Electromagnetic signal anti-sample generation method and system based on signal-flow mapping and dynamic disturbance

Abstract

The invention belongs to the field of wireless communication security, and discloses an electromagnetic signal confrontation sample generation method based on signal-flow mapping and dynamic disturbance modeling. The method comprises the following steps: firstly, acquiring an original signal by using software-defined radio, dividing a time period by a sliding window, and training a classifier by using TimesNet; extracting time domain and frequency domain characteristics of each fragment, mapping the time domain and frequency domain characteristics into a network flow structure, and selecting core characteristics to set a disturbance target; generating a smooth disturbance curve by adopting a damping second-order kinetic model, adjusting disturbance intensity and speed through parameters, and mapping back to a signal domain to obtain an adversarial sample conforming to physical constraints; and if the classifier fails to be attacked, iteratively optimizing the parameters until success. Experiments show that in wireless identification systems for electric vehicle keys, automobile unlocking and the like, the accuracy rate can be reduced from 95% to 15%, meanwhile, disturbance energy is low, signal forms are kept consistent, and the system has double values of actual combat and testing.

Description

Technical Field

[0001] The present invention relates to the field of adversarial machine learning and wireless signal processing technology, and specifically to a method and system for generating electromagnetic signal adversarial samples that combines cross-domain mapping with physical dynamics. The method is suitable for scenarios such as wireless communication security, spectrum resource protection, and intelligent signal interference analysis. Background Art

[0002] In intelligent communication systems, electromagnetic adversarial samples can subtly perturb signals while preserving the original signal's functionality or semantics, causing them to misclassify in deep learning models, such as spectrum identifiers and spectrum sensing systems. These samples serve not only as attack tools but also as a means of security testing and defense verification.

[0003] Traditional adversarial sample generation methods are mainly targeted at static scenes such as images and texts, and will face several key problems if directly used in the field of electromagnetic signals. The first is poor physical feasibility. Most algorithms generate adversarial samples from signals in simulation scenarios, lack consideration of realistic factors, and ignore the frequently changing physical conditions such as channel fading, noise, and multipath effects in real signal propagation scenarios, resulting in the generated samples being unable to be successfully transmitted or recognized in real environments; secondly, the perturbation concealment is insufficient. Current adversarial sample generation methods mostly rely on fixed-pattern perturbations or global perturbations at the signal level, achieving a high attack success rate through large-scale perturbations, but significantly changing the signal's spectrum or timing structure, making it easy to be captured by spectrum monitoring or intrusion detection systems. At the same time, existing methods generally ignore the sequence characteristics of electromagnetic signals and their coupling relationship with the physical laws of propagation.

[0004] In addition, unlike image or speech adversarial samples, electromagnetic signals must not only be misclassified, but also be able to reach the receiver intact through the wireless propagation link, while maintaining recognizable modulation modes and acceptable spectral characteristics. They have extremely strong physical constraints and engineering implementation requirements.

[0005] Traditional electromagnetic signal adversarial examples, such as those based on the Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Carlini & Wagner (CW), primarily perturb the overall waveform or spectrum of the electromagnetic signal, often assuming an ideal channel and a static environment. While these methods achieve high attack success rates on simulated datasets, they ignore the temporal correlation of signals and practical factors such as channel fading, frequency offset, and multipath interference in wireless propagation. Generated samples are difficult to propagate in real environments or be captured by spectrum systems, resulting in insufficient stealth and robustness. On the other hand, while methods based on Generative Adversarial Networks (GANs) improve the adaptability of perturbations, their perturbation structures lack physical constraints, are unexplainable, and are uncontrollable, making them difficult to meet the signal modulation integrity and protocol compliance requirements of real-world deployments. Consequently, existing methods still face significant bottlenecks in transferability and practicality in real-world scenarios.

[0006] The above analysis reveals the following problems and shortcomings of existing technologies: Most existing research focuses on perturbation design for the overall waveform or spectrum of the signal, ignoring the dynamic channel influence on the signal during propagation, resulting in the difficulty of the generated adversarial samples to function stably in real wireless environments. Most methods only achieve successful attacks under ideal simulation conditions, and fail to consider factors such as signal modulation structure, physical power constraints, and multipath fading. Although such methods can achieve strong attacks on specific models in a static environment, their generalization capabilities and physical feasibility are seriously insufficient in complex electromagnetic environments with multiple devices, multiple frequency bands, and multiple protocols coexisting, making them difficult to apply to signal security testing and attack deployment in multiple scenarios and dynamically evolving conditions. Summary of the Invention

[0007] In response to the problems existing in the prior art, the present invention provides a method and system for generating electromagnetic signal adversarial samples based on signal-traffic mapping and dynamic perturbations.

[0008] The present invention is implemented as follows: a method for generating electromagnetic signal adversarial samples based on signal-flow mapping and dynamic perturbations, comprising:

[0009] S1, uses wireless receiving equipment to collect the original electromagnetic signal and divides the signal into multiple time periods using a sliding window;

[0010] S2, input the segmented original electromagnetic signal samples to train a signal classifier;

[0011] S3, extracting time domain and frequency domain features for each time segment to form a structured feature sequence;

[0012] S4, flow mapping of the structured feature sequence, selecting the core features to set the perturbation target;

[0013] S5, generates a smooth perturbation curve through a damped second-order dynamic model, and uses different parameters to control the perturbation intensity and change speed;

[0014] S6, maps the perturbation curve back to the signal domain to generate adversarial samples that comply with physical constraints;

[0015] S7, input the generated adversarial sample into the classifier for attack testing. If it fails, it iteratively optimize the perturbation parameters. If it succeeds, it outputs the adversarial sample.

[0016] Furthermore, S1 specifically includes: using the software-defined radio receiving device HackRF One to collect electromagnetic signals. During the collection process, the center receiving frequency is set according to the transmission frequency band of the device under test, for example, 315.17MHz for car key signals and 2.4GHz for drone remote control signals. The front-end low-noise amplifier (LNA) gain is configured to 16 and the power amplifier (VGA) gain is configured to 10 to improve signal reception sensitivity. The sampling rate is set to 10MS / s (million samples per second) and the sampling time is set to 3 seconds to ensure that the timing characteristics and multipath changes of the signal are fully captured.

[0017] The collected signal is encoded in a complex IQ format (i.e., in-phase component I and quadrature component Q), which is expressed as:

[0018] x(t)=I(t)+jQ(t), t∈[0, T]

[0019] The data is saved as a binary file (.bin) in np.complex64 format. This data is continuously stored in a buffered manner and is suitable for subsequent batch processing.

[0020] After signal acquisition is completed, the continuous time domain signal is divided into "signal flow units" with boundary and structural properties, and its physical flow form is defined to achieve consistent expression of signal and flow space. The entire signal data set is regarded as a session in the flow. After the original continuous signal is segmented, a complete signal segment is regarded as a flow, a signal flow is divided into multiple time windows, and the signal in each window is regarded as a data packet. First, the original continuous signal is segmented into multiple time segments of equal length. Set the segment start and end time to [T i , T i+h ], is the sampling time interval of the i-th time segment, and the collected original continuous signal is represented as a series of signal segments x i , as follows:

[0021] x i =x(t),T∈[t,t i+1 ]

[0022] Where x(T) is the complex sample value at time t, including in-phase (I) and quadrature (Q) components. i Represents a signal stream of finite duration, providing a boundary structure for subsequent feature extraction and mapping operations. i Divide, let the total signal duration be t total , divided into N windows of equal length, each window is defined as follows:

[0023]

[0024] Each Window k As a basic signal unit, the divided signal sequence is:

[0025]

[0026] It represents the set of all time windows in a signal. The signal S i A set of features proposed for each time window.

[0027] Furthermore, S2 specifically includes: using the TimesNet time series network model to perform classification training on the segmented original electromagnetic signal samples:

[0028] (1) Input the pre-processed segmented raw electromagnetic signal samples;

[0029] (2) The input data is processed by TimesNet in multiple time blocks to extract temporal features. Each Times Block contains a time-domain convolutional encoder and a frequency-domain attention mechanism.

[0030] (3) After flattening and fully connected network, using ReLU activation function, the result of step (2) is gradually reduced in feature dimension through the fully connected layer;

[0031] (4) Normalize the result of step (3) and use the Softmax function to finally map the probability distribution of the five categories of features.

[0032] Furthermore, S3 specifically includes: for each divided time window Extract time domain and frequency domain features and express them as:

[0033]

[0034] The specific features extracted from the signal in each window are shown in Table 1:

[0035] Table 1 Signal feature extraction

[0036]

[0037]

[0038] Furthermore, S4 specifically includes: mapping the feature vector sequence extracted by S2 into the structural form of network traffic, and converting each signal segment x i Abstracted as a logical flow F i , the corresponding feature set is the "packet" set in the flow. The network traffic flow and data packets are defined as follows:

[0039]

[0040] in, Finger Flow F i The feature set of each data packet in Represented as flow F i The number of bytes transferred in Indicates the packet size, Indicates the time interval, Represents duration, with a total of k features. The feature set of each time window of the signal A feature set for each packet mapped to network traffic The specific feature mapping is shown in Table 2:

[0041] Table 2 Signal-traffic feature mapping

[0042]

[0043]

[0044] Among them, time domain energy and frequency domain power are specifically defined as:

[0045]

[0046] Where I(t) represents the real part, Q(t) represents the imaginary part, and t is time. FFT Indicates the number of FFT points. The denominator is used for normalization to ensure the physical meaning of the power spectrum density. i The time domain energy characteristics of each time window in the analogy flow F i The number of bytes transmitted for each data packet in the frequency domain power characteristic is analogous to the data packet size:

[0047]

[0048] Further, S5 specifically includes: generating a smooth perturbation curve through a damped second-order dynamics model, and controlling the intensity and change speed of the perturbation with different parameters. Introducing a dynamics system, mapping the discretized feature sequence into a continuous perturbation trajectory that conforms to physical laws, and accurately characterizing the dynamic attenuation, oscillation, and convergence behaviors of signal features through a damped second-order differential equation, so that the generated adversarial samples can not only inherit the physical constraints of the original signal but also accurately mislead the classifier through time-frequency and dual-domain collaborative perturbations. Extracting the features within the i-th time window from the above feature equation, we define its evolution to follow the following dynamics equation:

[0049]

[0050] This dynamics system is a typical damped second-order system, representing the change trajectory of features over time t, where W target is the increment of the feature, used to control the tendency value of the perturbation of this feature. ω n is the natural frequency of the system, determining the response speed of the change of the feature value. The larger the value, the faster the feature converges to the target value. ζ is the damping ratio, controlling the degree of attenuation or oscillation of the feature. When 0 < ζ < h, the system exhibits underdamped oscillation. When ζ = h, it is critically damped. When ζ > h, it is overdamped and there is no oscillation. in the solution of the dynamic equation indicates that the energy of the system gradually dissipates over time and finally tends to be stable. describes the oscillation behavior of the feature during the convergence process, and the frequency is determined by ω d

[0051] Adopting a randomized parameter strategy to generate a differentiated dynamics perturbation curve. In terms of parameter setting, the damping ratio ζ is sampled from a uniform distribution (ζ ∼ Uniform(0.3, h.2)), effectively avoiding the response anomalies caused by undamped oscillation ζ = 0 and severe overdamping ζ = 1.5. And the natural frequency (ω n ∼ LogUniform(0.1, h0)) allows the perturbation to span the time scale from milliseconds to seconds ω n → h0 and the slowly varying trend perturbation ω n → 0.1. Parameter randomization ensures that the perturbation patterns of each window are independent and unpredictable - from underdamped oscillation (ζ = 0.3, ω n = h0) to overdamped slow change (ζ = h.2, ω n = 0.h), covering the dynamic characteristics of multiple time scales.

[0052] ​Furthermore, S6 specifically includes: using an implicit reverse mapping mechanism to guide the perturbation back to the signal domain, gradually injecting modulation information from the perturbation trajectory in units of local time windows of the original signal, and achieving fine-grained control of the signal amplitude and structure. Within each local window, the modulation function guides the signal amplitude to make subtle adjustments, thereby embedding the characteristic perturbation into the signal itself in a physically reasonable way. According to the perturbation characteristic trajectory, the corresponding modulation function is constructed, and the local window of the original signal is fine-grainedly adjusted to generate an adversarial signal that meets the target characteristics. The modulation function is defined as follows:

[0053]

[0054] Among them, W(t k ) represents the characteristic disturbance value at the kth time point, W0 is the reference energy value, is the rate of change of the perturbation trajectory at that point. The two parts in the modulation function play different roles: the first It is used to adjust the amplitude response of the signal and simulate the natural fluctuation of the signal energy. The second term is the correction factor of the rate term, which is used to guide the disturbance to be more sensitive to transient changes, thereby enhancing its pertinence and the ability to express microscopic disturbances.

[0055] Furthermore, S7 specifically includes inputting the generated adversarial sample into the classification model to test the attack effect. If the attack fails, regenerate the perturbation and iterate the optimization. The adversarial signal sample generated after mapping the perturbation trajectory back to the signal domain in step S6 is Input into the trained signal classification model f(.) and observe the classification result y of the model adv If the model fails to produce misclassification, It is considered that the current perturbation has not achieved the attack target, and it is necessary to regenerate the perturbation trajectory and iteratively optimize the perturbation parameters. The perturbation parameter sampling mechanism is as follows:

[0056] Randomly sample the damping ratio ζ and natural frequency ω from the set perturbation parameter space n

[0057]

[0058] The sampled parameters are brought into the damped second-order system disturbance model to generate the disturbance trajectory and mapped into the modulation function to adjust the signal. Put the classification model again and judge whether the output category is consistent with the true label y true If the attack is not consistent, stop the iteration; otherwise, resample the perturbation parameters and repeat the above process until the attack is successful or the maximum number of iterations K is reached max .

[0059] Another object of the present invention is to provide a system for generating electromagnetic signal adversarial samples based on signal-flow mapping and dynamic perturbations, comprising:

[0060] The signal acquisition and preprocessing module is used to obtain the original electromagnetic signal through the wireless receiving device and divide the continuous signal into multiple time periods using the sliding window method;

[0061] A neural network training module is used to train a signal classifier based on the divided original electromagnetic signal samples, which is used to verify the attack effect of subsequent adversarial samples;

[0062] Feature extraction module, used to extract time domain and frequency domain features from each signal time segment and construct a structured feature sequence;

[0063] Traffic mapping module, used to map structured feature sequences into traffic forms and select core disturbance features as attack targets;

[0064] The disturbance generation module is used to generate continuous disturbance curves based on the damped second-order dynamic model, and control the disturbance intensity and change speed by setting different damping ratios and natural frequency parameters;

[0065] The signal reconstruction module is used to map the perturbation curve back to the signal domain and adjust the signal amplitude or spectral characteristics to generate adversarial samples that meet physical constraints;

[0066] The attack detection module is used to input the generated adversarial samples into the trained signal classifier model for attack verification and determine whether the attack is successful based on the model output results;

[0067] The optimization module iteratively optimizes the perturbation parameters if the attack fails, and outputs the final usable adversarial sample if it succeeds;

[0068] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps in the method for generating electromagnetic signal adversarial samples based on signal-flow mapping and dynamic perturbations are performed.

[0069] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps in the method for generating electromagnetic signal adversarial samples based on signal-flow mapping and dynamic perturbations.

[0070] Another object of the present invention is to provide an information data processing terminal, which is used to implement the electromagnetic signal adversarial sample generation system based on signal-flow mapping and dynamic perturbations to support the automatic generation, verification and output process of adversarial samples.

[0071] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0072] First, this invention introduces a flow-quantized mapping method for electromagnetic signals. Based on actual wireless signal samples, this method combines a sliding window method for preprocessing. It then extracts time-domain and frequency-domain features to construct structured time series data for subsequent adversarial sample generation. This design targets key perturbations of the original electromagnetic signal, such as energy and spectral power within a time segment. A perturbation trajectory generation method based on a damped second-order dynamic system is proposed, enabling continuous control and parameter adjustment of the signal perturbation process. Compared to traditional gradient perturbation methods such as FGSM and PGD, this method exhibits stronger perturbation controllability and temporal consistency, and the perturbation trajectory is physically interpretable and feasible for channel propagation. Experiments conducted on a typical electromagnetic modulation recognition dataset demonstrate that the adversarial samples generated by this method significantly reduce the model's recognition accuracy for the original signal, with the maximum accuracy reduction reaching 72.85%, demonstrating superior attack capability and robustness evaluation value.

[0073] The technical solutions of this invention offer the following technical effects and advantages: First, in experimental comparisons using the same dataset, our adversarial example generation method achieves certain advantages. Second, for the broad area of ​​electromagnetic signal adversarial examples, multi-domain models offer better detection results than single-domain models and simple hybrid-domain models, improving overall detection efficiency.

[0074] Second, the expected benefits lie in providing electromagnetic signal detection systems with an adversarial sample generation solution with practical deployable capabilities, significantly reducing the overall accuracy and robustness of existing wireless signal recognition models, and providing controllable and customizable adversarial attack modules for commercial systems such as communication security detection, drone countermeasures, and electromagnetic deception, which has the potential for industrial transformation.

[0075] The technical solution of the present invention fills the technical gap in the industry at home and abroad: based on the traditional electromagnetic countermeasure method that only focuses on signal strength or spectrum disturbance, it introduces structured traffic mapping and physical dynamic disturbance trajectory generation mechanism for the first time, and establishes a full-link system from signal structure to disturbance generation, solving the problem that existing methods are difficult to achieve continuous disturbance and universal disturbance of multiple types of real-scene signals, filling the technical gap in the physical transmissibility of signals in this field.

[0076] This invention addresses common issues with conventional adversarial examples in the electromagnetic signal domain, such as uncontrollable perturbations, unclear attack directions, and non-transmittable perturbation signals. By introducing a damped second-order dynamics system, this invention achieves structural control and target guidance during the perturbation process, significantly increasing the success rate of adversarial examples against target models.

[0077] This invention breaks through the inertial thinking that adversarial samples can only be generated in images, text or fixed-format traffic data, and extends adversarial disturbances to radio signals, a complex data type with physical continuity and propagation constraints. It combines the structured traffic mapping method to generate universal adversarial samples with temporal consistency, spectral rationality, and attack stability, effectively reducing the detection capability of mainstream classification models.

[0078] Third, by introducing a method for generating electromagnetic signal adversarial samples based on signal-flow mapping and dynamic perturbations, this invention solves a series of technical problems in the existing technology and achieves significant technological progress. These aspects are detailed below:

[0079] Technical issues solved:

[0080] 1. Limitations of Traditional Methods: Most existing methods for generating adversarial examples for electromagnetic signals directly perturb IQ signals or spectrograms. This perturbation process lacks structural control, making it difficult to maintain stability or concealment in real channels, and is generally ineffective against specific classification models. This paper proposes a signal-traffic mapping mechanism that structures continuous electromagnetic signals into a sequence of traffic features and perturbs key indicators in the feature space, making the perturbation process clearly targeted and interpretable.

[0081] 2. Poor Domain Adaptability: Existing methods for generating electromagnetic signal adversarial examples focus on specific task scenarios, such as fixed modulation type recognition or known model attacks. They lack adaptability to multi-task, multi-scenario, and multi-model environments, resulting in significantly reduced attack effectiveness against other signal types and difficulty in generalizing to more complex communication systems. This method, which employs a universal flow-based mapping approach and feature perturbation mechanism, demonstrates strong adversarial generalization capabilities across multiple domain scenarios.

[0082] 3. Poor physical feasibility and non-deployability: Existing signal perturbation methods are mostly limited to theoretical simulations. The perturbed signal spectrum may exhibit anomalies, excessive energy, or malfunction, limiting their application in real wireless communication systems. This invention generates continuous perturbation trajectories through a dynamic system and combines a signal reconstruction mechanism to map the perturbations back to the signal domain. This ensures that the final adversarial sample meets physical propagation requirements and possesses excellent spectrum constraints and energy control capabilities, thus resolving the key bottleneck that makes adversarial samples difficult to deploy.

[0083] Significant technological advancements achieved:

[0084] 1. Establish a structured perturbation mechanism: Through sliding window segmentation, feature extraction and flow mapping, the present invention represents the original electromagnetic signal as a structured flow sequence, providing a clear target structure for subsequent perturbations.

[0085] 2. Multiple types of real-world electromagnetic signals: The disturbance generation process designed in this invention is based on general feature abstraction and target parameterization. It does not rely on a specific model structure or a single signal type. This method is applicable to a variety of actual collected electromagnetic signal types.

[0086] 3. Achieving physical deployability: The present invention introduces a damped second-order dynamics model to generate perturbation trajectories and maps them back to the original signal domain, so that the generated adversarial sample signals meet the physical constraints of actual wireless communication systems in terms of amplitude, spectral range, power density, etc.

[0087] The method provided by this invention represents a significant technological advancement, effectively addressing several shortcomings of existing technologies and providing an innovative solution for generating adversarial examples for electromagnetic signals. These advances not only improve the accuracy and applicability of adversarial example attacks, but also advance research in electromagnetic information security and wireless communication security, providing practical technical support for future signal deception defense assessments and intelligent communication countermeasures.

[0088] Fourth, the technical design adopted by the electromagnetic signal adversarial sample generation method based on signal-flow mapping and dynamic perturbation in the present invention addresses several key technical problems in the existing technology and achieves significant technological progress.

[0089] 1. Signal-Flow Mapping Mechanism: By dividing the original electromagnetic signal into multiple time segments and extracting the time and frequency domain features of each segment, a structured feature sequence is constructed. This sequence is further mapped into a "flow-packet" structure similar to network traffic. Each time segment corresponds to a logical flow, and each feature dimension is analogous to a field or data packet in the flow, providing a unified structural foundation for subsequent perturbation generation.

[0090] 2. Dynamically generated disturbance trajectory: This paper introduces a damped second-order dynamic model as the core mechanism for disturbance generation. By adjusting the damping ratio and natural frequency, a smooth and adjustable disturbance trajectory is generated to guide the key features of the signal to evolve towards the target value, ensuring that the disturbance process is continuous, controllable and physically meaningful.

[0091] 3. Feature Dimension Perturbation Mapping and Signal Reconstruction: The generated perturbation trajectory is mapped back to the original signal domain, and a modulation function is used to adjust the signal amplitude or spectral components within the time segment. This process generates adversarial examples that maintain the propagation characteristics and modulation format of the original signal while maintaining controlled perturbation strength, making them suitable for transmission and reception in real wireless communication scenarios.

[0092] 4. Attack feedback and perturbation optimization mechanism: This paper designs a perturbation effect evaluation mechanism based on classifier feedback, tests the results of each perturbation, dynamically adjusts the perturbation parameters according to whether the attack is successful, and realizes iterative optimization of the perturbation. This mechanism enhances the model versatility and deployment flexibility of the method.

[0093] Technical issues solved:

[0094] 1. Lack of structural control over the original signal perturbation: Traditional adversarial example generation methods often rely on direct perturbations of the overall signal waveform or spectrum. This perturbation process lacks structural constraints, making it difficult to achieve precise attacks. This invention utilizes a signal-traffic mapping mechanism to convert electromagnetic signals into structured feature sequences, clearly defining the perturbation target and enabling more controlled and targeted attacks.

[0095] 2. Lack of cross-scenario adaptability: Existing methods often rely on specific signal types or model architectures, lacking the ability to adapt to diverse scenarios, different classifiers, or multiple types of electromagnetic signals. This paper proposes a universal perturbation generation mechanism that is applicable to a variety of electromagnetic signal types collected in real-world scenarios, improving the method's generalization and operational adaptability.

[0096] Technological advancements achieved:

[0097] 1. Improved attack effectiveness and disturbance accuracy: By introducing a damped second-order dynamics model to control the evolution trajectory of the disturbance in time and spectrum dimensions, the controllability of the counter-disturbance and the attack success rate are significantly improved, and the classification model's recognition accuracy of electromagnetic signals is reduced.

[0098] The present invention not only solves existing problems technically, but also demonstrates significant technical advantages in practical applications, providing an efficient, accurate and flexible new tool for electromagnetic signal adversarial samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0100] Figure 1 Schematic diagram of a method for generating electromagnetic signal adversarial samples based on signal-flow mapping and dynamic perturbations provided by an embodiment of the present invention;

[0101] Figure 2 This is a flow chart of a method for generating electromagnetic signal adversarial samples based on signal-flow mapping and dynamic perturbations provided by an embodiment of the present invention;

[0102] Figure 3 This is a structural diagram of the electromagnetic signal adversarial sample generation method based on signal-flow mapping and dynamic perturbation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0103] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0104] In response to the problems existing in the prior art, the present invention provides a method and system for generating electromagnetic signal anti-samples based on signal-flow mapping and dynamic perturbations. The present invention is described in detail below with reference to the accompanying drawings.

[0105] This paper presents a method for generating electromagnetic signal adversarial samples based on signal-flow mapping and dynamic perturbations. By combining feature-structured mapping with physically controllable perturbation technology, it provides an efficient and deployable solution for electromagnetic signal security testing and countermeasures. The following two specific examples demonstrate the application of this method in different scenarios.

[0106] Example 1: Generation of adversarial samples of electromagnetic signals from vehicle control systems

[0107] 1. Signal acquisition and preprocessing: The electromagnetic signals of the keyless start systems of Toyota Camry and Audi A3 are collected. After segmentation using a sliding window and normalization preprocessing, a standardized set of signal segments is obtained.

[0108] 2. Neural network training: Use the collected signal segments to train the signal classifier model. This is used to evaluate the effectiveness of subsequent adversarial attack samples.

[0109] 3. Feature extraction: Extract time domain features and frequency domain features for each time window segment to construct a structured feature vector sequence

[0110] 4. Organize the extracted multi-dimensional feature sequence into a structured "stream" structure, that is, regard each time segment as a logical stream, and each feature dimension corresponds to a "field" or "packet" in the stream, providing a unified representation format for generating perturbations.

[0111] 5. Perturbation trajectory generation: Introduce perturbations to the time-domain energy and frequency-domain power characteristics, generate perturbation trajectories through the dynamics model, and map them to the signal domain;

[0112] 6. Attack Iteration Optimization: The adversarial signal is input into the classifier for attack testing. If the attack fails, the perturbation trajectory parameters are adjusted for the next round of iteration until the attack is successful or the iteration limit is reached.

[0113] Example 2: Generation of adversarial samples of electromagnetic signals from smart remote control keys

[0114] 1. Signal acquisition and preprocessing: The electromagnetic signal of the Tesla Model S smart remote key is collected and segmented using a sliding window and normalized preprocessed to obtain a standardized set of signal segments.

[0115] 2. Neural network training: Use the collected signal segments to train the signal classifier model. This is used to evaluate the effectiveness of subsequent adversarial attack samples.

[0116] 3. Feature extraction: Extract time domain features and frequency domain features for each time window segment to construct a structured feature vector sequence

[0117] 4. Organize the extracted multi-dimensional feature sequence into a structured "stream" structure, that is, regard each time segment as a logical stream, and each feature dimension corresponds to a "field" or "packet" in the stream, providing a unified representation format for generating perturbations.

[0118] 5. Perturbation trajectory generation: Introduce perturbations to the time-domain energy and frequency-domain power characteristics, generate perturbation trajectories through the dynamics model, and map them to the signal domain;

[0119] 6. Attack Iteration Optimization: The adversarial signal is input into the classifier for attack testing. If the attack fails, the perturbation trajectory parameters are adjusted for the next round of iteration until the attack is successful or the iteration limit is reached.

[0120] These two embodiments demonstrate that the adversarial sample generation method of the present invention can be widely applied to a variety of electromagnetic signal scenarios, and can perform perturbation optimization on the characteristic structures of different signal types, thereby effectively improving the attack efficiency and cross-scenario adaptability of the classification model.

[0121] like Figure 1 、 Figure 2 As shown, the electromagnetic signal adversarial sample generation method based on signal-flow mapping and dynamic perturbation provided by the embodiment of the present invention includes:

[0122] S1, uses wireless receiving equipment to collect the original electromagnetic signal and divides the signal into multiple time periods using a sliding window;

[0123] S2, input the segmented original electromagnetic signal samples to train a signal classifier;

[0124] S3, extracting time domain and frequency domain features for each time segment to form a structured feature sequence;

[0125] S4, flow mapping of the structured feature sequence, selecting the core features to set the perturbation target;

[0126] S5, generates a smooth perturbation curve through a damped second-order dynamic model, and uses different parameters to control the perturbation intensity and change speed;

[0127] S6, maps the perturbation curve back to the signal domain to generate adversarial samples that comply with physical constraints;

[0128] S7, input the generated adversarial sample into the classifier for attack testing. If it fails, it iteratively optimize the perturbation parameters. If it succeeds, it outputs the adversarial sample.

[0129] In real-world scenarios where wireless communication security and intelligent perception intersect, electromagnetic signal classifiers play a fundamental role in numerous critical systems, such as drone identification, industrial control, and smart key authentication. However, these systems are highly vulnerable to adversarial attack. Traditional adversarial sample generation methods rely on perturbations of the pixel space of image data, making them difficult to transfer to the physically constrained electromagnetic signal domain. This is especially true when waveform continuity, time-frequency feature consistency, and modulation recognizability must be guaranteed. Existing technologies struggle to meet the dual requirements of both adversarial security and system stability.

[0130] To address this practical problem, this method proposes treating electromagnetic signals as complex entities with temporal boundaries and flow structures, thereby constructing a unified mapping between cyberspace and physical space. Specifically, after acquiring the signal, a sliding window is used to divide it into discrete time segments. Time and frequency domain features are extracted and defined as structured feature sequences, which are then further abstracted into logical packets equivalent to network traffic. This transformation ensures that each signal segment is bounded by both physical continuity and semantic boundaries, providing a context-controlled expression foundation for subsequent perturbation construction.

[0131] The design of the perturbation strategy breaks through the limitations of previous static vector noise addition methods and introduces a damped second-order system model to simulate the dynamic evolution of characteristic perturbations. In this model, the perturbation is viewed as a state response in the physical system, and its trajectory is driven by the natural frequency, damping ratio, and target eigenvalue. By analyzing the system response, perturbation trajectories with different attenuation or oscillation trends can be accurately generated, effectively perturbing the classifier decision boundary without compromising the usability of the original signal. This process significantly enhances the controllability and robustness of adversarial examples, making them highly valuable for practical deployment.

[0132] To reconstruct and inject the perturbation signal, this method constructs a modulation function to map the perturbation trajectory to the signal amplitude or spectral domain, achieving physically feasible and semantically discernible fine-grained adjustments. Within each signal window, a local gain function is constructed based on the amplitude and derivative of the perturbation and applied to the original signal. This approach preserves the global continuity of the signal while avoiding damage to the communication protocol or physical layer modulation structure. It is one of the few available mechanisms for countering attacks in practical wireless communication systems.

[0133] This method also incorporates a classifier feedback mechanism, dynamically sampling perturbation parameters to iteratively optimize unsuccessful attack samples. In each iteration, the classifier output provides feedback on the effectiveness of the perturbation, automatically adjusting the range of parameters such as the damping ratio and natural frequency, converging toward the model's vulnerable region, significantly improving attack efficiency. This closed-loop optimization approach not only improves the target matching of perturbation generation but also demonstrates strong model-agnostic (black-box) adaptability, making it particularly suitable for embedded communication devices deployed in complex scenarios.

[0134] To sum up, this method realizes the generation and dynamic adaptation of signal-level adversarial samples on the basis of maintaining the integrity of signal physical properties and protocol structure, solving the industry pain point that existing methods are difficult to simultaneously take into account physical feasibility and adversarial effectiveness. It has obvious system portability and scenario adaptability, and provides theoretical and practical support for building deep security protection in electromagnetic space.

[0135] The electromagnetic signal adversarial sample generation method based on signal-flow mapping and dynamic perturbation provided by the embodiments of the present invention involves the entire process from original electromagnetic signal acquisition to adversarial sample generation. The signal and data processing process in each step is explained in detail below:

[0136] S1: Use wireless receiving equipment to collect the original electromagnetic signal and divide the signal into multiple time periods using a sliding window

[0137] *Signal Collection: Using the HackRF One software-defined radio device, we capture electromagnetic signals in real time, including those from drones and remote control keys, and generate complex IQ signal streams.

[0138] *Data preprocessing: Bandpass filtering and power normalization are performed on the original signal to eliminate hardware nonlinear distortion and environmental noise.

[0139] *Window segmentation: Use a sliding window to divide the continuously collected signal data into multiple time segments to provide input samples for subsequent feature extraction.

[0140] S2: Training a signal classifier based on the segmented original electromagnetic signal samples

[0141] *Data partitioning: The segmented signal samples are divided into training set, validation set, and test set with a partition ratio of 7:2:1 for model training and evaluation.

[0142] *Model training: The time series neural network model is trained to classify electromagnetic signal samples, the test set is tested through classification, and the model parameters are continuously optimized to improve the model's detection performance for different electromagnetic signal types.

[0143] S3: Extract time domain and frequency domain features for each time segment to form a structured feature sequence

[0144] *Time domain feature extraction: The time domain includes calculating the signal amplitude statistics within the window, such as mean, variance, zero-crossing rate, etc.

[0145] *Frequency domain feature extraction: Use short-time Fourier transform (STFT) to calculate frequency domain features such as power and bandwidth to capture the signal spectrum structure.

[0146] S4: Structured feature sequence for flow mapping, select core features to set disturbance targets

[0147] * Cross-domain mapping mechanism: establish semantic alignment between signal features and network traffic statistics (see Table II)

[0148] *Core feature selection: Select the perturbation time domain energy and frequency domain power to avoid destroying local statistical features and causing functional failure.

[0149] S5: Generating smooth perturbation curves via a damped second-order dynamics model

[0150] *Disturbance trajectory generation: Use the damped second-order system equation to generate a smooth disturbance trajectory for the disturbance target, and adjust the damping ratio and natural frequency to control the disturbance speed and trend.

[0151] *Parameter combination strategy: A randomization strategy is used to explore the combination of disturbance parameters, and disturbance control curves are established in the time domain and frequency domain respectively to ensure that the disturbance process is continuously adjustable.

[0152] S6: Map the perturbation curve back to the signal domain to generate adversarial examples that conform to physical constraints

[0153] *Signal reconstruction: The perturbation features are mapped back to the original signal fragments through the amplitude gain function to reconstruct the perturbation signal.

[0154] S7: Input the generated adversarial sample into the classifier for attack testing. If it fails, iteratively optimize the perturbation parameters. If it succeeds, output the adversarial sample.

[0155] *Attack test: Input the reconstructed adversarial sample into the trained classifier to detect whether it can induce the classifier to output an incorrect result, thus achieving an adversarial attack.

[0156] * Perturbation optimization: If the attack fails, the perturbation parameters are adjusted based on the classification result feedback, the perturbation curve is regenerated and iteratively optimized until the attack is successful or the iteration limit is reached

[0157] S1 specifically involves using the HackRF One software-defined radio receiver to collect electromagnetic signals. During the acquisition process, the center receiving frequency is set according to the transmission frequency band of the device under test, such as 315.17 MHz for car key signals and 2.4 GHz for drone remote control signals. The front-end low-noise amplifier (LNA) gain is configured to 16, and the power amplifier (VGA) gain is configured to 10 to improve signal reception sensitivity. The sampling rate is set to 10 MS / s (million samples per second) and the sampling duration is set to 3 seconds to ensure that the timing characteristics and multipath variations of the signal are fully captured.

[0158] The collected signal is encoded in a complex IQ format (i.e., in-phase component I and quadrature component Q), which is expressed as:

[0159] x(t)=I(t)+jQ(t),t∈[0,T]

[0160] The data is saved as a binary file (.bin) in np.complex64 format. This data is continuously stored in a buffered manner and is suitable for subsequent batch processing.

[0161] After signal acquisition is completed, the continuous time domain signal is divided into "signal flow units" with boundary and structural properties, and its physical flow form is defined to achieve consistent expression of signal and flow space. The entire signal data set is regarded as a session in the flow. After the original continuous signal is segmented, a complete signal segment is regarded as a flow, a signal flow is divided into multiple time windows, and the signal in each window is regarded as a data packet. First, the original continuous signal is segmented into multiple time segments of equal length. Set the segment start and end time to [T i ,T i+h ], is the sampling time interval of the i-th time segment, and the collected original continuous signal is represented as a series of signal segments x i , as follows:

[0162] x i =x(t), T∈[t,t i+1 ]

[0163] Where x(T) is the complex sample value at time t, including in-phase (I) and quadrature (Q) components. i Represents a signal stream of finite duration, providing a boundary structure for subsequent feature extraction and mapping operations. iDivide, let the total signal duration be t total , divided into N windows of equal length, each window is defined as follows:

[0164]

[0165] Each Window k As a basic signal unit, the divided signal sequence is:

[0166]

[0167] It represents the set of all time windows in a signal. The signal S i A set of features proposed for each time window.

[0168] S2 specifically includes: using the TimesNet time series network model to classify and train the segmented original electromagnetic signal samples:

[0169] (1) Input the pre-processed segmented raw electromagnetic signal samples;

[0170] (2) The input data is processed by TimesNet in multiple time blocks to extract temporal features. Each Times Block contains a time-domain convolutional encoder and a frequency-domain attention mechanism.

[0171] (3) After flattening and fully connected network, using ReLU activation function, the result of step (2) is gradually reduced in feature dimension through the fully connected layer;

[0172] (4) Normalize the result of step (3) and use the Softmax function to finally map the probability distribution of the five categories of features.

[0173] S3 specifically includes: for each time window after division Extract time domain and frequency domain features and express them as:

[0174]

[0175] The specific features extracted from the signal in each window are shown in Table 1:

[0176] Table 1 Signal feature extraction

[0177]

[0178] S4 specifically includes: mapping the feature vector sequence extracted by S2 into the structural form of network traffic, and converting each signal segment x i Abstracted as a logical flow F i, the corresponding feature set is the "packet" set in the flow. The network traffic flow and data packets are defined as follows:

[0179]

[0180] in, Finger Flow F i The feature set of each data packet in Represented as flow F i The number of bytes transferred in Indicates the packet size, Indicates the time interval, Represents duration, with a total of k features. The feature set of each time window of the signal A feature set for each packet mapped to network traffic The specific feature mapping is shown in Table 2:

[0181] Table 2 Signal-traffic feature mapping

[0182]

[0183] Among them, time domain energy and frequency domain power are specifically defined as:

[0184]

[0185]

[0186] Where I(t) represents the real part, Q(t) represents the imaginary part, and t is time. FFT Indicates the number of FFT points. The denominator is used for normalization to ensure the physical meaning of the power spectrum density. i The time domain energy characteristics of each time window in the analogy flow F i The number of bytes transmitted for each data packet in the frequency domain power characteristic is analogous to the data packet size:

[0187]

[0188] S5 specifically includes: generating a smooth perturbation curve through a damped second-order dynamic model, and controlling the intensity and speed of the perturbation with different parameters. By introducing a dynamic system, the discretized feature sequence is mapped into a continuous perturbation trajectory that conforms to physical laws, and the dynamic attenuation, oscillation and convergence behavior of the signal characteristics are accurately characterized by a damped second-order differential equation, so that the generated adversarial sample can inherit the physical constraints of the original signal and accurately mislead the classifier through the coordinated perturbation of the time-frequency dual domain. The features within the i-th time window are extracted from the above characteristic equation, and we define its evolution to follow the following dynamic equation:

[0189]

[0190] This kinetic system is a typical damped second-order system, representing the change trajectory of the feature over time t, where W target is the increment of the feature, used to control the trend value of the feature perturbation. ω n is the natural frequency of the system, which determines the response speed of the eigenvalue change. The larger the value, the faster the feature converges to the target value. ζ is the damping ratio, which controls the degree of attenuation or oscillation of the feature. When 0 < ζ < h, the system exhibits underdamped oscillation. When ζ = h, it is critically damped. When ζ > h, it is overdamped and there is no oscillation. In the solution of the dynamic equation, represents that the system energy gradually dissipates over time and finally tends to be stable. describes the oscillation behavior of the feature during the convergence process, and the frequency is determined by ω d decides.

[0191] A randomized parameter strategy is adopted to generate a differentiated kinetic perturbation curve. In terms of parameter setting, the damping ratio ζ is sampled from a uniform distribution (ζ ∼ Uniform(0.3, h.2)), effectively avoiding the response anomalies caused by undamped oscillation ζ = 0 and severe overdamping ζ = h.5. And the natural frequency (ω n ~LogUniform(0.h, h0)) allows the perturbation to span the time scale from milliseconds to seconds ω n →h0 and the slowly varying trend perturbation ω n →0.1. Parameter randomization ensures that the perturbation patterns in each window are independent and unpredictable - from underdamped oscillation (ζ = 0.3, ω n = h0) to overdamped slow change (ζ = h.2, ω n = 0.h), covering the dynamic characteristics of multiple time scales.

[0192] S6 specifically includes: using an implicit inverse mapping mechanism to guide the perturbation back to the signal domain, taking the local time window of the original signal as a unit, and gradually injecting the modulation information from the perturbation trajectory to achieve fine-grained control of the signal amplitude and structure. In each local window, the modulation function guides the signal amplitude to make subtle adjustments, thereby embedding the feature perturbation into the signal itself in a physically reasonable way. According to the perturbation feature trajectory, a corresponding modulation function is constructed to make fine-grained adjustments to the local window of the original signal, thereby generating an adversarial signal that meets the target features. The modulation function is defined as follows:

[0193]

[0194] where, W(t k ) represents the feature perturbation value at the k-th time point, and W0 is the reference value of the benchmark energy, is the rate of change of the perturbation trajectory at that point. The two parts in the modulation function play different roles: the first It is used to adjust the amplitude response of the signal and simulate the natural fluctuation of the signal energy. The second term is the correction factor of the rate term, which is used to guide the disturbance to be more sensitive to transient changes, thereby enhancing its pertinence and the ability to express microscopic disturbances.

[0195] S7 specifically includes inputting the generated adversarial sample into the classification model to test the attack effect. If the attack fails, regenerate the perturbation and iterate the optimization. The adversarial signal sample generated by mapping the perturbation trajectory back to the signal domain in step S6 is Input into the trained signal classification model f(.) and observe the classification result y of the model adv If the model fails to produce misclassification, It is considered that the current perturbation has not achieved the attack target, and it is necessary to regenerate the perturbation trajectory and iteratively optimize the perturbation parameters. The perturbation parameter sampling mechanism is as follows:

[0196] Randomly sample the damping ratio ζ and natural frequency ω from the set perturbation parameter space n

[0197]

[0198] Substitute the sampled parameters into the damped second-order system disturbance model:

[0199] Generate a perturbation trajectory and map it to a modulation function to adjust the signal. Put the classification model again and judge whether the output category is consistent with the true label y true If the attack is not consistent, stop the iteration; otherwise, resample the perturbation parameters and repeat the above process until the attack is successful or the maximum number of iterations K is reached max .

[0200] like Figure 3 As shown, the electromagnetic signal adversarial sample generation system based on signal-flow mapping and dynamic perturbation provided by the embodiment of the present invention includes:

[0201] The signal acquisition and preprocessing module is used to obtain the original electromagnetic signal through the wireless receiving device and divide the continuous signal into multiple time periods using the sliding window method;

[0202] A neural network training module is used to train a signal classifier based on the divided original electromagnetic signal samples, which is used to verify the attack effect of subsequent adversarial samples;

[0203] Feature extraction module, used to extract time domain and frequency domain features from each signal time segment and construct a structured feature sequence;

[0204] Traffic mapping module, used to map structured feature sequences into traffic forms and select core disturbance features as attack targets;

[0205] The disturbance generation module is used to generate continuous disturbance curves based on the damped second-order dynamic model, and control the disturbance intensity and change speed by setting different damping ratios and natural frequency parameters;

[0206] The signal reconstruction module is used to map the perturbation curve back to the signal domain and adjust the signal amplitude or spectral characteristics to generate adversarial samples that meet physical constraints;

[0207] The attack detection module is used to input the generated adversarial samples into the trained signal classifier model for attack verification and determine whether the attack is successful based on the model output results;

[0208] The optimization module iteratively optimizes the perturbation parameters if the attack fails, and outputs the final usable adversarial sample if it succeeds;

[0209] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps in the method for generating electromagnetic signal adversarial samples based on signal-flow mapping and dynamic perturbations are performed.

[0210] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps in the method for generating electromagnetic signal adversarial samples based on signal-flow mapping and dynamic perturbations.

[0211] Another object of the present invention is to provide an information data processing terminal, which is used to implement the electromagnetic signal adversarial sample generation system based on signal-flow mapping and dynamic perturbations to support the automatic generation, verification and output process of adversarial samples.

[0212] Example:

[0213] A method for generating electromagnetic signal adversarial samples based on signal-flow mapping and dynamic perturbation is characterized by feature mapping and dynamic system perturbation. The method includes the following steps:

[0214] Step 1: Use the HackRF wireless receiving device to collect the raw electromagnetic signal data in the actual scene;

[0215] Step 2: Divide the original IQ data collected in step 1 into a sliding window with a fixed length of 128 sampling points to form multiple continuous window time segments, each segment contains 128 complex sample values;

[0216] Step 3: Train the signal classifier model using all time segments divided in Step 2 as input samples. Use the Adam optimizer with a learning rate of 0.0001 and a cross-entropy loss function for multi-classification training. Each batch processes 64 samples, for a total of 200 training rounds.

[0217] Step 4: For each window time segment in step 2, extract the time domain features and frequency domain features respectively.

[0218] Step 5: Construct the multidimensional features extracted in step 4 into a structured feature sequence, perform signal-traffic mapping, and form a logical flow structure;

[0219] Step 6: Select the time domain energy and spectral power of the key perturbation target dimension in the feature sequence of step 5 as the perturbation variable;

[0220] Step 7: Based on the time domain characteristics of each time window segment, generate the time domain disturbance trajectory based on the damped second-order dynamic model.

[0221] Step 8: Map the time domain disturbance trajectory back to the signal domain to generate the corresponding disturbance signal

[0222] Step 9: Right Perform fast Fourier transform to obtain frequency domain representation, select the frequency domain features of each time window segment, and generate the disturbance trajectory based on the damped second-order dynamic model;

[0223] Step 10: Map the frequency domain disturbance back to the signal domain to form the preliminary disturbance signal sample x after the complete disturbance adv ;

[0224] Step 11: Concatenate the adversarial signals of all window segments to reconstruct the entire adversarial sample x adv , and input it into the signal classifier model trained in step 3 for recognition to obtain the classification result

[0225] Step 12: If That is, if the classifier is not successfully fooled, the perturbation parameters are updated;

[0226] Step 13: Repeat steps 7 to 12 to regenerate the dual-stage perturbations in the time domain and frequency domain until the attack is successful or the maximum number of iterations K is reached. max ;

[0227] Step 14: The final generated successful adversarial sample x adv Save it locally and record its perturbation parameters and attack success model;

[0228] Step 15: The adversarial sample x constructed in step 14 advInput into the classifier trained in step 3 for final detection;

[0229] Step 15: The highest accuracy of the model detection is 91.20%; after using adversarial samples, the overall accuracy of the model is 28.10%.

[0230] The embodiments of the present invention have achieved some positive effects during the development or use process and indeed have great advantages over the existing technology. The following content describes them in conjunction with data, charts, etc. of the experimental process.

[0231] Table 3 Attack results

[0232]

[0233] As shown in the table above, the accuracy and SSIM metrics for classifying and detecting multiple device types in a real electromagnetic signal dataset using different attack methods are presented. The last column shows data from our experiments, where the classifier uses a trained model to identify various signal samples. "No Attack" represents baseline performance without an adversarial attack, with an overall accuracy of 91.20%. The attack method proposed in this paper reduces this accuracy to 28.10%, achieving significantly better results than similar attack methods.

[0234] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0235] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for generating electromagnetic signal adversarial samples based on signal-flow mapping and perturbation, characterized in that: The following steps are involved: S1, segmentation processing of the original electromagnetic signal; S2, extract the time domain and frequency domain features of each signal segment to form a structured feature sequence; S3, mapping the characteristic sequence into a logical traffic structure and constructing a disturbance target set consistent with the data packet; S4, generates a perturbation curve based on the selected features and maps it back to the signal domain to form an adversarial sample.

2. A system for generating electromagnetic signal adversarial samples based on dynamic perturbations, characterized in that: include: Acquisition and preprocessing module, used to convert raw signals into time segments; Feature mapping module, used to convert time-frequency features into traffic logical structure; Perturbation building module, used to generate perturbation curves; The signal reconstruction module is used to inject disturbances into the signal to form adversarial samples.

3. A method for generating an electromagnetic signal disturbance curve based on a damped second-order system, characterized in that: include: Construct the dynamical differential equations that describe the perturbation trajectory; Set the disturbance target value, damping ratio and natural frequency; Solve the system response function to obtain the disturbance curve; The output continuous perturbation features are used to physically reconstruct the signal.

4. The system according to claim 2, wherein: The disturbance construction module generates a controllable amplitude change curve based on the dynamic equation to control the convergence speed and change trend of the disturbance.

5. A method for generating adversarial sample perturbations based on classifier feedback optimization, characterized in that: include: Input the adversarial sample into the classification model to verify whether it is misclassified; If the classification result does not change, update the perturbation parameters; Generate a new perturbation curve based on the new parameters and map it back to the signal; The above process is repeated until a valid adversarial sample is generated or the set number of iterations is reached.

6. The system according to claim 2, wherein: It also includes a feedback optimization module for adjusting the perturbation parameters based on the classifier output to improve the attack effect.

7. The method according to claim 1, characterized in that Feature extraction includes extracting a feature set consisting of the maximum interval, average duration, energy, power, bandwidth mean, frequency domain amplitude, number of zero crossings and mean in each time segment.

8. The method according to claim 1, characterized in that Map structured signal characteristics into logical packet attributes, including packet size, interval, transmission rate and number of bytes, and abstract each signal segment into a logical flow.

9. The method according to claim 3, characterized in that The perturbation trajectory is constructed as a modulation function to adjust the signal amplitude in each time window to achieve fine-grained perturbation control.

10. The system according to claim 2, wherein: The signal reconstruction module works together with the disturbance construction module to apply the disturbance function to each local time period through the window synchronization mechanism to achieve distributed disturbance injection.

Images