Internet-of-things digital twinborn antagonism defense method and system facing electromagnetic spectrum

Abstract

The invention relates to the technical field of Internet of Things physical layer security, in particular to an Internet of Things digital twinborn antagonism defense method and system oriented to an electromagnetic spectrum. Comprising the following steps: constructing an electromagnetic spectrum digital twin of a target Internet of Things system to simulate a physical signal environment in a high-fidelity manner; in the twinborn body, an intelligent optimization algorithm is utilized to generate an antagonistic signal sample aiming at the communication model; carrying out adversarial training on a communication model by using the samples to enhance the robustness of the communication model, wherein a meta-learning framework can be adopted in the training process to improve the generalization adaptive capacity; establishing a dynamic synchronous closed loop between the digital twinborn body and a real physical environment, and realizing model calibration and real threat sample back feeding; when an attack is detected in a real environment, multiple defense strategies are simulated and evaluated in the twin, and an optimal strategy is selected and issued to physical equipment for execution. According to the invention, digital twinning, antagonistic machine learning and real-time closed-loop control are combined, and the anti-interference capability of the Internet of Things system is actively refined in advance at low cost.

Description

Technical Field

[0001] This invention relates to the fields of physics, chemistry, and biology, and specifically to an adversarial defense method and system for the Internet of Things (IoT) oriented towards the electromagnetic spectrum. Background Technology

[0002] With the rapid development and widespread application of IoT technology, massive numbers of smart devices are interconnected through wireless networks, extending the security boundary from the traditional network and data link layers to the physical layer. The electromagnetic spectrum, as the physical carrier of wireless communication, is highly vulnerable to physical layer attacks such as eavesdropping, interference, and deception. Traditional physical layer security technologies, such as spread spectrum, frequency hopping, and channel coding, rely heavily on static or pre-defined adversarial models, often proving inadequate against intelligent, adaptive, and ever-changing electromagnetic attacks. Furthermore, existing machine learning-based physical layer authentication and intrusion detection methods typically rely on training with attack samples collected in real-world environments. However, acquiring real attack samples is costly, time-consuming, and difficult to cover all unknown attack patterns, resulting in insufficient model generalization ability and a significant "zero-day attack" risk. Digital twin technology offers the possibility of high-fidelity reconstruction of physical systems in virtual space, but its current applications in IoT security are mostly focused on status monitoring, fault prediction, or network topology optimization. No mature solutions have yet been developed for its in-depth application in building dynamic, adversarial electromagnetic attack and defense training environments. Deeply integrating the high-fidelity simulation capabilities of digital twins, the intelligent countermeasure capabilities of machine learning, and the dynamic protection requirements of the IoT physical layer security to construct a proactive defense system capable of anticipating threats, continuously evolving defenses, and achieving real-time linkage between digital and physical spaces has become crucial for overcoming current IoT security bottlenecks and addressing future challenges of intelligent electromagnetic warfare. This invention aims to systematically solve these problems.

[0003] Therefore, existing technologies still need further development. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide an adversarial defense method and system for IoT digital twins oriented towards the electromagnetic spectrum, so as to solve the problems existing in the prior art.

[0005] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides an adversarial defense method for Internet of Things digital twins oriented towards the electromagnetic spectrum, comprising:

[0006] S1. Construct an electromagnetic spectrum digital twin of the target Internet of Things system, wherein the digital twin is used to simulate the wireless communication channel, device radio frequency characteristics and background electromagnetic noise in the real physical environment in a high-fidelity manner in the digital space.

[0007] S2. In the electromagnetic spectrum digital twin, generate adversarial signal samples against the communication model of the Internet of Things system, wherein the adversarial signal samples simulate intelligent interference or signal deception attacks.

[0008] S3. Using the generated adversarial signal samples, the communication model is adversarially trained in the electromagnetic spectrum digital twin to enhance the robustness and security of the communication model under adversarial conditions.

[0009] S4. Deploy the communication model parameters or defense strategies that have been enhanced through adversarial training to real IoT physical devices.

[0010] Specifically, in step S2, generating adversarial signal samples includes:

[0011] Based on preset interference attack types and parameter spaces, an intelligent optimization algorithm is used to search for performance degradation points in the communication model and generate adversarial signal samples that lead to communication performance degradation.

[0012] Specifically, the intelligent optimization algorithm is a reinforcement learning algorithm, the state space of the reinforcement learning algorithm includes the real-time performance index of the communication model, and the reward function is set to induce the communication model to make misjudgments or performance degradation.

[0013] Specifically, the reinforcement learning algorithm also accesses an electromagnetic spectrum threat feature database, which stores signal features of historical attacks; the reinforcement learning algorithm uses features in the threat feature database to initialize or guide the generation direction of adversarial signal samples.

[0014] Specifically, after step S1 and before step S4, step S0 is also included:

[0015] S0. Establish a dynamic synchronous closed loop between the electromagnetic spectrum digital twin and the real physical environment, collect electromagnetic spectrum data and communication equipment status of the real environment in real time, and use them to update the model parameters of the digital twin.

[0016] Specifically, in step S0, the dynamic synchronization closed loop further includes:

[0017] Abnormal interference signals captured in real physical environments are input into the electromagnetic spectrum digital twin as new adversarial signal samples to supplement or update the adversarial training sample set.

[0018] Specifically, in step S3, the adversarial training adopts a meta-learning framework, enabling the communication model to acquire the ability to quickly learn from a small number of adversarial samples and adapt to new and unknown attacks, thus achieving generalized defense capabilities.

[0019] Specifically, in step S4, the deployment of the application includes:

[0020] When a potential attack is detected in a real physical environment, multiple preset defense strategies are simulated and run in the electromagnetic spectrum digital twin, and the optimal defense strategy is selected based on the simulation results and sent to the corresponding real IoT physical device for execution.

[0021] Specifically, the defense strategy includes at least one of the following:

[0022] Communication carrier frequency hopping, adaptive adjustment of transmit power, switching of channel coding scheme, and weighted reconstruction of multi-antenna beamforming.

[0023] According to a second aspect of the present invention, an IoT digital twin adversarial defense system oriented towards the electromagnetic spectrum is provided, comprising:

[0024] A digital twin building block is used to build and maintain an electromagnetic spectrum digital twin of a target IoT system.

[0025] An adversarial sample generation module is used to generate adversarial signal samples targeting the communication model in the digital twin;

[0026] An adversarial training module is used to perform adversarial training on the communication model in the digital twin using the adversarial signal samples;

[0027] The synchronization update module is used to realize data synchronization and model update between the digital twin and the real physical environment;

[0028] The strategy deployment module is used to deploy the trained and enhanced communication model parameters or the optimal defense strategy evaluated based on digital twin simulation to real IoT physical devices.

[0029] Beneficial effects:

[0030] Compared with existing technologies, the IoT digital twin adversarial defense method and system oriented towards the electromagnetic spectrum provided by this invention has the following significant advantages:

[0031] First, this invention creates a safe, controllable, and extremely low-cost "virtual range" by constructing a high-fidelity electromagnetic spectrum digital twin. Within this range, various extreme and complex electromagnetic attack scenarios, including known and unknown intelligent interference modes, can be simulated risk-free. This overcomes the high-risk and high-cost limitations of conducting attack and defense exercises in real networks, providing a foundational platform for proactive defense.

[0032] Secondly, this invention creatively introduces adversarial machine learning concepts into the digital twin environment. By employing intelligent optimization algorithms such as reinforcement learning and evolutionary strategies, it proactively and purposefully searches for performance vulnerabilities in the communication model, intelligently generating efficient adversarial examples. This process transcends traditional training models that rely on limited historical attack data, systematically exposing and strengthening the model's deep-seated vulnerabilities, and significantly improving the inherent robustness of the communication model in the face of unknown and ever-changing attacks.

[0033] Third, this invention constructs a dynamic synchronous closed-loop and threat feedback mechanism between the digital twin and the real physical environment. By collecting channel states and abnormal signals from the real environment in real time, the twin model is dynamically calibrated to ensure its fidelity, and the captured real attack samples are fed back to the virtual training ground. This ensures that adversarial training always targets the most realistic and cutting-edge threats, achieving continuous online evolution of defense capabilities.

[0034] Fourth, by introducing a meta-learning framework, the communication model gains the ability to quickly learn from small samples and adapt to new and unknown attacks. This enables IoT devices to quickly adjust their parameters based on only a small number of samples when encountering new types of interference that have not appeared in the training data, achieving agile defense that can be "learned and used immediately," and effectively responding to zero-day attacks.

[0035] Finally, this invention achieves a collaborative proactive defense closed loop of "cloud-based simulation decision-making - rapid terminal execution." When a real attack is detected, multiple physical layer defense strategies are simulated and deduced in parallel and rapidly within a digital twin, and the optimal strategy is scientifically selected and executed in real time based on a comprehensive benefit function. This changes the traditional passive response mode, forming a complete OODA loop of "perception-simulation-decision-execution," greatly improving the intelligence, scientific nature, and timeliness of the defense response, and providing a complete technical path for building a next-generation IoT security system with immunity and self-healing capabilities. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the adversarial defense method for IoT digital twins oriented towards the electromagnetic spectrum provided in a specific embodiment of the present invention.

[0037] Figure 2 This is a schematic diagram of the system composition of an IoT digital twin adversarial defense system oriented towards the electromagnetic spectrum provided in a specific embodiment of the present invention. Detailed Implementation

[0038] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0039] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0040] Please see Figure 1 This invention provides an adversarial defense method for IoT digital twins oriented towards the electromagnetic spectrum, comprising:

[0041] S1. Construct an electromagnetic spectrum digital twin of the target Internet of Things system, wherein the digital twin is used to simulate the wireless communication channel, device radio frequency characteristics and background electromagnetic noise in the real physical environment in a high-fidelity manner in the digital space.

[0042] Further explanation is needed regarding the specific implementation of step S1, which is as follows: First, a physical model of the target IoT system (e.g., a LoRa WAN-based sensor network) is performed. This includes defining the network topology (location distribution of gateways and terminal nodes), the hardware RF model of each node (transmitter: power amplifier nonlinear characteristics, phase noise; receiver: noise figure, filter response, ADC quantization bits), and the antenna model (gain pattern). The wireless channel model adopts an improved 3G PPTR 38.901 Urban Micro (UMi) scenario path loss model, superimposed with a TDL-C channel model to simulate delay spread and multipath fading. The specific path loss PL(d) (unit: dB) calculation formula is as follows: in, This represents the path loss value at a distance of d. This represents the three-dimensional distance between the transmitter and receiver (unit: meters). The carrier frequency is represented in GHz. This model was chosen because it better reflects the propagation characteristics of densely deployed urban IoT devices. Background electromagnetic noise is generated by the superposition of Gaussian white noise and impulse noise following a generalized Pareto distribution. The probability of impulse noise occurrence is set to 0.01, and the shape parameter is set to 0.5 to simulate real-world sudden electromagnetic interference. The digital twin is implemented on a cloud server using MATLAB / Simulink or a customized C++ simulation platform, achieving a running speed more than 100 times faster than real-time.

[0043] S2. In the electromagnetic spectrum digital twin, generate adversarial signal samples against the communication model of the Internet of Things system, wherein the adversarial signal samples simulate intelligent interference or signal deception attacks.

[0044] It should be further explained that in step S2, the "generation of adversarial signal samples" is performed by a dedicated adversarial sample generation engine. The engine's input is the target communication model (e.g., a deep neural network for LoRa signal demodulation), and its output is a carefully constructed I / Q signal vector. First, the basic signal is generated. This can be a standard LoRa uplink signal. Then, the interference generator produces an interference signal. Adversarial examples It is the superposition of the two: ,in This is the interference strength coefficient, initially set to 0.1. Interference signal. The type is randomly selected from a preset type library, including: monotone interference ( Linear frequency sweep interference, partial frequency band noise interference with a bandwidth of 200kHz, and waveforms generated based on generative adversarial networks (GANs) are used to simulate unknown intelligent interference.

[0045] S3. Using the generated adversarial signal samples, the communication model is adversarially trained in the electromagnetic spectrum digital twin to enhance the robustness and security of the communication model under adversarial conditions.

[0046] It should be further noted that the adversarial training in step S3 is performed in the training environment of the digital twin. A batch of adversarial examples generated in step S2 (batch size set to 32) along with their corresponding clean labels (the original sent data) are input into the communication model. The Adam optimizer with gradient clipping is used, with an initial learning rate set to... To stabilize the training process, the loss function is a weighted sum of cross-entropy loss and adversarial loss: .in, It is the cross-entropy loss of the standard classification task. It encourages the model to output consistent loss values ​​for adversarial and clean samples, such as KL divergence and weighting coefficients. Set it to 0.5 to balance performance on the main task and robustness. Training should last at least 50 epochs, or until accuracy no longer improves on independent validation sets.

[0047] S4. Deploy the communication model parameters or defense strategies that have been enhanced through adversarial training to real IoT physical devices.

[0048] It should be further explained that step S4 deployment is divided into two modes: model deployment and policy deployment. Model deployment is suitable for devices with strong terminal computing capabilities. The trained communication model parameters (such as neural network weight files) are sent to the device via secure OTA updates to replace the original signal processing algorithm. Policy deployment is suitable for simple terminals. After the digital twin determines the optimal defense strategy (such as "immediately hop to channel CH47") in the cloud simulation, concise policy instructions are sent to the terminal for execution via the control channel.

[0049] Specifically, in step S2, generating adversarial signal samples includes: based on preset interference attack types and parameter spaces, searching for performance degradation points of the communication model through an intelligent optimization algorithm, and generating adversarial signal samples that lead to communication performance degradation.

[0050] It should be further explained that the preset interference attack type library T includes: Type 1: Single-tone interference, Type 2: Linear frequency sweep interference, Type 3: Partial frequency band noise interference, Type 4: Impulse noise interference, and Type 5: GAN generative interference. Each type corresponds to a parameter space P. For example, for single-tone interference, its parameter space... Including: interference amplitude (Range: [0.01, 1.0]), Interference frequency offset (Relative to the signal center frequency, range: [-125kHz, +125kHz]), initial phase (range: [0,2) For linear frequency sweeping interference, its parameter space In addition to amplitude and initial phase, it also includes the starting frequency. End frequency and sweep frequency cycle .

[0051] Furthermore, the intelligent optimization algorithm employs an adaptive evolutionary strategy based on the covariance matrix. The specific steps are as follows:

[0052] 1. Initialization: Randomly generate parameters within the parameter space for each type of interference. A number of candidate solutions (i.e., parameter combinations) constitute the initial population. Let... Calculate the fitness value for each candidate solution. The fitness function is defined as the increment of the bit error rate (BER) of the communication model after adding the interference corresponding to that parameter, relative to the BER of the clean signal. The goal is to maximize the fitness value.

[0053] 2. Selection and Recombination: Selecting from the current population (set) The individual with the highest fitness is selected as the parent. This is used to calculate... The weighted average of each individual value is used as the mean of the next-generation search distribution. .

[0054] 3. Adaptive Update: Update the search covariance matrix based on the distribution of parent individuals. and step length This allows the algorithm to adaptively adjust the direction and scope of the parameter search, performing a fine-grained search in regions with good performance and exploring the entire system.

[0055] 4. Sampling: Based on the updated mean Covariance matrix and step length Sample to generate new Each offspring individual forms a new generation of population.

[0056] 5. Iteration: Repeat steps 2-4 until the maximum number of iterations is reached (e.g., 100 generations) or the fitness value increases less than the threshold within 10 consecutive generations. .

[0057] 6. Output: Extract the interference parameters corresponding to the individual with the highest fitness in each generation of the population, and use them to generate the final adversarial signal sample.

[0058] Understandably, this method, through targeted optimization, can systematically discover "vulnerable" combinations of interference parameters that degrade the performance of a specific communication model, making it dozens of times more efficient than randomly generating samples.

[0059] Specifically, the intelligent optimization algorithm is a reinforcement learning algorithm, the state space of the reinforcement learning algorithm includes the real-time performance index of the communication model, and the reward function is set to induce the communication model to make misjudgments or performance degradation.

[0060] It should be further noted that this invention employs a deep deterministic policy gradient algorithm as the reinforcement learning algorithm. Its specific implementation is as follows:

[0061] 1. Environment: The environment refers to the communication simulation link in the digital twin. The actions of the intelligent agent (jammer) will change the parameters of the interference signal, thereby affecting the performance of the communication model at the receiving end.

[0062] 2. State Space At time step t, the state is a 12-dimensional vector, including: the current channel signal-to-noise ratio estimate (SNR, in dB), the output probability distribution of the communication model for the current received signal (taking the probability values ​​of the top-3 categories, a total of 3 dimensions), the bit error rate (BER, statistics of the past 10 symbols) of the current demodulation result, and the current parameters of the interference signal (normalized amplitude). Frequency offset Type encoding The system comprises three dimensions: the confidence level of the communication model (output maximum probability value), the peak-to-average power ratio (PAPR) of the received signal, and the average throughput over the past three time steps. This state information comprehensively reflects the current status of the communication link and interference.

[0063] 3. Motion space The action is a 4-dimensional continuous vector, with each dimension ranging from [-1, 1]. The adjustments to the interference parameters are as follows: (Amplitude adjustment, mapped to the actual change in [-0.1, +0.1]) (Frequency offset adjustment, mapped to [-10kHz, +10kHz]) (Phase adjustment amount, mapped to [- / 10,+ [ / 10]), and an action to trigger an interference type switch with a certain probability. When the type switch action value is greater than 0.5, the agent randomly selects another type from the interference type library.

[0064] 4. Reward Function The reward function is designed as follows: in: This represents the bit error rate of the communication model under adversarial sample conditions at time step t. This represents the bit error rate at the previous time step. This represents the cumulative energy consumed by the jammer up to time step t (related to jamming power and time integral). This represents the absolute value (in kHz) of the offset (of the center frequency of the interference signal relative to the carrier frequency of the communication signal) at time step t. (Constant term) These are small penalties at each step, encouraging the agent to complete tasks quickly. The rationale for the parameter selection is as follows: The first parameter, with a coefficient of 10 (maximum), emphasizes that improving the bit error rate is the core objective. The second parameter, with a coefficient of 0.5, encourages energy conservation and seeks efficient interference points. The third parameter, with a coefficient of -0.1, penalizes excessive frequency offset, aiming to guide interference energy to concentrate near the signal frequency band, improving stealth and efficiency. The constant term, -0.01, prevents the agent from taking meaningless and ineffective actions that delay the process.

[0065] 5. Network Structure and Training: Both the Actor and Critic networks are 3-layer fully connected neural networks with 256 neurons per layer, using the ReLU activation function. The experience replay buffer size is... Discount factor Soft update parameters Each training epoch contains 200 time steps, with a total of 5000 training epochs. The optimizer uses Adam, and the learning rate is set to... .

[0066] Understandably, through the above settings, the reinforcement learning agent learns to dynamically adjust the interference parameters to maximize the disruptive effect on the communication model with minimal energy cost.

[0067] Specifically, the reinforcement learning algorithm also accesses an electromagnetic spectrum threat feature database, which stores signal features of historical attacks; the reinforcement learning algorithm uses features in the threat feature database to initialize or guide the generation direction of adversarial signal samples.

[0068] It should be further explained that this invention introduces a prior knowledge base to accelerate and guide learning. The electromagnetic spectrum threat feature database is a relational database, and each record contains the following fields: signal ID, time-domain I / Q sampling data (1024 points in length), spectral feature vector (dimension reduced to 20 dimensions by PCA from 128-point FFT amplitude spectrum), interference type label, optimal countermeasure parameters, and attack effectiveness score.

[0069] Furthermore, during the initialization of the reinforcement learning agent, it no longer starts exploring from completely random parameters. The specific guidance method is as follows:

[0070] 1. Feature-based reward shaping: In the basic reward function In addition, add an auxiliary reward based on feature similarity. : in: Calculate cosine similarity. . It is the spectral feature vector of the interference signal generated at the current time step. It is the spectral feature vector of the k-th record in the threat feature database. It is an index set of all records in the feature library. This is the scaling factor, set to 0.2. When This reward is only calculated when the similarity is at a certain level. A threshold of 0.7 was chosen to require at least moderate similarity, providing guidance without unduly restricting the exploration of novel interference patterns.

[0071] 2. Course Learning Initialization: In the early stages of training, the agent's experience replay buffer is pre-filled with the interaction history (state, action, reward, new state) corresponding to the "best adversarial parameters" recorded in the threat feature database, with a filling ratio of 20%. This provides the agent with high-quality initial exploration experience.

[0072] 3. Guided movement exploration: During the training process, using... Instead of selecting the action output by the Actor network, the agent randomly chooses a record from the threat feature database and executes it using its "optimal adversarial parameters." This ensures that the agent always has a certain probability of utilizing known effective attack patterns throughout the training process.

[0073] Understandably, the threat signature database itself is updated through two avenues: first, by using efficient adversarial examples discovered through reinforcement learning from digital twins; and second, by capturing and validating anomalous signals from real-world physical environments. This mechanism allows the generation of adversarial examples to combine inheritance with innovation.

[0074] Specifically, after step S1 and before step S4, step S0 is also included: S0, establishing a dynamic synchronous closed loop between the electromagnetic spectrum digital twin and the real physical environment, collecting electromagnetic spectrum data and communication equipment status of the real environment in real time, and using them to update the model parameters of the digital twin.

[0075] It should be further explained that this invention establishes a synchronization link between the digital twin and the physical world. The specific data flow and synchronization algorithm for step S0 are as follows:

[0076] 1. Data Acquisition: Lightweight state-aware firmware is embedded in the baseband processing chip of each IoT terminal node and gateway node. This firmware is updated every [number] days. Data packets are collected every second (synchronization cycle), encapsulated, and then uploaded to the cloud-based digital twin platform via a dedicated low-frequency, low-power control channel (such as LoRa's Class B mode) or by utilizing gaps in the data channel. The collected data packets include:

[0077] ① Channel sounding information: By sending a known pilot sequence, the receiver estimates the channel frequency response (CFR) and takes the complex channel gain on 64 subcarriers.

[0078] ② Background noise floor: Statistical values ​​(mean and variance) of Received Signal Strength Indication (RSSI) are measured during time gaps when there is no communication activity.

[0079] ③ Device status: Transmit power level of this node, remaining battery power, GPS location (if available).

[0080] ④ Communication performance metrics: average packet error rate (PER) and throughput within the most recent time window (e.g., 100 data packets).

[0081] 2. Online Model Parameter Update: After receiving data, the digital twin platform initiates a synchronous update routine. Let the original channel model in the digital twin be... The measured channel frequency response reported from the i-th terminal node is The channel model parameters are updated using the recursive least squares method: the channel is modeled as a model containing L paths, with the frequency response being... ,in and These are the complex gain and time delay of the l-th path, respectively. The RLS algorithm minimizes the cost function. To iteratively update and The forgetting factor was set to 0.99 to balance the impact of remembering old data with the impact of new data.

[0082] 3. Synchronization Triggering and Verification: Define synchronization error metrics Where N is the number of active terminal nodes, It is the actual packet error rate reported by node i. This is the packet error rate calculated by simulating the corresponding node in a digital twin under the same communication load. A synchronization error threshold is set. When continuous Calculated in one synchronization cycle (i.e., 3 seconds) Both are greater than When the digital twin becomes out of sync with the physical world, the system triggers a "deep synchronization": the platform instructs all terminal nodes to perform a centralized channel probe, collecting denser channel information, and uses the collected batch data to perform a complete recalibration of the digital twin's channel and noise models. and This is to avoid false triggering caused by noise from a single measurement while being able to respond promptly to real changes in the channel environment.

[0083] Specifically, in step S0, the dynamic synchronization closed loop further includes: inputting the abnormal interference signals captured in the real physical environment as new adversarial signal samples into the electromagnetic spectrum digital twin to supplement or update the adversarial training sample set.

[0084] It should be further explained that this invention enables real-world threats to "feed back" into the digital space. The specific implementation process is as follows:

[0085] 1. Anomaly Detection and Capture: A lightweight anomaly detector is integrated into the terminal node's sensing firmware. This detector continuously monitors the following characteristics of the received signal: the duration for which instantaneous energy exceeds a preset threshold (e.g., 10 dB above the average noise floor), the symmetry of the signal spectrum, and cyclostationary characteristics. When multiple characteristics simultaneously exceed the threshold, an "anomaly capture mode" is triggered. In this mode, the device records I / Q data before and after the anomaly signal segment at the highest sampling rate (e.g., 2 Msps) for a recording duration of [duration missing]. The recording time is milliseconds to ensure complete capture of interference events. Recorded data includes timestamps, center frequency, and geographic location tags, and is marked as "suspected interference" before being uploaded to the cloud.

[0086] 2. Cloud-based Verification and Feature Extraction: After receiving a "suspicious interference" sample, the cloud platform initiates the verification process. First, it recreates the normal communication scenario of the terminal node under the same time, geographical location, and channel conditions in the digital twin. Then, the captured abnormal I / Q data is injected into the simulation as additional interference. If the packet error rate of the simulation after injection... Compared to the baseline packet error rate without interference Increased beyond the threshold If the abnormal signal is deemed a genuine and valid attack sample, it is accepted. Next, a 32-dimensional feature vector is extracted from the accepted sample. This includes: normalized power spectral density (16-dimensional, obtained by FFT amplitude spectrum downsampling), time-domain peak-to-average power ratio, signal bandwidth (bandwidth occupying 90% of the energy), and signal time-domain burst length. This feature vector... Its corresponding original I / Q data fragments and the verified interference effect ( They are stored together in the "Real Capture Adversarial Sample Library".

[0087] 3. Sample Library Management and Usage: This sample library employs a first-in, first-out (FIFO) strategy, with a maximum capacity of 10,000 records. During adversarial training (step S3), each training batch consists of 70% algorithm-generated samples and 30% randomly sampled samples from the "Real Capture Adversarial Sample Library." This ensures that the model can handle both extreme cases explored by the algorithm and learn potentially more complex interference patterns existing in the real world. Furthermore, every 24 hours, the system uses all data from the "Real Capture Adversarial Sample Library" to fine-tune the generator of the Generative Adversarial Network (GAN), enabling the GAN to generate new samples that more closely resemble the statistical characteristics of real-world interference, thereby continuously enriching the diversity of algorithm-generated samples.

[0088] Specifically, in step S3, the adversarial training adopts a meta-learning framework, enabling the communication model to acquire the ability to quickly learn from a small number of adversarial samples and adapt to new and unknown attacks, thus achieving generalized defense capabilities.

[0089] It should be further noted that this invention employs a model-independent meta-learning algorithm framework to train the communication model, enabling it to adapt rapidly. The specific training steps are as follows:

[0090] 1. Task Construction: Construct a large number of different meta-learning tasks from the overall adversarial example library (containing generated examples and real captured examples). Each task simulates a specific interference environment. For example, task Strong single-tone interference environment (SIR between -10dB and 0dB); mission : Fast time-varying multipath channel superimposed impulse noise environment; mission : An environment with specific GAN-generated waveform interference. Each task They all have their own data distribution .

[0091] 2. Inner Loop Adaptation: Let the parameters of the communication model (e.g., a CNN classifier) ​​be... For a sampled batch of tasks (one batch contains 4 tasks, i.e., batchsize=4), each task We can determine the distribution of these distributions. Sampling a support set (Include (sample-label pairs) and a query set (Contains 15 sample-label pairs). For each task, the model performs inner loop adaptation using its support set. Specifically, the model is computed on the support set... loss function on (Using cross-entropy loss). Then, the model parameters are updated once using gradient descent to obtain the task-specific adapted parameters. : in, This is the inner loop learning rate, set to... K=5 was chosen to simulate a small sample size scenario. This is to achieve rapid but not drastic parameter adjustments.

[0092] 3. Outer loop optimization: After the inner loop adapts, the model optimizes the query set for each task. Calculate loss The goal of meta-learning is to minimize the sum of the losses of all tasks on their respective query sets after inner loop adaptation. Therefore, the loss of the outer loop (meta-update) is: Then, calculate this loss with respect to the initial parameters. The gradient (which requires calculating the second derivative, but MAML does this via automatic differentiation) is calculated, and the initial parameters are updated. : in, It is the outer loop learning rate (meta-learning rate), set to .choose This is for the purpose of stable meta-optimization.

[0093] 4. Iterative Training: Repeat steps 2 and 3 until the meta-loss converges. After training, the model learns the initial parameters. It has excellent cross-task generalization ability.

[0094] 5. Rapid Adaptation to New Interference: When a novel type of interference never seen before is detected in a real-world environment, the device can quickly collect a small number of samples (e.g., 5-10 samples) under this interference to form a support set. Initialize parameters using pre-trained meta-learning. Perform 1-5 inner loop gradient descent updates on the device (using the inner loop learning rate). This allows for the rapid acquisition of model parameters optimized for this new disturbance. This immediately improves robustness to the interference and enables "rapid adaptation to small samples".

[0095] Specifically, in step S4, the deployment of the application includes: when a potential attack is detected in a real physical environment, simulating and running multiple preset defense strategies in the electromagnetic spectrum digital twin, and selecting the optimal defense strategy based on the simulation results and sending it to the corresponding real IoT physical device for execution.

[0096] It should be further explained that this invention achieves collaborative defense through cloud-based decision-making and edge / terminal execution. The specific decision-making and execution closed loop is as follows:

[0097] 1. Attack Detection and Triggering: In a real-world environment, the terminal or gateway continuously monitors the Bit Error Rate (BER). When the BER continuously exceeds a certain threshold... The threshold was exceeded in each data packet transmission cycle. If the Received Signal Strength Indication (RSSI) is normal, it is initially determined that the issue is a potential malicious interference attack, rather than a simple channel depth fading. The device then sends an "attack alert" to the cloud-based digital twin platform, along with a snapshot of the key status information at the current moment. This includes: the estimated current channel matrix Preliminary spectral characteristics of the interference signal (center frequency, bandwidth), and a list of available defensive actions for this device (such as a list of switchable channels, available transmission power levels, etc.).

[0098] 2. Cloud-based strategy simulation: Upon receiving an alarm, the cloud platform immediately loads the virtual mirror node corresponding to the alarm device into the digital twin and initializes its state. The platform then selects a set (e.g., 6) of feasible defense strategies from the "defense strategy library". Perform parallel simulation and deduction. Examples of policies in the policy library:

[0099] ① Switch to the backup channel CH_alt1 while maintaining the original modulation and coding scheme (MCS).

[0100] ② In the current channel, reduce the transmit power from Upgraded to .

[0101] ③ Switch to a more robust MCS (e.g., downgrade from 16QAM to QPSK, reducing the coding rate from 3 / 4 to 1 / 2).

[0102] ④ Combination strategy: Switch the channel to CH_alt2 and simultaneously adopt... MCS.

[0103] : Activate frequency hopping mode, the frequency hopping pattern is Pattern_A.

[0104] ⑥ Combination strategy: Activate frequency hopping mode Pattern_B and simultaneously increase power by 6dBm.

[0105] 3. Strategy Evaluation and Selection: For each strategy... Digital twins with Real-time speed simulation of the future The communication process takes seconds. In the simulation, the interference model is continuously generated based on preliminary characteristics. After the simulation, the overall benefit under this strategy is calculated. : in: It represents the average throughput (bps) achieved during the simulation. This is the additional energy consumed by the device during the simulation (relative to the baseline, in joules). This is the performance loss (unit: bps·s) resulting from the handover latency and signaling overhead incurred by implementing this strategy. These are weights, set to... .choose Ensuring communication performance is the top priority; Energy consumption should be considered while performance allows; Punish complex strategies that cause excessively long communication outages.

[0106] 4. Strategy Distribution and Execution: The cloud platform compares all... Choose the strategy with the highest return. This is encoded into a short command (e.g., "CMD:FH_Pattern_B;PWR+6dBm") and sent to the attacked real device via a secure control channel (e.g., using AES-128 encryption). The device atomically executes the strategy in the next predefined signaling slot after receiving the command. After execution, the device continues to monitor BER and feeds the results back to the cloud, forming a complete "perception-decision-execution-evaluation" closed loop.

[0107] Specifically, the defense strategy includes at least one of the following: communication carrier frequency hopping, adaptive adjustment of transmit power, channel coding scheme switching, and multi-antenna beamforming weight reconstruction.

[0108] It should be further explained that this invention clarifies the specific technical means and their parameterized implementation in the defense strategy library:

[0109] 1. Communication carrier frequency hopping: Applicable to protocols whose physical layer supports frequency hopping (such as Bluetooth, IEEE 802.15.4). The policy parameter is the frequency hopping sequence number. and frequency hopping rate The digital twin's policy library pre-stores multiple pseudo-random frequency hopping sequences, each with a period of 256 hops. (Frequency hopping rate...) The selectable values ​​are 50 hops, 100 hops, and 200 hops per second. During simulation, the twin is determined based on the spectral characteristics of the interference (such as narrowband interference bandwidth). Choose a strategy. If They tend to choose fast frequency hopping ( Frequency hopping is used to quickly escape interference frequencies (256 hops / second); if the interference is broadband, slower frequency hopping or a combination of other strategies is chosen. Choosing a period of 256 hops and a rate of 50-200 hops / second represents an engineering balance between anti-interference capability, synchronization overhead, and equipment complexity.

[0110] 2. Adaptive adjustment of transmit power: The strategy parameter is the power adjustment amount. The value ranges from +3, +6, +10, -3, to -6 dBm (relative to the current power). In the simulation, the twin not only evaluates the improvement in receiver signal-to-noise ratio (SNR) caused by increasing power. ,in This is an estimate of the path loss increment caused by possible small-scale location changes, and it also assesses the impact on device battery life. A penalty term is included. It will be added to the payoff function In the calculation, This is the power penalty coefficient, set to 0.05, to suppress unnecessary power increases. 3. Channel coding scheme switching: The strategy parameter is the Modulation and Coding Scheme Index (MCS Index). For example, switching from a high-order MCS (e.g., MCS-9, 64QAM, code rate 5 / 6) to a lower-order but more robust MCS (e.g., MCS-4, QPSK, code rate 1 / 2). In the simulation, the twin will look up the pre-stored SIR-MCS performance mapping table based on the currently estimated signal-to-interference ratio (SIR), selecting the scheme that guarantees a packet error rate lower than [previous value]. The highest rate MCS. This mapping table was obtained through prior offline simulation or field calibration.

[0111] 4. Multi-antenna beamforming weighted reconstruction: Suitable for devices equipped with at least two antennas (such as gateways). The policy parameter is the beamforming weight vector. When interference is detected from a direction At that time, the twin uses the current channel estimate H and interference to... Recalculate the weights in the simulation. To maximize the received signal strength in the direction of the target user, while... The direction forms a null trap. The optimization problem can be formalized as:

[0112]

[0113]

[0114] in, It is an array manifold vector. It is a very small value (e.g.) dB). This represents the conjugate transpose. The twin is obtained by solving this optimization problem to obtain a new... The effect is simulated. This strategy can suppress interference spatially, but its computational complexity is high, and it is usually only used on nodes with computing resources, such as gateways.

[0115] Understandably, digital twins evaluate these strategies or combinations thereof in parallel simulations, ultimately selecting the overall benefit. The strategy with the highest value.

[0116] Please see Figure 2 The present invention provides another embodiment, which provides an IoT digital twin adversarial defense system oriented towards the electromagnetic spectrum. The IoT digital twin adversarial defense system oriented towards the electromagnetic spectrum includes:

[0117] Digital twin building module 100 is used to build and maintain an electromagnetic spectrum digital twin of a target Internet of Things system;

[0118] Adversarial sample generation module 200 is used to generate adversarial signal samples targeting the communication model in the digital twin;

[0119] The adversarial training module 300 is used to perform adversarial training on the communication model in the digital twin using the adversarial signal samples.

[0120] The synchronization update module 400 is used to realize data synchronization and model update between the digital twin and the real physical environment;

[0121] The strategy deployment module 500 is used to deploy the trained and enhanced communication model parameters or the optimal defense strategy evaluated based on digital twin simulation to real IoT physical devices.

[0122] It should be further noted that the system is deployed as a microservice on a cloud or edge server cluster, and the modules communicate with each other through message queues and API interfaces, as detailed below:

[0123] 1. Digital Twin Building Module 100: Composed of a simulation engine and a model repository. The simulation engine is based on OMNeT++ or a custom discrete event simulation kernel, integrating an RF front-end model library (containing equivalent baseband models of power amplifiers, mixers, filters, and ADCs / DACs), a channel model library (containing mathematical models of path loss, shadowing fading, and multipath fading), and a communication protocol stack simulation model (accurate to the physical layer frame structure). The model repository uses a MongoDB database to store RF parameter configuration files, network topology configuration files, and historical channel snapshots for different IoT device types (such as LoRa terminals and NB-IoT modules). This module can quickly instantiate a virtual IoT network containing thousands of nodes based on the configuration files, achieving a simulation time speedup of up to 100:1.

[0124] 2. Adversarial Example Generation Module 200: This module includes a reinforcement learning agent service and an example synthesizer. The agent service runs the DDPG algorithm based on the TensorFlow or PyTorch framework, with both its Actor and Critic networks being 3-layer fully connected neural networks (256-128-64 neurons). This service obtains the simulation environment interface from the digital twin construction module and threat feature library data from the synchronization update module. The example synthesizer, based on the action parameters output by the agent or parameters randomly selected from the library, calls GNU Radio or a custom waveform generation function library to generate I / Q sampling data of adversarial signals in real time and injects it into the simulation link.

[0125] 3. Adversarial Training Module 300: This is a centralized model training platform based on Kubernetes for resource scheduling. It manages multiple training task queues. Each training task specifies the communication model architecture used (e.g., a one-dimensional convolutional neural network with the structure: Conv1D(64,kernel=3)-ReLU-MaxPool-Conv1D(128,kernel=3)-ReLU-GlobalAvgPool-Dense(128)-Dense(num_classes)), the training data source (from the adversarial example generation module and the real captured sample library), and the training hyperparameters (e.g., the inner / outer loop learning rate for meta-learning). Training tasks are distributed to compute nodes equipped with NVIDIA V100 GPUs for execution. The trained models are versioned and stored in the model registry.

[0126] 4. Synchronization Update Module 400: Composed of a data access gateway, a stream processing engine, and a model calibrator. The data access gateway receives uplink data streams from real devices via the MQTT protocol. The stream processing engine (such as an Apache Flink job) processes these data streams in real time, calculates indicators such as channel characteristics and packet error rate, and compares them with the predicted values ​​of the digital twin to determine whether synchronization is triggered. The model calibrator implements online learning algorithms such as recursive least squares. When synchronization is triggered, it consumes a batch of the latest data from the data stream, calculates the incremental update of the channel model parameters, and pushes the updated parameters to the simulation engine of the digital twin construction module via API calls.

[0127] 5. Policy Deployment Module 500: This module includes a policy simulation engine, a policy evaluator, and a command issuance service. The policy simulation engine is a lightweight, fast-starting copy of the digital twin building block, specifically designed to receive real-time state snapshots from the synchronization update module and load policies from the policy library for rapid parallel simulation. The policy evaluator is a microservice that reads the result metrics of each simulation from the Redis cache and, according to the definitions of this invention... The function calculates the benefit and selects the optimal strategy. The command delivery service then securely sends the optimal strategy instructions to the target device via an HTTPS RESTful API or CoAP protocol. The device has a corresponding lightweight strategy executor responsible for receiving, parsing, and executing the instructions.

[0128] It is understandable that all the above modules work together to form a complete autonomous system of perception, simulation, training, decision-making, and execution, providing the Internet of Things with continuously evolving and proactive physical layer security defense capabilities.

[0129] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising:

[0130] The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the described electromagnetic spectrum-oriented IoT digital twin adversarial defense method. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0131] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0132] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0133] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0134] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A digital twin adversarial defense method for the Internet of Things (IoT) oriented towards the electromagnetic spectrum, characterized in that, Includes the following steps: S1. Construct an electromagnetic spectrum digital twin of the target Internet of Things system, wherein the digital twin is used to simulate the wireless communication channel, device radio frequency characteristics and background electromagnetic noise in the real physical environment in a high-fidelity manner in the digital space. S2. In the electromagnetic spectrum digital twin, generate adversarial signal samples against the communication model of the Internet of Things system, wherein the adversarial signal samples simulate intelligent interference or signal deception attacks. S3. Using the generated adversarial signal samples, the communication model is adversarially trained in the electromagnetic spectrum digital twin to enhance the robustness and security of the communication model under adversarial conditions. S4. Deploy the communication model parameters or defense strategies that have been enhanced through adversarial training to real IoT physical devices.

2. The method according to claim 1, characterized in that, In step S2, generating adversarial signal samples specifically includes: Based on preset interference attack types and parameter spaces, an intelligent optimization algorithm is used to search for performance degradation points in the communication model and generate adversarial signal samples that lead to communication performance degradation.

3. The method according to claim 2, characterized in that, The intelligent optimization algorithm is a reinforcement learning algorithm. The state space of the reinforcement learning algorithm includes the real-time performance index of the communication model, and the reward function is set to induce the communication model to make misjudgments or performance degradation.

4. The method according to claim 3, characterized in that, The reinforcement learning algorithm also accesses an electromagnetic spectrum threat feature database, which stores signal features of historical attacks; the reinforcement learning algorithm uses features in the threat feature database to initialize or guide the generation direction of adversarial signal samples.

5. The method according to claim 1, characterized in that, After step S1 and before step S4, step S0 is also included: S0. Establish a dynamic synchronous closed loop between the electromagnetic spectrum digital twin and the real physical environment, collect electromagnetic spectrum data and communication equipment status of the real environment in real time, and use them to update the model parameters of the digital twin.

6. The method according to claim 5, characterized in that, In step S0, the dynamic synchronization closed loop further includes: Abnormal interference signals captured in real physical environments are input into the electromagnetic spectrum digital twin as new adversarial signal samples to supplement or update the adversarial training sample set.

7. The method according to claim 1, characterized in that, In step S3, the adversarial training adopts a meta-learning framework, enabling the communication model to acquire a generalized defense capability that can quickly learn from a small number of adversarial samples and adapt to new and unknown attacks.

8. The method according to claim 1 or 7, characterized in that, In step S4, the deployment of the application specifically includes: When a potential attack is detected in a real physical environment, multiple preset defense strategies are simulated and run in the electromagnetic spectrum digital twin, and the optimal defense strategy is selected based on the simulation results and sent to the corresponding real IoT physical device for execution.

9. The method according to claim 8, characterized in that, The defense strategy includes at least one of the following: Communication carrier frequency hopping, adaptive adjustment of transmit power, switching of channel coding scheme, and weighted reconstruction of multi-antenna beamforming.

10. An IoT digital twin adversarial defense system oriented towards the electromagnetic spectrum, for implementing the method as described in any one of claims 1-9, characterized in that, include: A digital twin building block is used to build and maintain an electromagnetic spectrum digital twin of a target IoT system. An adversarial sample generation module is used to generate adversarial signal samples targeting the communication model in the digital twin; An adversarial training module is used to perform adversarial training on the communication model in the digital twin using the adversarial signal samples; The synchronization update module is used to realize data synchronization and model update between the digital twin and the real physical environment; The strategy deployment module is used to deploy the trained and enhanced communication model parameters or the optimal defense strategy evaluated based on digital twin simulation to real IoT physical devices.

Images