them with confidence correction, it is possible to identify and isolate covert poisoning nodes that continuously forge small-amplitude routing parameters within a millisecond to second timescale, thus preventing malicious parameters from participating in routing decisions for a long time.

[0014] 3. By segmenting data into fragments and transmitting them concurrently along multiple non-overlapping secure paths with pseudo-random time slots, traffic observers on a single link cannot obtain complete and regular packet size and transmission time characteristics, thereby blocking side-channel analysis attacks based on traffic characteristics. Compared with heavy defense solutions such as blockchain consensus, this significantly reduces computational and communication overhead. Attached Figure Description

[0015] Figure 1 This is a flowchart of the physical layer feature zero-knowledge proof verification subsystem of the present invention;

[0016] Figure 2 This is a flowchart of the spatiotemporal graph neural network anti-poisoning routing control subsystem of the present invention;

[0017] Figure 3 This is a flowchart of the multipath microstate injection transmission subsystem of the present invention. Detailed Implementation

[0018] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0019] I. System Overall Architecture

[0020] The IoT secure data transmission and routing control system provided by this invention is generally divided into three layers: the terminal node layer, the edge control layer, and the cloud training layer, forming a cross-layer security architecture that coordinates cloud, edge, and terminal. In the terminal node layer, each IoT terminal node is equipped with a channel state information acquisition module, a routing metric parameter acquisition module, and an authentication tag generation module. During data transmission, the terminal node corresponds to a sending node, a relay node, or a destination node, depending on its role. There is no essential difference in hardware configuration among the three; the only difference lies in the functions it performs in specific transmission tasks. The edge control layer is deployed on a software-defined network controller or edge gateway and includes a proof verification module, a spatiotemporal graph neural network routing decision module, and a multipath transmission scheduling module. The cloud training layer is equipped with an offline training and periodic model distribution process for the spatiotemporal graph neural network model, which is used to continuously provide updated inference model parameters to the edge control layer.

[0021] In this embodiment, confidence is used as one of the input features of the spatiotemporal graph neural network routing control subsystem to construct the dynamic graph, and the secure routing decision result is used as the basis for the multipath transmission scheduling module to allocate paths, thereby forming a tight data coupling and cross-layer linkage relationship between the physical layer, network layer, control layer and transport layer.

[0022] II. Specific Implementation Methods of the Physical Layer Zero-Knowledge Proof Verification Subsystem

[0023] On the terminal node side, node i collects the original feature vector of local physical layer channel state information. It is mapped to a binary feature string using a preset quantization function:

[0024] ;

[0025] in For feature quantization function, This is the original feature vector of the channel state information collected at node i. This is the binary feature string obtained after quantization.

[0026] Node i uses an error correction coding function to generate auxiliary error correction data from the quantized bit string:

[0027] ;

[0028] in For error correction coding functions, To assist in error correction data, it is used to tolerate channel state information bit errors caused by differences in measurement time or measurement location between the verifier and the node.

[0029] Node i reads the current network layer routing metric parameter vector (Includes remaining battery power, cache utilization, and average forwarding latency components) and the current timestamp. Combined with a pre-set symmetric key Calculate the binding commitment value:

[0030] ;

[0031] in For cryptographic hash functions, the symbol This represents a bit concatenation operation. This is a commitment value for binding physical layer features with routing metric parameters.

[0032] Node i generates random numbers Calculate the authentication label:

[0033] ;

[0034] in To authenticate the label, node i sends a routing request message to neighboring nodes or the edge controller. The message content includes... , , , , It does not contain the original channel state information vector. Quantization bit string and key This enables verifiability without disclosing the original physical layer privacy data.

[0035] On the edge controller side, the verifier uses its own measured channel state information reference vector. Performing the same quantization operation, we obtain:

[0036] ;

[0037] And utilize the received auxiliary error correction data right Error correction decoding is performed to correct bit differences introduced by channel measurement noise, resulting in candidate quantized bit strings. The verifier combines the key held locally. (and Calculate candidate commitments and candidate tags (either obtained through key negotiation or by correspondence):

[0038] ;

[0039] ;

[0040] The verifier compares candidate tags With the received tag Are they consistent? If they are consistent, determine the routing metric parameters reported by node i. If it matches its actual physical layer state, output the first confidence level. If there is a discrepancy, a second confidence level lower than the first confidence level is output, and the node is marked as a suspicious node. In one specific implementation, the first confidence level is 1 and the second confidence level is 0. However, those skilled in the art will understand that the specific value of the confidence level is not limited to the above numerical form, as long as it can characterize the relative level of the authenticity of the routing metric parameter. The verification party will set the confidence level... The node identifier and timestamp are reported to the edge controller as auxiliary confidence feature inputs for subsequent spatiotemporal graph neural network models to perform anomaly detection.

[0041] It should be noted that the physical layer feature binding method based on channel state information described above is suitable for mobile or semi-mobile topology scenarios with rich channel multipath features and high distinguishability. In fixed topology IoT scenarios where the channel state is relatively stable and the channel feature distinguishability is insufficient, device fingerprints based on physically non-cloning functions can be used to replace or supplement channel state information features. That is, the unique fingerprint of the device is generated by utilizing the inherent process randomness of the terminal node hardware circuit, and the fingerprint and routing metric parameters are processed according to the same binding commitment and authentication tag generation process described above. This can also achieve the verification of the authenticity of the routing metric parameters without relying on channel measurement conditions and without decrypting the original device fingerprint data.

[0042] III. Specific Implementation Methods of the Spatiotemporal Graph Neural Network Anti-Poisoning Routing Control Subsystem

[0043] The edge controller periodically collects the network topology connections and routing metric parameters reported by each terminal node. Combined with the confidence level output by the aforementioned proof verification module , building time Animated GIF ,in For a set of nodes, Let be the set of edges at that moment. The Behavioral node feature vectors, containing routing metric parameters Confidence level and the historical behavioral statistical characteristics of node i .

[0044] For the picture For each node i, a graph attention network is used to aggregate neighbor information layer by layer. The layer node representation is calculated as follows:

[0045] ;

[0046] in Describe the set of one-hop neighbors of node i. For the first Layer learnable weight matrix, It is a non-linear activation function. For the first Layer nodes The expression .

[0047] Attention weight The following results were obtained by scoring and normalizing the node representations:

[0048] ;

[0049] in For the first The layer can learn attention parameter vector, symbol This represents the vector concatenation operation. Indicates to Perform the transpose operation.

[0050] Place node i in a continuous Spatial embedding sequence of historical moments Inputting a temporal coding unit yields a joint embedding that fuses spatiotemporal information:

[0051] ;

[0052] in For gated cyclic unit timing coding function, For node i at time... Spatiotemporal joint embedding.

[0053] Based on joint embedding Embedded with the historical normal behavior baseline of this node Calculate the anomaly deviation score for node i:

[0054] ;

[0055] in Represents the L2 norm operation. As shown in the above formula, the abnormal deviation score is simultaneously affected by the weighted correction of the deviation magnitude of spatiotemporal behavior and the confidence level. The lower the confidence level, the higher the score is obtained under the same deviation magnitude. That is, for nodes that have been marked as suspicious in the physical layer verification process, their abnormal behavior score will be amplified accordingly, thereby realizing the joint decision-making of physical layer verification results and network layer behavior analysis results.

[0056] like ,in If a node is identified as a suspected poisoning node, the edge controller will immediately isolate it from the candidate next-hop node set and will no longer accept its reported routing parameters for subsequent decision-making. Otherwise, the node will remain in the candidate next-hop node set.

[0057] For the set of candidate next-hop nodes that have not been isolated, the routing policy network Embedded in the current spacetime Given the state as input, output the value estimate of each candidate next-hop action:

[0058] ;

[0059] And select the action with the highest value as the secure routing decision:

[0060] ;

[0061] in As a candidate for the next jump action, To select the optimal action, the controller uses a comprehensive reward function:

[0062] ;

[0063] Policy network parameters Perform gradient updates, where This is the end-to-end delay for the path. The average energy consumption of the path nodes. This represents the mean confidence score of each node along the path. , , In one specific implementation, to pre-set weighting coefficients, The value is 0.4. The value is 0.3. The value is set to 0.3, meaning that end-to-end latency and average energy consumption of path nodes have a negative contribution to the comprehensive reward function, while the mean path confidence has a positive contribution to the comprehensive reward function. This guides the policy network to select forwarding paths composed of high-confidence nodes while optimizing latency and energy consumption indicators, thereby achieving continuous online optimization of routing decisions.

[0064] In another specific implementation, the combined structure of graph attention network and gated recurrent unit can be replaced by a spatiotemporal self-attention network structure. The self-attention mechanism is used to simultaneously model the spatial adjacency relationship and historical time series of nodes. By replacing the graph attention aggregation coefficient with a multi-head self-attention weight matrix, the consistency of node historical behavior and the spatiotemporal evolution law of local topology can be cross-validated. Furthermore, it can be combined with federated learning to collaboratively train model parameters among multiple edge controllers without having to aggregate the raw sensitive data collected by each edge controller to the cloud training layer, thus supplementing the centralized cloud training method.

[0065] IV. Specific Implementation Methods of the Multipath Microstate Injection Transmission Subsystem

[0066] The edge controller selects the candidate path based on the evaluation results output by the aforementioned spatiotemporal graph neural network module. These non-intersecting secure paths form a set of transmission paths, denoted as . .

[0067] The sending node will transmit sensitive data packets. Divided into segments according to the preset segmentation strategy A data fragment, denoted as Each fragment is randomly padded to eliminate any discernible pattern in the original data packet size. The sending node uses a pseudo-random number generator seeded with a pre-shared key to generate a sequence of transmission time slots corresponding to each fragment, denoted as . This is used to disrupt the regularity of fragment transmission.

[0068] Sending nodes for each fragment Assign a serial number identifier that has undergone symmetric encryption. Only the destination node can decrypt it, based on a pseudo-random mapping function. Will Assign to a path in the path set:

[0069] ;

[0070] in The path number assigned to this fragment. This represents the modulo operation, where the sending node operates according to the time slot corresponding to each fragment. Fragments are sequentially sent along the assigned paths, so that the traffic observed on any single path exhibits uniform pseudo-random characteristics in both packet size distribution and transmission frequency distribution.

[0071] When each fragment is forwarded by a relay node along its assigned path, the relay node only performs regular hop-by-hop forwarding and is unaware of the complete data transmission task to which the packet belongs, thus preventing any single-point observer from obtaining a complete traffic profile.

[0072] The destination node receives from After decrypting the fragments along the path, decrypt the sequence number identifier of each fragment. The system then caches and sorts the arriving fragments according to their sequence numbers. The destination node uses a lightweight aggregation function to reassemble the sorted fragments and restore the original data packet.

[0073] ;

[0074] in This is a lightweight aggregation and reassembly function. If some fragments fail to arrive within the set timeout window, a retransmission request mechanism for the corresponding path is triggered to ensure the complete and reliable delivery of data.

[0075] V. Overall Closed-Loop Working Process of the System

[0076] In one specific embodiment of the present invention, the overall system operation process is as follows: the terminal node collects channel state information and routing metric parameters and generates authentication tags, sends routing request messages to the edge controller, the edge controller verifies the authentication tags and outputs confidence scores, the confidence scores and routing metric parameters are jointly input into the dynamic graph construction, graph attention aggregation and temporal coding stages, after abnormal deviation scoring and threshold judgment, some nodes are isolated and the candidate next-hop node set is updated, the nodes that are not isolated output secure routes through reinforcement learning routing decisions, the sending nodes perform data fragmentation, pseudo-random time slots and path allocation according to the secure routes and send concurrently, the destination node completes decryption, sorting and aggregation and recombination, restores the original data, the recombination results and path transmission quality statistics are further fed back to the online reward update stage of the policy network, forming a complete closed loop from physical layer feature collection to online policy optimization. The above closed loop operation process enables the system to continuously adapt to the dynamic changes in IoT network topology and node behavior, and achieves both routing decision reliability assurance and transmission confidentiality protection while ensuring low computational and communication overhead.

[0077] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. An Internet of Things (IoT) secure data transmission and routing control system, characterized in that, This includes a physical layer feature zero-knowledge proof verification subsystem, a spatiotemporal graph neural network anti-poisoning routing control subsystem, and a multipath microstate injection transmission subsystem; The physical layer feature zero-knowledge proof verification subsystem is deployed on IoT terminal nodes and edge controllers. It is used to collect local physical layer channel state information features on the terminal node side and perform quantization encoding. It generates a commitment value that binds the physical layer features and the routing metric parameters by combining the currently collected routing metric parameters and the current timestamp. It generates an authentication tag based on the commitment value and a random number. It sends a routing request message containing the routing metric parameters, the random number, the authentication tag, and quantization auxiliary error correction data to the edge controller. On the edge controller side, it collects the locally measured channel state information reference features and uses the quantization auxiliary error correction data to perform error correction decoding to obtain candidate quantization features. It generates candidate tags by combining the candidate quantization features, the routing metric parameters, and the current timestamp. It outputs a confidence level that characterizes the authenticity of the routing metric parameters by comparing the candidate tags with the received authentication tags. The spatiotemporal graph neural network anti-poisoning routing control subsystem is deployed on the edge controller. It is used to construct a dynamic graph based on the network topology connection relationship, the routing metric parameters reported by each terminal node, and the confidence level. For each node in the dynamic graph, the node spatial embedding is obtained by aggregating the neighbor node information layer by layer using a graph attention network. The node spatial embeddings of multiple consecutive historical moments are input into a gated recurrent unit to obtain the spatiotemporal joint embedding of the node. The abnormal deviation score of the node is calculated based on the difference between the spatiotemporal joint embedding and the historical normal behavior baseline embedding of the node, combined with the confidence level. Nodes with abnormal deviation scores exceeding a preset threshold are isolated from the candidate next-hop node set. For nodes in the candidate next-hop node set that are not isolated, a reinforcement learning policy network is used to output a safe routing decision based on the spatiotemporal joint embedding. The multipath microstate injection transmission subsystem is deployed at the sending node, relay node, and destination node. It is used to divide the data packet to be transmitted into multiple data fragments and randomly fill each data fragment. It generates a transmission time slot corresponding to each data fragment according to a pseudo-random number generator, and allocates each data fragment to multiple non-intersecting secure paths output by the secure routing decision according to the pseudo-random mapping relationship. It sends each data fragment along the allocated secure path according to the transmission time slot. The relay node performs hop-by-hop forwarding of the data fragments it passes through. After receiving the data fragments from the multiple secure paths, the destination node decrypts the encrypted sequence number of each data fragment and sorts them according to the sequence number. It then uses an aggregation and reassembly function to reassemble the sorted data fragments back into the original data packet. The confidence level serves as the input feature for the spatiotemporal graph neural network anti-poisoning routing control subsystem to construct the dynamic graph, and the secure routing decision serves as the path basis for the multipath microstate injection transmission subsystem to allocate data fragments.

2. The system according to claim 1, characterized in that, The physical layer zero-knowledge proof verification subsystem generates commitment values ​​on the terminal node side in the following manner: The commitment value is obtained by concatenating the quantized physical layer features, routing metric parameters, current timestamp, and preset symmetric key, followed by a hash operation. The authentication label is generated by concatenating the commitment value with a random number and then performing a hash operation.

3. The system according to claim 2, characterized in that, After obtaining the candidate quantization features, the edge controller concatenates the candidate quantization features, the routing metric parameters, the current timestamp, and the locally held key, and then performs a hash operation to obtain the candidate commitment value. The candidate commitment value is then concatenated with the received random number and then performed a hash operation to obtain the candidate label. If the candidate label matches the received authentication label, the routing metric parameter is determined to be consistent with the actual physical layer state of the terminal node and the first confidence level is output. If they are inconsistent, the second confidence level is output and the corresponding terminal node is marked as a suspicious node, where the first confidence level is greater than the second confidence level.

4. The system according to claim 1, characterized in that, The node feature vector of the dynamic graph includes the routing metric parameters, confidence level, and historical behavior statistical features of the corresponding terminal node. When the graph attention network aggregates neighbor node information layer by layer, it calculates attention weights based on the current layer node representation and the neighbor node representations, and then performs a weighted summation of the neighbor node representations based on the attention weights before passing them through a nonlinear activation function to obtain the next layer node representation.

5. The system according to claim 4, characterized in that, The anomaly deviation score of a node is the product of the norm of the difference between the node's spatiotemporal joint embedding and the historical normal behavior baseline embedding, and a correction coefficient determined by the confidence level, wherein the lower the confidence level, the larger the correction coefficient.

6. The system according to claim 1, characterized in that, The reinforcement learning policy network updates its parameters based on a comprehensive reward function composed of end-to-end latency, average energy consumption of path nodes, and average confidence of each node on the path. In the comprehensive reward function, the weight coefficients corresponding to end-to-end latency and average energy consumption of path nodes are both negative, while the weight coefficients corresponding to the average confidence of each node on the path are positive.

7. The system according to claim 1, characterized in that, The multipath microstate injection transmission subsystem allocates secure paths to each data fragment in the following manner: The security path number assigned to a data fragment is obtained by performing a modulo operation on the index of the data fragment using a pseudo-random mapping function.

8. The system according to claim 7, characterized in that, If the destination node does not receive all data fragments within the set timeout window, it triggers a retransmission request to the corresponding secure path. The encrypted sequence number is generated by the sending node using a symmetric encryption method, and only the destination node can decrypt it.

9. A method for secure data transmission and routing control in the Internet of Things, characterized in that, include: The terminal node collects local physical layer channel state information features and performs quantization encoding. It combines the collected routing metric parameters and the current timestamp to generate a commitment value that binds the physical layer features and routing metric parameters. Based on the commitment value and a random number, it generates an authentication tag and sends a routing request message containing the routing metric parameters, the random number, the authentication tag, and quantization auxiliary error correction data to the edge controller. The edge controller collects locally measured channel state information reference features, uses the quantization-assisted error correction data to perform error correction decoding to obtain candidate quantization features, combines the candidate quantization features, the routing metric parameters and the current timestamp to generate candidate labels, and outputs a confidence level characterizing the authenticity of the routing metric parameters by comparing the candidate labels with the received authentication labels. The edge controller constructs a dynamic graph based on the network topology, routing metric parameters reported by each terminal node, and the confidence level. For each node in the dynamic graph, the neighbor node information is aggregated layer by layer using a graph attention network to obtain the node spatial embedding. The node spatial embeddings of multiple consecutive historical moments are input into a gated recurrent unit to obtain the spatiotemporal joint embedding of the node. Based on the difference between the spatiotemporal joint embedding and the historical normal behavior baseline embedding of the node, and combined with the confidence level, the abnormal deviation score of the node is calculated. Nodes with abnormal deviation scores exceeding a preset threshold are isolated from the candidate next-hop node set. For nodes in the candidate next-hop node set that are not isolated, a reinforcement learning policy network is used to output a safe routing decision based on the spatiotemporal joint embedding. The sending node divides the data packet to be transmitted into multiple data fragments and performs random filling on each data fragment. It generates a transmission time slot corresponding to each data fragment according to a pseudo-random number generator. Based on the pseudo-random mapping relationship, it allocates each data fragment to multiple non-intersecting secure paths output by the secure routing decision. It then sends each data fragment along the allocated secure path according to the transmission time slot. The relay node performs hop-by-hop forwarding on the data fragments it passes through. After receiving the data fragments from the multiple secure paths, the destination node decrypts the encrypted sequence number identifier of each data fragment and sorts them according to the sequence number. It then uses an aggregation and reassembly function to reassemble the sorted data fragments back into the original data packet.

10. The method according to claim 9, characterized in that, Also includes: The edge controller updates the parameters of the reinforcement learning policy network based on a comprehensive reward function composed of end-to-end latency, average energy consumption of path nodes, and average confidence of each node on the path.