Wireless communication adaptive security protection method and system based on edge computing

Abstract

The present application relates to the technical field of wireless communication protection, in particular to a wireless communication adaptive security protection method and system based on edge computing; physical layer channel state information when an Internet of Things terminal device accesses a network gateway for the first time, functional area semantic identification of the network gateway and a current timestamp are captured to generate an anonymous session credential; operation trajectory data of the anonymous session credential is tracked to construct a trajectory operation segment; based on trajectory uniqueness, spatial information sensitivity and behavior correlation analysis of the trajectory operation segment, a privacy risk assessment value is obtained; when the privacy risk assessment value exceeds a preset risk threshold, a trajectory disconnection mechanism is triggered; based on privacy protection processing and the trajectory disconnection mechanism, privacy-enhanced trajectory data is obtained; a virtual data generation model is trained through the privacy-enhanced trajectory data; after the virtual data generation model converges, a virtual data generator in the virtual data generation model is driven to generate a virtual operation and maintenance data stream, thereby realizing adaptive protection of wireless communication data.

Description

Technical Field

[0001] This invention relates to the field of wireless communication protection technology, specifically to an adaptive security protection method and system for wireless communication based on edge computing. Background Technology

[0002] In modern computing technology, processing, analyzing, and utilizing the massive, high-speed, real-time operational data streams generated by distributed IoT scenarios is a core issue. Current mainstream computing architectures typically follow a centralized processing model, aggregating raw data from various IoT terminals through data acquisition interfaces to a central server or cloud data center to build a large-scale database; for example, a digital twin supporting the operation of a smart factory, where upper-layer applications perform unified computation and analysis. However, centralized data processing architectures inherently suffer from data security deficiencies when handling highly sensitive IoT data.

[0003] First, from a computing system architecture perspective, centrally storing all raw IoT data creates a massive and complex data volume. Performing global queries, correlation analysis, and model training on this data volume requires enormous computing resources and storage I / O, leading to longer response times for computing tasks and making it difficult to meet the real-time control and decision-making needs of IoT scenarios. More importantly, this architecture creates a single point of failure risk in data security. A highly centralized raw database containing complete operational details of the factory would be catastrophic if illegally accessed.

[0004] To address these challenges, existing data processing methods attempt to preprocess data before input, such as using hash-based or static pseudonym-based algorithms to anonymize IoT terminal identifier fields in data records. However, these algorithms only make superficial modifications to the data structure and do not alter the inherent, deep statistical relationships between data records. Advanced data analysis algorithms, such as pattern recognition and sequence mining, can still inversely infer key, high-value industrial processes or equipment operating status by analyzing the behavioral characteristics, temporal relationships, and contextual associations of this anonymized data. This indicates that existing data processing procedures lack the ability to protect the deep semantic privacy of IoT data at the algorithmic level.

[0005] To address this, an adaptive security protection method and system for wireless communication based on edge computing is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive security protection method and system for wireless communication based on edge computing. The method trains a virtual data generation model using privacy-enhancing trajectory data. After the virtual data generation model converges, it drives the virtual data generator to generate a virtual operation and maintenance data stream, thereby achieving adaptive protection of wireless communication data.

[0007] To achieve the above objectives, the present invention provides an adaptive security protection method for wireless communication based on edge computing, comprising:

[0008] The system captures the physical layer channel state information, the functional area semantic identifier of the network gateway, and the current timestamp when the IoT terminal device first accesses the network gateway, and generates anonymous session credentials based on hash operations.

[0009] The operation trajectory data of the anonymous session credentials is tracked in real time, and trajectory operation segments are constructed; the privacy risk assessment model is called to identify the trajectory operation segments, and the privacy risk assessment value is obtained based on the trajectory uniqueness, spatial information sensitivity and behavioral correlation analysis of the trajectory operation segments;

[0010] The trajectory operation segments are processed for privacy protection, and the privacy risk assessment value is further judged by a threshold. When the preset risk threshold is exceeded, the trajectory disconnection mechanism is triggered. Based on the privacy protection processing and the trajectory disconnection mechanism, privacy-enhanced trajectory data is obtained.

[0011] The preset virtual data generation model is trained using privacy-enhanced trajectory data; after the virtual data generation model converges, the virtual data generator within it is driven to produce a virtual operation and maintenance data stream equivalent to the real production and operation and maintenance mode.

[0012] The steps for constructing the trajectory operation segment include: maintaining a real-time state machine for each anonymous session credential in the edge computing node, and dynamically defining the start and end points of the trajectory operation segment by listening to business events;

[0013] The triggering conditions for the business events include: the wireless signal of the IoT terminal device switching from one network gateway to another network gateway representing a different functional area, constituting a region crossing event; the data interaction between the IoT terminal device and the control interface of the key industrial facility through a preset application layer protocol, constituting a device interaction event; and the preset state transition occurring when the internal operating status field is parsed from the telemetry data reported by the IoT terminal device, constituting a state change event.

[0014] When a business event is triggered, the state machine terminates the trajectory operation segment before the trigger and generates a new segment; the trajectory operation segment contains a spatiotemporal coordinate sequence, a business event type code used to identify the reason for termination, and associated metadata.

[0015] The identification process of the privacy risk assessment model includes static risk assessment, dynamic risk assessment, and risk fusion decision-making.

[0016] The static risk assessment receives the current position coordinates of the trajectory operation segment and queries the functional area confidentiality map stored in the edge computing node; the functional area confidentiality map is a key-value pair data structure, where the key is the polygon coordinates of the spatial region, the value is the pre-calculated static risk score, and the matching static risk score is output.

[0017] The dynamic risk assessment extracts the trajectory uniqueness, spatial information sensitivity, and behavioral correlation analysis features of the trajectory operation segment to form a multi-dimensional behavioral feature vector; then the multi-dimensional behavioral feature vector is input into the pre-trained dynamic risk identification module to output a normalized dynamic risk score.

[0018] The risk fusion decision receives static risk scores and dynamic risk scores, and weights them to obtain a privacy risk assessment value.

[0019] The process of determining the static risk score in the functional area confidentiality map includes:

[0020] The IoT environment is divided into spatial regions, and a functional assessment is conducted based on the physical purpose and business functions of each spatial region.

[0021] The functional assessment includes: the confidentiality level of the data generated and processed by the equipment within the space area in the data classification standard; the critical path weight of the process links carried by the space area in the overall production flow chart; and the stringency level of the physical security and network access control strategies deployed in the space area.

[0022] Quantitative weighting coefficients are configured for functional assessment, and static risk scores are calculated for each spatial region.

[0023] The trajectory uniqueness is determined by calculating the continuous spatiotemporal coordinates of the trajectory operation segment and mapping them to a discretized spatial region grid to obtain a discrete state sequence; based on historical anonymous trajectory data, a group behavior baseline model including a state transition probability matrix and a trajectory sequence module is constructed; the discrete state sequence is identified through the group behavior baseline model to obtain a trajectory uniqueness score, thereby quantifying the trajectory uniqueness;

[0024] The spatial information sensitivity is calculated by comparing the dwell time of the trajectory operation segment with the duration probability density function defined by the standard operating procedure of the functional area, and calculating the Z-score of the anomaly.

[0025] The behavioral association analysis calculates the cumulative association weight of the node sequence path traversed by the trajectory operation segment in a process flow knowledge graph that uses equipment, regions, and materials as nodes, in order to quantify the inferability of its behavior.

[0026] For all constructed trajectory operation segments, the basic processing module applies privacy protection processing, including: for position coordinates, performing coordinate normalization operation to align with the center point of the preset grid; for telemetry values, calculating the Laplace distribution parameters according to the preset privacy budget and adding corresponding noise.

[0027] When the privacy risk assessment value of the trajectory operation segment exceeds the preset risk threshold, the enhanced processing module initiates the trajectory disconnection mechanism on top of the basic processing; the trajectory disconnection mechanism generates a session update credential for the IoT terminal device and performs subsequent trajectory recognition.

[0028] The steps for training the virtual data generation model include:

[0029] The privacy-enhancing trajectory data is read into the volatile memory of the edge computing node in the form of a data stream. After each backpropagation gradient calculation is completed, the data in the volatile memory is marked as erasable and overwritten by the memory management unit. Furthermore, in the step of gradient aggregation to update model weights, Gaussian noise calculated based on the global privacy budget and gradient sensitivity is added to the aggregated gradient vector to achieve provable privacy protection defined by differential privacy.

[0030] An edge computing-based adaptive security protection system for wireless communication includes:

[0031] The session credential generation module captures the physical layer channel state information, the functional area semantic identifier of the network gateway, and the current timestamp when the IoT terminal device first accesses the network gateway, and generates anonymous session credentials based on hash operations.

[0032] The privacy risk assessment module tracks the operation trajectory data of the anonymous session credentials in real time and constructs trajectory operation segments; it calls the privacy risk assessment model to identify the trajectory operation segments, and obtains the privacy risk assessment value based on the trajectory uniqueness, spatial information sensitivity and behavioral correlation analysis of the trajectory operation segments;

[0033] The privacy data processing module performs privacy protection processing on trajectory operation segments and further performs threshold judgment on privacy risk assessment values. When the preset risk threshold is exceeded, the trajectory chain disconnection mechanism is triggered. Based on the privacy protection processing and the trajectory chain disconnection mechanism, privacy-enhanced trajectory data is obtained.

[0034] The virtual data generation module trains a preset virtual data generation model using privacy-enhanced trajectory data. After the virtual data generation model converges, it drives the virtual data generator within it to produce a virtual operation and maintenance data stream equivalent to the real production and operation and maintenance mode.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] 1. This invention introduces a business event-driven state machine to construct trajectory operation segments, achieving intelligent and semantic segmentation of the original data stream. Compared to traditional fixed-window segmentation methods, the segments generated by this solution have inherent logical integrity, with each segment corresponding to a specific and complete business operation. This significantly improves the accuracy of subsequent privacy risk assessments because the assessment model can analyze high-quality data units with clear business contexts, thereby more accurately identifying abnormal behavior and potential privacy leaks, and avoiding information fragmentation and misjudgment of risks caused by improper data segmentation.

[0037] 2. This solution proposes a dual-channel privacy risk assessment model, combining pre-calculated static spatial risks with real-time calculated dynamic behavioral risks. This ensures that the risk assessment possesses both stability and objectivity based on the physical environment, as well as sensitivity and adaptability to specific behavioral patterns. By using a functional area confidentiality map, the prior knowledge of domain experts is embedded into the system, guaranteeing the rationality of the risk assessment baseline. Meanwhile, the dynamic risk identification module can capture subtle but crucial behavioral anomalies that static rules cannot cover. The weighted fusion of these two approaches makes the final risk assessment results more comprehensive, accurate, and robust, providing a reliable decision-making basis for subsequent adaptive privacy protection strategies.

[0038] 3. This invention constructs a three-dimensional and in-depth behavioral risk profile by quantifying three orthogonal dimensions: trajectory uniqueness, spatial information sensitivity, and behavioral correlation analysis. This dynamic risk assessment method based on multi-dimensional feature fusion greatly enhances the system's ability to identify advanced and covert privacy threats, making privacy protection no longer a blind blurring of data, but a precise strike against specific risk points, thus maximizing the usability of data while protecting privacy. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the adaptive security protection method for wireless communication based on edge computing according to the present invention.

[0040] Figure 2 This is a data logic diagram of the edge computing-based adaptive security protection method for wireless communication according to the present invention.

[0041] Figure 3 This is a schematic diagram of the structure of the edge computing-based wireless communication adaptive security protection system of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Example 1:

[0044] This invention proposes an adaptive security protection method for wireless communication based on edge computing, the process of which is as follows: Figure 1 As shown, its data logic is as follows: Figure 2 As shown, it specifically includes:

[0045] The system captures the physical layer channel state information, the functional area semantic identifier of the network gateway, and the current timestamp when the IoT terminal device first accesses the network gateway, and generates anonymous session credentials based on hash operations.

[0046] The operation trajectory data of the anonymous session credentials is tracked in real time, and trajectory operation segments are constructed; the privacy risk assessment model is called to identify the trajectory operation segments, and the privacy risk assessment value is obtained based on the trajectory uniqueness, spatial information sensitivity and behavioral correlation analysis of the trajectory operation segments;

[0047] The trajectory operation segments are processed for privacy protection, and the privacy risk assessment value is further judged by a threshold. When the preset risk threshold is exceeded, the trajectory disconnection mechanism is triggered. Based on the privacy protection processing and the trajectory disconnection mechanism, privacy-enhanced trajectory data is obtained.

[0048] The preset virtual data generation model is trained using privacy-enhanced trajectory data; after the virtual data generation model converges, the virtual data generator within it is driven to produce a virtual operation and maintenance data stream equivalent to the real production and operation and maintenance mode.

[0049] Preferably, the steps for constructing the trajectory operation segment include: maintaining a real-time state machine for each anonymous session credential in the edge computing node, and dynamically defining the start and end points of the trajectory operation segment by listening to business events;

[0050] This solution defines three types of core business events, and the triggering conditions for these business events include:

[0051] When an IoT terminal device's wireless signal switches from one network gateway to another representing a different functional area, it constitutes a region crossing event. When an IoT terminal, such as an inspection robot, switches its wireless signal from one wireless network gateway to another, the edge controller receives a handover signal. Since each gateway is pre-assigned a functional area semantic identifier, this handover constitutes a region crossing event with a clear business meaning.

[0052] Data interactions between IoT terminal devices and the control interfaces of critical industrial facilities via preset application-layer protocols constitute device interaction events. In industrial IoT scenarios, terminal devices often communicate with critical industrial facilities, such as PLCs, robotic arm controllers, and CNC machine tools, through application-layer protocols such as OPC-UA and MQTT. Edge nodes can monitor these interactions by deploying lightweight protocol parsing probes. When a specific command, such as "read the pressure value of a high-pressure vessel" or "start the conveyor belt," is detected, a device interaction event is constituted.

[0053] Parsing telemetry data reported by IoT terminal devices reveals a transition in internal operating status fields to a preset state, constituting a state change event. Many IoT terminals periodically report their own telemetry data, such as temperature, humidity, and vibration frequency. Edge nodes can set rules so that when a critical status field undergoes a drastic change or crosses a preset threshold, such as a motor temperature transitioning from normal to alarm, a state change event is constituted.

[0054] When a business event is triggered, the state machine terminates the trajectory operation segment before the trigger and generates a new segment; the trajectory operation segment contains a spatiotemporal coordinate sequence, a business event type code used to identify the reason for termination, and associated metadata.

[0055] The associated metadata includes the preceding and following regions traversed, information about the interacting devices, and the degree of change in telemetry data.

[0056] This invention introduces a business event-driven state machine to construct trajectory operation segments, achieving intelligent and semantic segmentation of the original data stream. Compared to traditional fixed-window segmentation methods, the segments generated by this solution possess inherent logical integrity, with each segment corresponding to a specific and complete business operation. This significantly improves the accuracy of subsequent privacy risk assessments because the assessment model can analyze high-quality data units with clear business contexts, thereby more accurately identifying abnormal behavior and potential privacy breaches, and avoiding information fragmentation and misjudgment of risks caused by improper data segmentation.

[0057] Preferably, the identification process of the privacy risk assessment model includes static risk assessment, dynamic risk assessment, and risk fusion decision-making;

[0058] The static risk assessment receives the current position coordinates of the trajectory operation segment and queries the functional area confidentiality map stored in the edge computing node; the functional area confidentiality map is a key-value pair data structure, where the key is the polygon coordinates of the spatial region, the value is the pre-calculated static risk score, and the matching static risk score is output.

[0059] The dynamic risk assessment extracts the trajectory uniqueness, spatial information sensitivity, and behavioral correlation analysis features of the trajectory operation segment to form a multi-dimensional behavioral feature vector; then the multi-dimensional behavioral feature vector is input into the pre-trained dynamic risk identification module to output a normalized dynamic risk score.

[0060] The risk fusion decision receives static risk scores and dynamic risk scores, and weights them to obtain a privacy risk assessment value.

[0061] Preferably, the process of determining the static risk score in the functional area confidentiality map includes:

[0062] The IoT environment is divided into spatial regions, and a functional assessment is conducted based on the physical purpose and business functions of each spatial region.

[0063] The functional assessment includes: the confidentiality level of the data generated and processed by the equipment within the space area in the data classification standard; the critical path weight of the process links carried by the space area in the overall production flow chart; and the stringency level of the physical security and network access control strategies deployed in the space area.

[0064] Data confidentiality level: Based on the company's internal data security management standards (such as secret, confidential, top secret, general), the data types that are routinely generated or processed in this area are classified. For example, the process parameter data of the R&D laboratory is "top secret", while the temperature and humidity telemetry data of the warehouse is "general".

[0065] Critical path weight: In the overall production flow diagram of the factory (such as a directed acyclic graph), analyze the process stage in which this area is located. If this stage is a node on the critical path that affects the final output, its weight is high.

[0066] Physical and cybersecurity level: Assess the existing security measures in the area. For example, are there access control systems and surveillance cameras? Is the network physically isolated? Are strict network access controls deployed? The higher the security level, the greater its importance.

[0067] Quantitative weighting coefficients were configured for the functional assessment, and a static risk score was calculated for each spatial region. These quantitative weighting coefficients were determined using the analytic hierarchy process (AHP) in conjunction with expert review.

[0068] Preferably, the dynamic risk assessment step further includes a cross-document spatiotemporal correlation analysis submodule;

[0069] The workflow of the submodule is as follows: a real-time anonymous session credential proximity graph is maintained in the edge computing node. When two or more different anonymous session credentials appear simultaneously within a preset time window and spatial distance threshold, an associated edge is established for them in the graph. The submodule continuously monitors the generation pattern of the edges in the graph and compares it with normal collaborative behavior. When a credential combination pattern that has never occurred in history or has an extremely low frequency appears, it is determined as an abnormal collaboration event. A significant penalty term is superimposed on the dynamic risk score of the current trajectory operation segment of all anonymous credentials participating in the event to reflect the higher-dimensional potential privacy leakage risk generated by multi-agent collaboration.

[0070] For example, if two anonymous credentials that should be working on different production lines co-occur for an extended period in a non-public area (such as a maintenance corridor), the system will identify this as an abnormal collaboration event and add a significant penalty to the dynamic risk score of the current trajectory operation segment of all anonymous credentials involved in the event, in order to reflect the higher-dimensional potential privacy leakage risk generated by multi-agent collaboration.

[0071] This invention, building upon individual behavioral risk assessment, adds a monitoring dimension for abnormal behavior in multi-agent collaboration, significantly enhancing the breadth and depth of privacy risk assessment. Traditional privacy protection methods typically focus only on the behavioral trajectory of a single device, failing to identify more covert information theft or process inference attacks accomplished collaboratively by multiple devices. For example, analyzing the collaborative trajectories of multiple devices may reveal a critical, confidential process requiring multi-tasking coordination. This upgrade from "individual risk" to "relationship risk" assessment effectively defends against more advanced, correlation-based attacks, filling a gap in multi-agent privacy security and providing more comprehensive and multi-dimensional security protection for complex industrial scenarios.

[0072] Furthermore, this solution proposes a dual-channel privacy risk assessment model, creatively combining pre-calculated static spatial risks with real-time calculated dynamic behavioral risks. This structure enables the risk assessment to possess both physical environment-based stability and objectivity, as well as sensitivity and adaptability to specific behavioral patterns. Through a "functional area confidentiality map," the prior knowledge of domain experts is embedded into the system, ensuring the rationality of the risk assessment benchmark. Meanwhile, the dynamic risk identification module can capture subtle but crucial behavioral anomalies that static rules cannot cover. The weighted fusion of these two approaches makes the final risk assessment results more comprehensive, accurate, and robust, providing a reliable decision-making basis for subsequent adaptive privacy protection strategies.

[0073] The trajectory uniqueness is determined by calculating the continuous spatiotemporal coordinates in the trajectory operation segment and mapping them to a discretized spatial region grid to obtain a discrete state sequence; based on historical anonymous trajectory data, a group behavior baseline model including a state transition probability matrix and a trajectory sequence module is constructed; the discrete state sequence is identified through the group behavior baseline model to obtain a trajectory uniqueness score, and the trajectory uniqueness is quantified.

[0074] The process of identifying discrete state sequences by the group behavior baseline model includes:

[0075] From the state transition probability matrix of the baseline model, query the historical probability corresponding to each state transition in the discrete state sequence, and calculate the probability of all steps according to the Shannon entropy formula to obtain an entropy value that characterizes the unpredictability of the trajectory path.

[0076] In the trajectory sequence module of the baseline model, the historical probability of the discrete state sequence as a whole and its decomposed N-gram subsequences is queried, and the reciprocal or negative logarithm of the probability is taken to obtain a rarity score that characterizes the rarity of the trajectory pattern.

[0077] The calculated information entropy value and the rarity score are subjected to maximum and minimum normalization, and then weighted and summed using preset weights to form a single, quantified trajectory uniqueness score.

[0078] This invention, by weighting and fusing these two dimensions, can accurately distinguish different types of anomalies, avoiding misjudgments that may arise from single-dimensional assessments. For example, it can treat a rare but structurally simple normal maintenance path as the same as a complex and disordered potentially malicious detection path. This refined measurement makes subsequent privacy risk assessment decisions more comprehensive and accurate, enabling more effective identification of advanced and covert privacy threats. It maximizes data availability while precisely targeting privacy risks.

[0079] Furthermore, the baseline model for group behavior is specifically a first-order Markov model. This model is constructed by statistically analyzing at least 30 days of historical anonymous trajectory data. First, the physical space is divided into a discrete grid of 10 meters by 10 meters as the state space. Then, the transition frequency of all trajectories between the grids is statistically analyzed, and the state transition probability matrix is ​​calculated. The trajectory sequence module is a hash table, where the key is a trajectory subsequence in the form of N-gram (preferably, N=3), and the value is the frequency of the subsequence in the historical data.

[0080] The spatial information sensitivity is calculated by comparing the dwell time of the trajectory operation segment with the duration probability density function defined by the standard operating procedure of the functional area, and then calculating the Z-score of the anomaly.

[0081] The duration probability density function approximates a Gaussian distribution, i.e., the mean duration and standard deviation. For a trajectory segment, its dwell time in a certain region is calculated. Then, the Z-score is calculated. The larger the absolute value of the Z-score, the higher the degree of anomaly in the dwell time, and the higher the spatial information sensitivity score.

[0082] The behavioral association analysis calculates the cumulative association weight of the node sequence path traversed by the trajectory operation segment in a process flow knowledge graph that uses equipment, regions, and materials as nodes, in order to quantify the inferability of its behavior.

[0083] The SoData Process Knowledge Graph is a semantic network stored in the form of a graph database, describing the various entities and their relationships in the production process. Nodes in the graph include: equipment, areas, and materials; edges represent the relationships between them. Each edge can be assigned a weight, representing the importance or sensitivity of that relationship within the overall process. Edge types include "interacting with," "located in," and "transporting," and the weight of each edge is initialized by domain experts based on its importance within the process.

[0084] A trajectory operation segment is mapped to a path on a knowledge graph. The behavioral association analysis score of this trajectory segment is the cumulative association weight of all edges on that path. A path connecting multiple high-weight nodes means that the behavior reveals a key link in the core production process, its inferability is strong, and its association analysis score is high.

[0085] This invention constructs a comprehensive and in-depth behavioral risk profile by quantifying three orthogonal dimensions: trajectory uniqueness, spatial information sensitivity, and behavioral correlation analysis. This dynamic risk assessment method based on multi-dimensional feature fusion greatly enhances the system's ability to identify advanced and covert privacy threats, making privacy protection no longer a blind blurring of data, but a precise strike against specific risk points, maximizing data usability while protecting privacy.

[0086] Furthermore, the aforementioned process flow knowledge graph can also be configured with dynamic self-learning and evolution capabilities;

[0087] Specifically, the implementation involves running an association rule mining module in the background. This module takes anonymous trajectory operation fragments as input transactions and periodically uses algorithms such as Apriori or FP-Growth to mine frequently occurring itemsets across space, time, equipment type, and business events. When the confidence and support of the mined association rule exceed preset thresholds, the system abstracts this rule into a new weighted edge or updates the weights of existing edges, automatically injecting it into the process flow knowledge graph. This enables the knowledge graph to learn autonomously and reflect the long-term evolution or short-term changes in the factory's operating model, maintaining its timeliness in relation to current business logic.

[0088] This invention addresses the fundamental problem of rigid knowledge bases in traditional behavioral association analysis, which are unable to adapt to changes in business processes, by introducing a dynamic knowledge graph self-learning mechanism. In modern flexible manufacturing environments, production processes, equipment layouts, and operating procedures are frequently adjusted, causing static knowledge graphs to quickly become outdated and leading to a sharp decline in the accuracy of risk assessment. This solution, through background association rule mining, enables the knowledge graph to have self-evolution capabilities, continuously and automatically discovering and absorbing new, high-confidence behavioral patterns and entity associations from real data streams. This ensures that the "knowledge benchmark" upon which behavioral association analysis relies remains synchronized with the real world, greatly improving the adaptability and long-term effectiveness of risk assessment.

[0089] Preferably, for all constructed trajectory operation segments, the basic processing module applies privacy protection processing, including: for position coordinates, performing coordinate normalization operation to align with the center point of a preset grid; for telemetry values, calculating the Laplace distribution parameters according to a preset privacy budget and adding corresponding noise.

[0090] When the privacy risk assessment value of the trajectory operation segment exceeds the preset risk threshold, the enhanced processing module initiates the trajectory disconnection mechanism on top of the basic processing; the trajectory disconnection mechanism generates a session update credential for the IoT terminal device and performs subsequent trajectory recognition.

[0091] The coordinate normalization is a position generalization technique. For example, the physical space is divided into a 10m x 10m grid, and any coordinate point falling into this grid is uniformly processed as the coordinates of the center point of the grid.

[0092] Laplace noise is a type of random noise with a specific form that satisfies the definition of differential privacy. Its distribution is peak-shaped and can provide mathematically provable privacy protection for data while changing the statistical characteristics of the data less than Gaussian noise.

[0093] This invention follows the security principles of minimal privilege and defense in depth, which means that all data is assumed to have a certain risk and requires basic protection; for high-risk data that has been identified, stronger measures must be taken to isolate and block it.

[0094] Basic privacy protection: All completed trajectory operation segments, regardless of their risk assessment value, are sent to the basic processing module. This module performs two operations: for each location coordinate in its spatiotemporal coordinate sequence, it performs coordinate normalization to reduce position accuracy. For the telemetry data it carries, it calculates the scale parameter of the Laplace distribution based on a preset global privacy budget, and samples a noise value from this distribution to add to the original value.

[0095] Enhanced privacy protection: After basic processing, the system compares the privacy risk assessment value of the trajectory segment with the preset risk threshold. If the privacy risk assessment value is greater than the preset risk threshold, enhanced processing is triggered, and the trajectory disconnection mechanism is activated.

[0096] Trajectory Disconnection Execution: The core of this mechanism is to destroy the current identity credentials and force the device to re-authenticate. Specifically, the edge node can send a control command to the target IoT terminal, forcing it to disconnect from the current network gateway and immediately reconnect. At the moment the device initiates the reconnection request, the edge node recaptures its instantaneous Physical Layer Channel State Information (CSI), current timestamp, and functional area identifier, and repeats the hash operation to generate a completely new session update credential, unrelated to the previous one. All subsequent trajectories will be bound to this new credential, thus logically severing the connection between high-risk behavior and its subsequent actions.

[0097] This invention establishes a dynamic, hierarchical privacy protection system, achieving an optimal balance between resource efficiency and protection strength. The basic processing module ensures all data meets fundamental privacy compliance requirements, forming a security baseline. Meanwhile, a risk threshold-based trajectory disconnection mechanism proactively cuts off attackers' ability to profile and track behavior over long periods upon detecting high-risk actions. This adaptive strategy avoids overprotecting all data, thus providing robust privacy security while maximizing the analytical value of low-risk data.

[0098] Preferably, the steps for training the virtual data generation model include:

[0099] The privacy-enhancing trajectory data is read into the volatile memory of the edge computing node in the form of a data stream. After each backpropagation gradient calculation is completed, the data in the volatile memory is marked as erasable and overwritten by the memory management unit. Furthermore, in the step of gradient aggregation to update model weights, Gaussian noise calculated based on the global privacy budget and gradient sensitivity is added to the aggregated gradient vector to achieve provable privacy protection defined by differential privacy.

[0100] The virtual data generation model is a generative machine learning model whose goal is to learn the inherent distribution patterns of real data and generate entirely new data with similar statistical characteristics but completely fictitious content. In this embodiment, a generative adversarial network (GAN) is employed, particularly a variant designed for processing time-series data, such as a recurrent GAN or a temporal GAN.

[0101] Volatile memory, also known as RAM, contains data that exists only while power is on and is lost when power is off.

[0102] Differential privacy is a strong privacy protection model that provides a mathematical framework to quantify the risk of privacy breaches. Its core idea is to introduce a suitable amount of randomness into the algorithm's output, making it impossible for attackers to determine whether a specific individual's data exists in the original input dataset simply by observing the output.

[0103] The global privacy budget is a key differential privacy parameter used to control the total amount of privacy consumed throughout the training process. The smaller the global privacy budget, the stronger the privacy protection.

[0104] Model architecture and training process:

[0105] A recurrent GAN is used as the core of the virtual data generation model architecture;

[0106] Generator: Recurrent neural network, which takes a random noise vector as input and outputs sequence data with a format and length similar to real privacy-enhanced trajectory data.

[0107] Discriminator: Another recurrent neural network that receives a trajectory data, which may be real or faked by the generator. Its task is to determine its authenticity and output a probability value.

[0108] Design principles and process:

[0109] Memory safety: All privacy-enhancing trajectory data obtained through the preceding steps will not be written to disk. The data loader reads them directly into the volatile memory of the edge computing node in batches as a data stream for model training.

[0110] Training iterations: In each training iteration, the generator produces a batch of fake data, while the discriminator simultaneously receives this batch of fake data and a batch of real data retrieved from memory, and makes a judgment. Then, the loss is calculated based on the discriminator's judgment, and the network weights of the generator and discriminator are updated using the backpropagation algorithm.

[0111] Immediate sample destruction: Once a batch of real data has completed its computational tasks in both forward and backward propagation—that is, gradient calculation is complete—the memory manager immediately marks this memory area as erasable or overwriteable. This means that the use of this batch of real data in the model is strictly limited to a single gradient update, with an extremely short lifespan, greatly reducing the risk of accidental data leakage.

[0112] Differential privacy injection employs a differential privacy stochastic gradient descent algorithm during the backpropagation stage for updating model weights. After aggregating the gradients of all samples in a batch to form the final updated gradient vector, the system adds carefully calculated Gaussian noise to this aggregated gradient vector. The variance of this noise depends on parameters such as the global privacy budget and gradient sensitivity (controlled by gradient clipping). This step is crucial for achieving the mathematical guarantees of differential privacy, ensuring that the final trained model parameters do not overfit or memorize any specific real-world trajectory information.

[0113] This invention achieves data minimization and zero persistence principles from both physical and logical perspectives by strictly confining the model training process to volatile memory and combining it with a "use-and-destroy" strategy for samples, significantly reducing the risk of data leakage during data storage and processing. More importantly, by injecting differential privacy noise into the gradient updates of the training algorithm, it provides a mathematically provable and rigorous privacy protection commitment for the final generated model. Even if the model itself is stolen and analyzed, attackers cannot reverse engineer any individual information about the original training data, ensuring that the final output virtual operation and maintenance data stream is not only statistically usable but also achieves the highest harmony between data utilization and privacy protection.

[0114] Example 2:

[0115] This invention proposes an edge computing-based adaptive security protection system for wireless communication, deployed on the edge computing server of a smart factory. It aims to protect the privacy and analyze the operational data of various wireless IoT terminals within the factory, such as automated guided vehicles (AGVs), inspection drones, and handheld terminals. The following section uses the operation of an AGV-007 as an example to explain in detail the collaborative workflow of each module in the system.

[0116] The structure of the system is as follows Figure 3 As shown, it specifically includes:

[0117] The session credential generation module captures the physical layer channel state information, the functional area semantic identifier of the network gateway, and the current timestamp when the IoT terminal device first accesses the network gateway, and generates anonymous session credentials based on hash operations.

[0118] Process Start Point: When AGV-007 first enters the "Workshop-Raw Material Area" from the dormant area and connects to the wireless network gateway of this area, the system's session credential generation module is activated. Detailed Implementation:

[0119] Feature capture and extraction: The physical layer channel state information (CSI) of AGV-007 at the moment of access is captured through the wireless interface of the network gateway. To ensure the uniqueness and repeatability of the features, the module extracts the amplitude values ​​on all 52 subcarriers and arranges them into a one-dimensional floating-point vector.

[0120] Multi-source information fusion and hash operation: The module strings the above vector and concatenates it in an ordered manner with the pre-configured functional area semantic identifier (FIDE) of the network gateway and the current precise Unix timestamp. The concatenation format is "CSI feature string | FIDE identifier | timestamp". Subsequently, the module uses the SHA-256 hash algorithm to calculate a unique long string after concatenation, generating a 256-bit hexadecimal string. This string is then established as the anonymous session credential for AGV-007 during this session period for subsequent tracking.

[0121] The privacy risk assessment module tracks the operation trajectory data of the anonymous session credentials in real time and constructs trajectory operation segments; it calls the privacy risk assessment model to identify the trajectory operation segments, and obtains the privacy risk assessment value based on the trajectory uniqueness, spatial information sensitivity and behavioral correlation analysis of the trajectory operation segments.

[0122] AGV-007 begins to move within the factory and perform tasks, while the system's privacy risk assessment module tracks the location and status data corresponding to its anonymous session credentials in real time.

[0123] Construction of trajectory operation segments: The module maintains a real-time state machine for this credential. When AGV-007 travels from the "raw material area" to the "first production line," and its wireless signal switches from the gateway of the raw material area to the gateway of the production line, this "area crossing event" is detected by the state machine. The state machine immediately encapsulates the continuous trajectory data (including spatiotemporal coordinate sequence, associated metadata, etc.) before the switch (within the raw material area) into the first trajectory operation segment and begins recording new segments.

[0124] For the newly generated raw material area trajectory operation segment, the module calls the privacy risk assessment model:

[0125] Static Risk Assessment: The module extracts the location information of the fragment and queries the built-in "Functional Area Confidentiality Map". The map shows that the confidentiality level of the "Raw Material Area" is average, and its pre-calculated static risk score is 0.4.

[0126] Dynamic risk assessment: The module calculates three dynamic features from the fragment: trajectory uniqueness (whether the path is rare), spatial information sensitivity (whether the dwell time is abnormal), and behavioral correlation analysis (whether it is associated with key process nodes). These three feature values ​​are input into a pre-trained dynamic risk identification module; the dynamic risk identification module is a multilayer perceptron neural network containing two hidden layers with 16 and 8 neurons respectively, using the ReLU activation function. The dynamic risk identification module outputs a normalized dynamic risk score, for example, 0.5.

[0127] Risk fusion: The module calculates the weighted sum of the two scores based on the weights predetermined by the analytic hierarchy process to obtain the final privacy risk assessment value for the trajectory segment.

[0128] The privacy data processing module performs privacy protection processing on trajectory operation segments and further performs threshold judgment on privacy risk assessment values. When the preset risk threshold is exceeded, the trajectory disconnection mechanism is triggered. Based on the privacy protection processing and the trajectory disconnection mechanism, privacy-enhanced trajectory data is obtained.

[0129] Basic privacy protection: Basic protection is applied to the trajectory operation segments. The position coordinates of the points are normalized and uniformly aligned to the center of a 10*10 meter grid. At the same time, for the telemetry data (such as motor load) carried within, the Laplace distribution parameters are calculated and corresponding noise is added based on the privacy budget adaptively set for the current area (raw material area, low static risk).

[0130] Enhanced privacy protection: The risk score is compared with a dynamically set risk threshold. This threshold is derived from the 95th percentile of all risk values ​​over the past 24 hours. If the risk score is below the threshold, enhanced protection is not triggered. If the risk score is exceeded, the trajectory disconnection mechanism is immediately activated. A forced re-authentication command is sent to AGV-007 via the network controller, causing it to disconnect and immediately reconnect to the network. At the moment of reconnection, the session credential generation module generates a completely new, anonymous session credential unrelated to the previous one, thus completely severing the tracking link between high-risk behavior and subsequent behavior.

[0131] The virtual data generation module trains a preset virtual data generation model using privacy-enhanced trajectory data. After the virtual data generation model converges, it drives the virtual data generator within it to produce a virtual operation and maintenance data stream equivalent to the real production and operation and maintenance mode.

[0132] All privacy-enhanced trajectory data obtained through the privacy data processing module will not be directly stored or uploaded to the cloud, but will be sent to the virtual data generation module.

[0133] This privacy-enhancing data is directly read into the volatile memory of the edge server in the form of a data stream to train a virtual data generation model. In this embodiment, the virtual data generation model is a recurrent generative adversarial network (RGAN). Both the generator and discriminator of this RGAN consist of two-layer LSTM networks, enabling it to process time-series data. During the gradient update phase of training, the system injects Gaussian noise calculated based on a global privacy budget to achieve differential privacy protection. Each batch of data is immediately erased from memory after completing one gradient calculation.

[0134] The generation and application of virtual data streams: Once the virtual data generation model training converges and can generate trajectory data that is highly similar to the real operation and maintenance mode in statistical characteristics (such as the average speed of AGVs, path heatmaps, and dwell time distribution), the system stops using real data for training. Thereafter, it drives the trained generator to continuously generate high-quality, completely fictitious virtual operation and maintenance data streams. These virtual data streams can be securely uploaded to the cloud data center for upper-layer applications to perform global production efficiency analysis, bottleneck diagnosis, and process optimization, eliminating the risk of leakage of real individual trajectory data at the source.

[0135] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive security protection method for wireless communication based on edge computing, characterized in that, include: The system captures the physical layer channel state information, the functional area semantic identifier of the network gateway, and the current timestamp when the IoT terminal device first accesses the network gateway, and generates anonymous session credentials based on hash operations. Real-time tracking of the operation trajectory data of the anonymous session credentials, and construction of trajectory operation segments; The privacy risk assessment model is invoked to identify the trajectory operation segments. Based on the trajectory uniqueness, spatial information sensitivity, and behavioral correlation analysis of the trajectory operation segments, a privacy risk assessment value is obtained. The trajectory operation segments are processed for privacy protection, and the privacy risk assessment value is further judged by a threshold. When the preset risk threshold is exceeded, the trajectory disconnection mechanism is triggered. Based on the privacy protection processing and the trajectory disconnection mechanism, privacy-enhanced trajectory data is obtained. The preset virtual data generation model is trained using privacy-enhanced trajectory data; after the virtual data generation model converges, the virtual data generator within it is driven to produce a virtual operation and maintenance data stream equivalent to the real production and operation and maintenance mode.

2. The adaptive security protection method for wireless communication based on edge computing according to claim 1, characterized in that: The steps for constructing the trajectory operation segment include: maintaining a real-time state machine for each anonymous session credential in the edge computing node, and dynamically defining the start and end points of the trajectory operation segment by listening to business events; The triggering conditions for the business events include: the wireless signal of the IoT terminal device switching from one network gateway to another network gateway representing a different functional area, constituting a region crossing event; the data interaction between the IoT terminal device and the control interface of the key industrial facility through a preset application layer protocol, constituting a device interaction event; and the preset state transition occurring when the internal operating status field is parsed from the telemetry data reported by the IoT terminal device, constituting a state change event. When a business event is triggered, the state machine terminates the trajectory operation segment before the trigger and generates a new segment; the trajectory operation segment contains a spatiotemporal coordinate sequence, a business event type code used to identify the reason for termination, and associated metadata.

3. The adaptive security protection method for wireless communication based on edge computing according to claim 1, characterized in that: The identification process of the privacy risk assessment model includes static risk assessment, dynamic risk assessment, and risk fusion decision-making. The static risk assessment receives the current location coordinates of the trajectory operation segment and queries the functional area confidentiality map stored in the edge computing node. The functional area confidentiality map is a key-value pair data structure, where the key is the polygon coordinates of the spatial area, the value is the pre-calculated static risk score, and the output is the matched static risk score. The dynamic risk assessment extracts the trajectory uniqueness, spatial information sensitivity, and behavioral correlation analysis features of the trajectory operation segment to form a multi-dimensional behavioral feature vector; then the multi-dimensional behavioral feature vector is input into the pre-trained dynamic risk identification module to output a normalized dynamic risk score. The risk fusion decision receives static risk scores and dynamic risk scores, and weights them to obtain a privacy risk assessment value.

4. The adaptive security protection method for wireless communication based on edge computing according to claim 3, characterized in that: The process of determining the static risk score in the functional area confidentiality map includes: The IoT environment is divided into spatial regions, and a functional assessment is conducted based on the physical purpose and business functions of each spatial region. The functional assessment includes: the confidentiality level of the data generated and processed by the equipment within the space area in the data classification standard; the critical path weight of the process links carried by the space area in the overall production flow chart; and the stringency level of the physical security and network access control strategies deployed in the space area. Quantitative weighting coefficients are configured for the functional assessment, and a static risk score is calculated for each of the spatial regions.

5. The adaptive security protection method for wireless communication based on edge computing according to claim 3, characterized in that: The trajectory uniqueness is determined by calculating the continuous spatiotemporal coordinates of the trajectory operation segment and mapping them to a discretized spatial region grid to obtain a discrete state sequence; based on historical anonymous trajectory data, a group behavior baseline model including a state transition probability matrix and a trajectory sequence module is constructed; the discrete state sequence is identified through the group behavior baseline model to obtain a trajectory uniqueness score, thereby quantifying the trajectory uniqueness; The spatial information sensitivity is calculated by comparing the dwell time of the trajectory operation segment with the duration probability density function defined by the standard operating procedure of the functional area, and calculating the Z-score of the anomaly. The behavioral association analysis calculates the cumulative association weight of the node sequence path traversed by the trajectory operation segment in a process flow knowledge graph that uses equipment, regions, and materials as nodes.

6. The adaptive security protection method for wireless communication based on edge computing according to claim 1, characterized in that: For all constructed trajectory operation segments, the basic processing module applies privacy protection processing, including: for position coordinates, performing coordinate normalization operation to align with the center point of the preset grid; for telemetry values, calculating the Laplace distribution parameters according to the preset privacy budget and adding corresponding noise. When the privacy risk assessment value of the trajectory operation segment exceeds the preset risk threshold, the enhanced processing module initiates the trajectory disconnection mechanism on top of the basic processing; the trajectory disconnection mechanism generates a session update credential for the IoT terminal device and performs subsequent trajectory recognition.

7. The adaptive security protection method for wireless communication based on edge computing according to claim 1, characterized in that: The steps for training the virtual data generation model include: The privacy-enhancing trajectory data is read into the volatile memory of the edge computing node in the form of a data stream; after each backpropagation gradient calculation is completed, the data in the volatile memory is marked as erasable and overwritten by the memory management unit; and in the step of gradient aggregation to update model weights, Gaussian noise calculated based on the global privacy budget and gradient sensitivity is added to the aggregated gradient vector.

8. A wireless communication adaptive security protection system based on edge computing, characterized in that, include: The session credential generation module captures the physical layer channel state information, the functional area semantic identifier of the network gateway, and the current timestamp when the IoT terminal device first accesses the network gateway, and generates anonymous session credentials based on hash operations. The privacy risk assessment module tracks the operation trajectory data of the anonymous session credentials in real time and constructs trajectory operation segments; The privacy risk assessment model is invoked to identify the trajectory operation segments. Based on the trajectory uniqueness, spatial information sensitivity, and behavioral correlation analysis of the trajectory operation segments, a privacy risk assessment value is obtained. The privacy data processing module performs privacy protection processing on trajectory operation segments and further performs threshold judgment on privacy risk assessment values. When the preset risk threshold is exceeded, the trajectory chain disconnection mechanism is triggered. Based on the privacy protection processing and the trajectory chain disconnection mechanism, privacy-enhanced trajectory data is obtained. The virtual data generation module trains a preset virtual data generation model using privacy-enhanced trajectory data. After the virtual data generation model converges, it drives the virtual data generator within it to produce a virtual operation and maintenance data stream equivalent to the real production and operation and maintenance mode.

Images