Radio high-frequency band malicious signal identification and adaptive filtering method

Abstract

The invention discloses a radio high-frequency band malicious signal identification and self-adaptive filtering method, which comprises the steps of extracting multidimensional features, synchronously performing time-frequency domain, modulation domain and statistical domain analysis on baseband signals, extracting feature indexes including short-time Fourier transform energy concentration degree, cyclic spectrum symmetry and the like, performing feature alignment driven by environmental perception, and performing adaptive filtering. A prior interference environment model is constructed by using a frequency spectrum image, a historical interference label and geographic position information, and distributed alignment and dimension reduction are performed on feature vectors through a semantic embedding algorithm based on a graph structure. Therefore, digital signature verification, fingerprint comparison and path consistency analysis are finally completed on the FPGA platform, and safe traceability is carried out on the signal source. The problems that the false alarm rate of fixed threshold detection is high, a filtering mechanism is difficult to suppress nonlinear interference, a model lacks adaptability, a source verification means is weak and the like are effectively solved, and accurate recognition, deep suppression and security defense of malicious signals in a complex high-frequency electromagnetic environment are achieved.

Description

Technical Field

[0001] The present invention relates to the field of radio technology, and in particular to a method for identifying malicious signals and adaptive filtering in the radio high-frequency band. Background Art

[0002] In the field of malicious signal identification and filtering in the radio high-frequency band (3 - 30 MHz), the existing technologies mainly adopt energy detection based on fixed thresholds and conventional filtering methods. For example, the spectral characteristics of signals are analyzed through fast Fourier transform (FFT), and interference signals are suppressed by combining matched filters or adaptive linear filters. In addition, some solutions introduce machine learning algorithms (such as support vector machines) for signal classification, or use spatial domain beamforming technology to enhance target signals. However, the performance of these methods still has great room for improvement in high-dynamic and complex electromagnetic environments, especially in terms of nonlinear interference suppression, real-time processing capabilities, and anti-spoofing capabilities; During actual use, fixed-threshold detection is difficult to cope with complex electromagnetic environments, with high false alarm rates and missed detection rates. Conventional linear filters have poor suppression capabilities for nonlinear interference (such as swept-frequency signals), and the signal-to-interference-plus-noise ratio (SINR) improvement is limited. Traditional algorithms have high computational complexities and are difficult to meet the real-time processing requirements of high-frequency signals. There is a lack of in-depth exploration of physical layer fingerprint features, and they are vulnerable to forged signal attacks. Most existing methods have static parameter settings and cannot dynamically adapt to changes in the electromagnetic environment. Summary of the Invention

[0003] The present invention aims to solve at least one of the technical problems in the related technologies to some extent.

[0004] To this end, the object of the present invention is to propose a method for identifying malicious signals and adaptive filtering in the radio high-frequency band, which realizes high-precision malicious signal detection through an improved deep residual network, significantly improves the filtering effect by adopting a three-stage cascaded filtering architecture (spatial domain beamforming + nonlinear Volterra filtering + blind source separation), and introduces a dynamic reconstruction mechanism and an electromagnetic environment cognition engine to enhance environmental adaptability. In addition, through physical layer fingerprint feature verification and blockchain evidence storage mechanism, as well as FPGA hardware acceleration and dynamic partial reconfiguration technology, the problems of low detection accuracy, poor filtering effect, insufficient real-time performance, and weak anti-spoofing ability in the existing technologies are solved, and high-precision and high-reliability malicious signal identification and filtering are achieved.

[0005] To achieve the above object, the present invention proposes a method for identifying malicious signals and adaptive filtering in the radio high-frequency band, including the following steps: S1. Signal perception and capture, receiving signals through a broadband reconfigurable RF front-end, generating I / Q two-channel baseband signals using a dual-channel quadrature down-conversion structure, and the dynamic adjustment range of the sampling rate is 5 - 100 MS / s; S2. Multi-dimensional feature extraction: Perform time-frequency domain, modulation domain, and statistical domain analyses on the baseband signal synchronously, and extract feature indicators including the energy concentration degree of the short-time Fourier transform, cyclic spectrum symmetry, etc. S2a. Environment perception-driven feature alignment: Construct a prior interference environment model using spectral images, historical interference labels, and geographical location information, and perform distributed alignment and dimensionality reduction on the feature vectors through a graph-structure-based semantic embedding algorithm to improve feature discrimination ability and environmental adaptability. S3. Malicious signal detection: Construct a classification model based on an improved deep residual network, which embeds a channel attention module on the basis of ResNet-18, inputs the aligned feature vectors, outputs the probability estimate of malicious signals, and the detection threshold is dynamically set to 0.65 - 0.85. S3a. Abnormality perception and model self-learning mechanism: Establish a confidence evaluation mechanism between the detection results and historical sample labels, use an anomaly detection module to judge the credibility of samples based on feature density and reconstruction error, automatically trace back to the original signal when the credibility is lower than the preset threshold, and trigger an incremental training process to dynamically update the classification model parameters to achieve unsupervised online adaptive optimization. S4. Adaptive filtering processing and security protection verification: Adopt a three-stage cascaded filtering architecture. The first stage is spatial beamforming based on a 16-element virtual array, the second stage is a Volterra filter including third-order non-linear terms, and the third stage is a blind source separation module based on the joint diagonalization algorithm. The coefficients of the three-stage filters are updated online through the conjugate gradient method. Implement digital signature verification and whitelist filtering mechanisms on the FPGA hardware platform, and perform propagation path backtracking and transmitter feature matching on the filtered signals. The matching includes signal fingerprint matching, propagation path consistency verification, and transmitter certificate verification.

[0006] A method for identifying malicious signals and adaptive filtering in the radio high-frequency band of the present invention constructs an intelligent processing system integrating sensing, recognition, suppression, and protection around the processing chain of "signal perception - multi-dimensional feature extraction - environment modeling - deep recognition - anomaly self-learning - cascaded filtering - source verification". In terms of the working principle, the system first receives and down-converts through a reconfigurable radio frequency structure to generate I / Q signals, and then performs multi-dimensional feature extraction including short-time Fourier transform, cyclic spectrum, and high-order statistic analysis. Subsequently, a graph neural network-driven environment modeling mechanism is introduced to perform distributed alignment and dimensionality reduction on the feature space to improve the robustness of the model in multi-scenarios. In the recognition stage, a deep residual network embedded with an attention module is used to classify the signals and output the malicious probability estimate. If the confidence level of the detection result is low, the original signal is traced back and retrained to realize the online self-learning optimization of the classification model. After signal recognition, interference suppression is performed through a three-stage filtering architecture composed of spatial domain beam, non-linear Volterra filtering, and blind source separation, and the filter parameters are dynamically adjusted online. Finally, digital signature verification, fingerprint comparison, and path consistency analysis are completed on the FPGA platform to perform secure traceability of the signal source, effectively solving problems such as high false alarm rate in fixed threshold detection, difficulty in suppressing non-linear interference by the filtering mechanism, lack of adaptability of the model, and weak source verification means, and realizing accurate recognition, deep suppression, and security defense of malicious signals in complex high-frequency electromagnetic environments.

[0007] Specifically, the multi-dimensional feature extraction in step S2 includes: calculating the energy entropy of the time-frequency matrix using the short-time Fourier transform with an overlap rate of 80% for time-frequency domain features, extracting the symmetry index of the cyclic spectrum density function for modulation domain features, and calculating the normalized skewness value of the fourth-order cumulant for statistical domain features.

[0008] Specifically, the environment modeling in step S2a constructs a prior model using a multi-source data fusion method including spectrum image, interference type, and geographical location information, and performs alignment and dimensionality reduction processing on the features through graph structure learning.

[0009] Specifically, the improved deep residual network in step S3 uses ResNet-18 as the basic structure and embeds a channel attention module at the backend of its convolutional layer to enhance the high-weight feature expression ability.

[0010] Specifically, the anomaly perception mechanism in step S3a calculates the confidence level based on the joint analysis of the historical sample distribution density and the feature reconstruction error, and the confidence threshold is set to 0.4.

[0011] Specifically, when the confidence level in step S3a is lower than the preset threshold, the system quickly traces back the original I / Q signal and re-extracts the features, and calls the incremental learning module to adjust the parameters of the classification network.

[0012] Specifically, the three - stage cascaded filtering structure in step S4 includes: the spatial resolution of the spatial domain beamforming module is better than 5°, the Volterra filter is a third - order structure, and the blind source separation module adopts the joint diagonalization algorithm with a separation error less than 0.1 dB.

[0013] Specifically, the filter coefficients are updated online using the conjugate gradient method with an update period not exceeding 1 millisecond to adapt to the complex electromagnetic interference environment.

[0014] Specifically, in the security protection verification, the digital signature verification adopts the ECDSA algorithm, the correlation coefficient of the signal fingerprint matching is higher than 0.9, and the time delay difference of the propagation path consistency verification is less than 100 nanoseconds.

[0015] Specifically, the FPGA hardware platform integrates a dynamically partially reconfigurable module, supports the rapid switching of three processing modes of feature extraction, filtering processing, and signal verification, and the reconfiguration time does not exceed 10 milliseconds.

[0016] Additional aspects and advantages of the present invention will be given in part in the following description, will become apparent in part from the following description, or will be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The above - mentioned and / or additional aspects and advantages of the present invention will become apparent and easy to understand from the following description of the embodiments in conjunction with the drawings, where: Figure 1 is a flowchart of the method for identifying malicious signals and adaptive filtering in the radio high - frequency band of the present invention; Figure 2 is a data diagram of Embodiment 1 of the method for identifying malicious signals and adaptive filtering in the radio high - frequency band of the present invention; Figure 3 is a data diagram of Embodiment 2 of the method for identifying malicious signals and adaptive filtering in the radio high - frequency band of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are intended to explain the present invention, and should not be construed as a limitation of the present invention. On the contrary, the embodiments of the present invention include all changes, modifications, and equivalents falling within the spirit and scope of the appended claims.

[0019] The method for identifying malicious signals and adaptive filtering in the radio high - frequency band of the embodiments of the present invention will be described below in conjunction with the drawings.

[0020] As Figures 1 - 3As shown, the method for identifying malicious signals and adaptive filtering in the radio high-frequency band according to the embodiments of the present invention includes: S1. Signal perception and capture: Receive signals through a broadband reconfigurable RF front-end, and generate I / Q two-channel baseband signals using a two-channel quadrature down-conversion structure. The dynamic adjustment range of the sampling rate is 5 - 100 MS / s.

[0021] It should be noted that in the signal perception and capture stage described in this embodiment, signals are received through a broadband reconfigurable RF front-end (frequency coverage: 3 - 30 MHz, instantaneous bandwidth ≥ 10 MHz, dynamic range ≥ 70 dB), and I / Q two-channel baseband signals are generated using a two-channel quadrature down-conversion structure. Among them, the local oscillator frequency is precisely controlled by a fractional-N phase-locked loop (PLL), with phase noise ≤ -110 dBc / Hz @ 10 kHz. At the same time, an automatic gain control (AGC) module is integrated, with a dynamic adjustment range of -40 dB to +30 dB to ensure stable signal strength. The dynamic adjustment range of the sampling rate is 5 - 100 MS / s, and frequency fine-tuning is realized by a numerically controlled oscillator (NCO), with a step accuracy ≤ 1 Hz. Digital down-conversion (DDC) processing is implemented in the FPGA, and the decimation factor can be configured as 1 - 64 to adapt to different bandwidth requirements. The RF front-end adopts a multi-stage filtering architecture, including a pre-bandpass filter (insertion loss ≤ 1 dB), an image rejection filter (rejection ratio ≥ 60 dB), and an anti-aliasing filter (stopband attenuation ≥ 80 dB) to ensure signal purity. In addition, by real-time monitoring the signal power spectral density (PSD) and signal-to-noise ratio (SNR), the sampling rate and gain parameters are dynamically adjusted to achieve optimal signal capture performance.

[0022] S2. Multi-dimensional feature extraction: Synchronously perform time-frequency domain, modulation domain, and statistical domain analysis on the baseband signals, and extract feature indicators including short-time Fourier transform energy concentration, cyclic spectrum symmetry, etc. S2a. Feature alignment driven by environmental perception: Construct a prior interference environment model using spectral images, historical interference labels, and geographical location information, and perform distributed alignment and dimensionality reduction on the feature vectors through a graph-structure-based semantic embedding algorithm to improve feature discrimination ability and environmental adaptability.

[0023] It should be noted that in this embodiment, for the S2. multi-dimensional feature extraction and feature alignment driven by environmental perception, time-frequency domain, modulation domain, and statistical domain analyses are performed on the I / Q baseband signal synchronously. A 17-dimensional composite feature vector including the energy aggregation degree of the short-time Fourier transform (STFT), cyclic spectrum symmetry, skewness of the high-order cumulant, amplitude histogram features, etc. is constructed. Among them, for the time-frequency domain features, the STFT with an overlap rate of 80% and a window length of 256 points is used to calculate the time-frequency matrix, and the energy entropy and instantaneous frequency variance are extracted; for the modulation domain features, based on the cyclic spectrum density function at 1 / 4 of the carrier frequency offset, the symmetry index and the cyclic stationarity parameter α are extracted; the statistical domain features include the normalized skewness value, kurtosis of the fourth-order cumulant, and the amplitude distribution histogram under 2048-point sampling. To improve the expression ability and robustness of the feature vector in complex environments, a feature alignment mechanism driven by environmental perception is further introduced. An environmental prior model is constructed using the spectrum image, interference label, and geographical location information, and a graph structure semantic map of multi-source information is formed through a graph neural network (GNN). Graph convolution operations are used to extract global feature embeddings, and the graph attention mechanism is combined to weight-update the feature nodes to complete the distributed alignment and dimensionality reduction processing of the feature space across environmental scenarios; the GNN module supports periodic updates every 30 seconds to adapt to the dynamic changes of the spectrum environment. The feature selection module dynamically adjusts the feature weights based on the information gain ratio, and improves the consistency and discrimination of the feature input through normalization and principal component analysis to ensure the input stability and detection accuracy of the subsequent deep classification model.

[0024] S3. Malicious signal detection, constructing a classification model based on an improved deep residual network, which embeds a channel attention module on the basis of ResNet-18, inputs the aligned feature vector, and outputs the probability estimate of the malicious signal. The detection threshold is dynamically set to 0.65 - 0.85; S3a. Abnormality perception and model self-learning mechanism, by establishing a confidence evaluation mechanism between the detection result and the historical sample label, using the anomaly detection module to judge the sample credibility based on the feature density and reconstruction error. When the credibility is lower than the preset threshold, the original signal is automatically traced back, and the incremental training process is triggered to dynamically update the classification model parameters to achieve unsupervised online adaptive optimization.

[0025] It should be noted that in this embodiment, for the S3. Malicious signal detection and model self-learning mechanism, a joint classification model based on an improved deep residual network is first constructed. This network uses ResNet-18 as the backbone structure, and a channel attention mechanism module (Squeeze-and-Excitation Block, SEBlock) is embedded after each residual module. Through global average pooling and per-channel scaling, weighted regulation of the feature dimension is achieved, thereby enhancing the expression ability of important feature channels and improving the sensitivity to fine-grained signal features. The network accepts a 17-dimensional feature vector aligned from the previous environment as input, outputs the probability estimate of the corresponding malicious signal, and the detection threshold is dynamically set in the range of 0.65 to 0.85, and can be adaptively adjusted through the Q-learning reinforcement learning algorithm according to the change of the electromagnetic environment. To enhance the system's ability to identify unknown or mutated malicious signals, an anomaly perception and model self-learning mechanism is further introduced. A confidence evaluation module is constructed based on the historical sample library. This module comprehensively considers the feature density estimate of the current sample and the reconstruction error calculated based on the variational autoencoder (VAE), and fuses the two into a joint confidence index. When the confidence of the detected sample is lower than the preset threshold of 0.4, the system automatically triggers a backtracking mechanism, reverts to the original I / Q signal level and reextracts the feature vector, and at the same time enables a lightweight incremental training module to fine-tune and update the deep classification model. The model update adopts a sliding window strategy to ensure sample quality and training efficiency, supports the model self-evolution process without introducing manual annotation, and the overall delay does not exceed 50 ms, effectively improving the system's generalization ability, adaptive ability and response ability to rare interference patterns in a complex electromagnetic environment.

[0026] S4. Adaptive filtering processing and security protection verification. A three-stage cascaded filtering architecture is adopted. The first stage is spatial beamforming based on a 16-element virtual array, the second stage is a Volterra filter containing third-order non-linear terms, and the third stage is a blind source separation module based on the joint diagonalization algorithm. The coefficients of the three-stage filter are updated online through the conjugate gradient method. Digital signature verification and whitelist filtering mechanisms are implemented on the FPGA hardware platform, and the propagation path of the filtered signal is traced back and the transmitter characteristics are matched. The matching includes signal fingerprint matching, propagation path consistency verification and transmitter certificate verification.

[0027] It should be noted that in this embodiment, the adaptive filtering process and security protection verification adopt a three-level cascaded filtering architecture in cooperation with a multi-dimensional security mechanism. The filtering part includes: the first level is the spatial domain beamforming module, which constructs a spatial domain pattern based on a 16-element virtual array, realizes the direction positioning and interference suppression of spatially distributed interference signals through steering vector construction and covariance matrix reconstruction, with a spatial resolution better than 5° and a sidelobe suppression ratio not lower than 30 dB; the second level is the Volterra filter module containing third-order nonlinear terms, which models non-Gaussian interferences such as frequency sweep and amplitude modulation using a nonlinear kernel function, sets the filter memory depth to 10 symbol periods, updates the weights through the least mean square error algorithm, and has the capabilities of nonlinear suppression and dynamic order reduction; the third level is the blind source separation module based on the joint diagonalization algorithm, which uses the JADE or FastICA algorithm to perform unsupervised separation of mixed signals, with the separation error controlled within 0.1 dB; the above three-level filtering modules cooperate in series to form a cascaded filtering link, and online update all filter weights through the conjugate gradient method, with an update period less than 1 ms, and has the ability to dynamically adapt to different interference scenarios. In terms of security protection verification, multiple security verification modules are deployed based on the FPGA hardware platform, including three sub-functions: digital signature verification, whitelist filtering, and emission source tracking; among them, the digital signature uses the elliptic curve digital signature algorithm (ECDSA) with a key length of 256 bits to ensure the legitimacy of the signal source; the whitelist mechanism is compared based on the pre-registered identity fingerprint of the transmitting device and enables fast verification through hardware; in the signal tracking link, the propagation path of the filtered signal is traced back and the emission source features are matched, specifically including physical layer signal fingerprint matching based on carrier frequency offset and error vector magnitude (EVM) (correlation coefficient higher than 0.9), propagation path consistency verification based on the joint positioning of time difference of arrival (TDOA) and Doppler frequency shift (time delay difference less than 100 nanoseconds, positioning error not exceeding 50 meters), and emission source certificate verification recorded based on the blockchain deposit mechanism. The triple mechanism jointly completes the source confirmation and anti-spoofing verification of malicious signals; when any verification fails, the system automatically triggers the spectrum interference injection mechanism, emits a co-frequency interference signal (power adjustable range -20 dBm to +10 dBm) through the RF synthesis module configured in the FPGA, and the continuous interference time is automatically adjusted according to the threat level (100 ms to 10 s), forming a closed-loop anti-attack protection system, significantly improving the system's recognition accuracy, filtering depth, and signal security in complex scenarios.

[0028] Specifically, by constructing a full - process processing link that includes signal perception, feature extraction, semantic alignment, intelligent detection, self - learning optimization, filtering suppression, and security protection, the efficient recognition and source tracing of malicious signals in a complex interference environment in the high - frequency band are realized, significantly solving the problems existing in the prior art, such as fixed thresholds, weak non - linear interference suppression ability, poor real - time performance, and insufficient anti - spoofing ability. First, in the signal perception and capture stage, the system adopts a broadband reconfigurable RF front - end and a dual - channel quadrature down - conversion structure to dynamically receive high - frequency signals, supporting a flexible sampling rate of 5 - 100 MS / s to adapt to different electromagnetic bandwidth requirements. Subsequently, it enters the multi - dimensional feature extraction module, which jointly analyzes the I / Q signals in the time - frequency domain, modulation domain, and statistical domain, and extracts various feature indicators including the energy aggregation degree of the short - time Fourier transform, cyclic spectrum symmetry, etc., covering complex patterns such as non - Gaussian interference and swept - frequency signals. To further improve the generalization ability of the model to different electromagnetic environments, the system introduces an environment - perception - driven feature alignment mechanism, constructs a prior environment model using multi - source data such as spectral images, geographical locations, and historical interference labels, and realizes the distributed alignment and dimensionality reduction of the feature space among multiple scenarios through semantic embedding and graph - structure propagation of the graph neural network, thus solving the problem of the sharp drop in the recognition rate of traditional models during environment switching. In the malicious signal detection link, the present invention constructs an improved deep residual network as the main classification model, integrates a channel attention module based on ResNet - 18, enables the model to strengthen the expression of key features when facing complex interference patterns, improves the detection accuracy, and adapts the output judgment in different scenarios through dynamic threshold adjustment. Aiming at the problem that the existing models lack feedback and self - adaptation ability, the system further introduces an anomaly perception and self - learning mechanism, establishes a confidence evaluation system, integrates the feature density estimation and the reconstruction error of the variational auto - encoder. When the confidence is lower than the preset threshold, it automatically triggers sample backtracking and incremental training to realize the unsupervised online evolution of the classification model, effectively overcoming the limitation that fixed models cannot adapt to channel changes and the new types of interference. At the back - end of signal processing, to ensure the filtering effect and system response speed, the system designs a three - stage cascaded adaptive filtering architecture, which respectively uses spatial beamforming, non - linear Volterra filters, and blind source separation modules to work together, and combines the conjugate gradient method to realize the real - time update of the filter parameters, significantly improving the comprehensive suppression ability of spatial interference, non - linear interference, and mixed - source signals. In terms of security protection, the system relies on the FPGA platform to implement a hardware - level digital signature verification and white - list filtering mechanism, and performs propagation path backtracking and transmitter source verification after the signal is filtered. Through triple authentication means of signal fingerprints, path consistency, and certificates, it completes source right confirmation and anti - forgery defense, thus solving the problem that existing systems are easily deceived by forged signals and lack a source tracing mechanism, and finally realizing an efficient, accurate, and traceable high - frequency band malicious signal recognition and processing solution.

[0029] Furthermore, as Figure 1As shown, the multi-dimensional feature extraction in step S2 includes: for time-frequency domain features, the energy entropy of the time-frequency matrix is calculated using the short-time Fourier transform with an 80% overlap rate; for modulation domain features, the symmetry index of the cyclic spectral density function is extracted; for statistical domain features, the normalized skewness value of the fourth-order cumulant is calculated.

[0030] It should be noted that in this embodiment, for the time-frequency domain features, the energy entropy of the time-frequency matrix, the instantaneous frequency change rate, and the spectral center drift are calculated using the short-time Fourier transform with an 80% overlap rate. A Hamming window with a window length of 256 points is used to perform local windowing on the signal in the time domain to ensure the spectral detail resolution. At the same time, the frame shift overlap method is adopted to enhance the ability to capture the dynamic changes of the spectrum; the modulation domain feature extraction includes the symmetry index of the cyclic spectral density function, the cyclic stationarity test coefficient, and the normalized modulation degree change range. In the cyclic spectrum calculation, a complex exponential kernel function is used and the modulation cyclic component at 1 / 4 of the carrier frequency is introduced as the main criterion to enhance the recognition ability of AM, FM, and frequency sweep type modulation signals; for the statistical domain features, the high-order statistics of the baseband amplitude sequence are comprehensively considered, the normalized skewness value and kurtosis of the fourth-order cumulant are calculated, and the amplitude histogram distribution density function under 2048-point equidistant sampling is introduced. The non-Gaussianity and non-symmetry levels of the signal are judged by the density center of gravity and the degree of tail extension. The above features are combined by the feature fusion module to construct a 17-dimensional multi-component feature vector, and in the subsequent feature selection stage, the sub-feature set is dynamically screened according to the maximum mutual information criterion to improve the effectiveness and representativeness of the subsequent model input, so as to ensure that the entire system has high-robust front-end feature extraction ability in the scenarios of dealing with complex background interference, multi-type signal mixing, and mutant malicious signals.

[0031] Furthermore, as Figure 1 shown, the environmental modeling in step S2a uses a multi-source data fusion method including spectral image, interference type, and geographical location information to construct a prior model, and the features are aligned and dimension-reduced through graph structure learning.

[0032] It should be noted that in this embodiment, the environmental modeling in step S2a constructs a prior model by using a multi-source data fusion method that includes spectral images, interference types, and geographical location information. First, the evolution characteristics of interference patterns in the time domain and frequency domain are extracted from the spectral image sequence in the historical record, and their spatial distribution and frequency aggregation patterns are extracted through a two-dimensional convolutional neural network. Then, a perturbation label matrix is constructed in combination with interference type labels (such as pulse interference, continuous wave, frequency sweep perturbation, etc.), and the geographical location information (latitude and longitude, terrain annotation, estimated emission source area, etc.) at the time of signal acquisition is introduced as an auxiliary attribute to form heterogeneous graph data containing topological structure, node semantics, and spatial attributes. On this basis, an embedding model is constructed for the multi-source features extracted by the graph structure learning method. A graph neural network (GNN) is used to construct the semantic association edge weight matrix between nodes, and the feature information of adjacent nodes is aggregated by graph convolution operations to form a global environmental perception representation vector. The graph attention mechanism (GAT) is used to assign higher weights to important nodes to highlight key interference features, thus completing the distributed alignment and dimensionality reduction processing of the original multi-dimensional feature vector. This graph structure model supports periodic updates according to changes in the current electromagnetic environment. The update period is default set to 30 seconds. The latest spectral state data is continuously absorbed through the window sliding mechanism, and the edge weights and embedding parameters are updated, so as to enhance the stability and environmental adaptability of the feature representation under cross-region, cross-time period, and cross-interference type conditions, and provide a more consistent, anti-interference, and transferable input feature expression for the subsequent malicious signal detection module.

[0033] Further, as Figure 1 shown, the improved deep residual network in step S3 uses ResNet-18 as the basic structure, and a channel attention module is embedded at the backend of its convolutional layer to enhance the high-weight feature expression ability.

[0034] It should be noted that in this embodiment, in step S3, the improved deep residual network uses ResNet-18 as the basic structure, and a channel attention module is embedded at the backend of the convolutional layer in each residual block (Residual Block) to enhance the high-weight feature expression ability. The channel attention module is constructed based on the Squeeze-and-Excitation mechanism (SE module). Through global average pooling operation, the features of each channel are compressed, and after extracting its global statistical information, a non-linear inter-channel dependence relationship is constructed through two fully connected layers. The channel weight coefficient is output through the Sigmoid activation function, and the original feature map is weighted channel by channel to enhance the key feature expression and suppress redundant features. This module is integrated into each residual unit of the ResNet backbone network, effectively making up for the lack of the original ResNet structure in identifying fine-grained feature differences, enabling the model to have higher recognition sensitivity to weak features in malicious signals such as non-linear perturbations and variant carrier frequency jitters. In addition, to adapt to the feature differences in different electromagnetic environments, the model supports dynamically adjusting the scaling factor and compression ratio of the attention module. The default compression ratio is 16, and this value can be adaptively adjusted through environmental perception parameters to optimize the trade-off between computational cost and recognition accuracy. When the entire network is trained, the cross-entropy loss function is used and supplemented with the L2 regularization term to enhance the generalization ability of the model. The final output of the model is the probability estimation of the malicious signal, and together with the dynamically set detection threshold (automatically adjusted between 0.65 and 0.85), it realizes the classification and decision of signals in different scenarios, effectively improving the recognition accuracy and robustness of the system in a high-dynamic interference environment.

[0035] Furthermore, as Figure 1 shown, in step S3a, the anomaly perception mechanism calculates the confidence based on the joint analysis of the historical sample distribution density and the feature reconstruction error, and the confidence threshold is set to 0.4.

[0036] It should be noted that in step S3a described in this embodiment, the anomaly perception mechanism calculates the confidence based on the joint analysis of the historical sample distribution density and the feature reconstruction error. First, a historical sample library is constructed, and density modeling is performed on the feature vectors of known normal and abnormal signals. The probability density function is constructed by using the kernel density estimation (KDE) method to describe the distribution trend of the feature space. For the input sample, the local density index under this distribution is calculated as the first sub-component of its confidence. At the same time, a variational autoencoder (VAE) is introduced as an unsupervised reconstruction model to compress and encode the input features and then reconstruct them. The reconstruction error is obtained by calculating the mean square error between the original features and the reconstructed features, and this error is used as the second sub-component of the confidence, indicating the deviation degree of the sample from the training distribution. The system composes a joint confidence index by weighted fusion of the density score and the reconstruction error score, and sets a dynamically updated fusion weight to adapt to different environmental characteristics. Finally, when the comprehensive confidence of the sample is lower than the set threshold of 0.4, it is determined as a low-confidence sample. At this time, the anomaly response mechanism is automatically triggered, the original I / Q signal corresponding to this sample is traced back, the feature extraction and alignment process is re-executed, and the incremental learning module is activated to fine-tune and update the main classification model. This incremental learning process uses a sliding window mechanism to collect the recent confidence anomaly sample set and performs backpropagation in a mini-batch manner to ensure the stability and real-time performance of the model parameter adjustment. The entire process can complete online adaptive evolution without the participation of manual annotation, thereby enhancing the system's continuous learning and anti-interference capabilities against unknown interference, new camouflage signals, and electromagnetic environment changes, and significantly improving the generalization performance and robustness of the overall system in dynamic scenarios.

[0037] Furthermore, as Figure 1 shown, when the confidence in step S3a is lower than the preset threshold, the system quickly traces back the original I / Q signal and reextracts the features, and calls the incremental learning module to adjust the parameters of the classification network.

[0038] It should be noted that in this embodiment, when the confidence level in step S3a is lower than the preset threshold, the system retraces the original I / Q signal quickly and re - extracts features, and invokes the incremental learning module to adjust the parameters of the classification network. First, the system triggers the retracing mechanism to re - extract the I / Q signal segment corresponding to the current low - confidence sample in the original sampling buffer, and regenerates the complete time - frequency domain, modulation domain, and statistical domain feature vectors based on the latest multi - dimensional feature extraction algorithm. At the same time, it performs feature alignment processing with reference to the current environment perception model to ensure the consistency and context matching of the input data. Subsequently, the incremental learning module is started. This module takes the low - confidence samples processed in a labeled or weakly supervised manner as the core, combines the historical high - confidence samples within the sliding time window to construct a small - sample training set, and uses a lightweight fine - tuning strategy to incrementally optimize the parameters of the last few layers of the deep classification network. Among them, the front - layer feature extraction layer is frozen, and only the channel weight coefficients in the decision layer and the attention module are updated to reduce the risk of model drift. During the training process, the Adam optimizer based on learning rate annealing is used to control the parameter adjustment rate. At the same time, the confidence change rate is introduced as a stop - condition judgment index. When the confidence improvement amplitude is less than 1% in three consecutive training rounds, the update is terminated to avoid overfitting and waste of computing resources. This incremental update process can be executed in parallel on the local edge device or the soft - core processor integrated in the FPGA, and the overall delay is controlled within 100 milliseconds, effectively realizing the fast response and adaptation of the model to new or mutated interference signals, thereby ensuring that the system still maintains a high recognition accuracy and real - time performance in a high - dynamic and high - adversarial radio environment.

[0039] Furthermore, as Figure 1 shown, the three - stage cascaded filtering structure in step S4 includes: the spatial resolution of the spatial beamforming module is better than 5°, the Volterra filter is a third - order structure, and the blind source separation module uses the joint diagonalization algorithm with a separation error less than 0.1 dB.

[0040] It should be noted that the three - stage cascaded filtering structure in step S4 described in this embodiment includes: the spatial resolution of the spatial domain beamforming module is better than 5°, the Volterra filter is a third - order structure, the blind source separation module adopts the joint diagonalization algorithm, and the separation error is less than 0.1 dB. Among them, the first - stage spatial domain beamforming module is constructed based on a 16 - element virtual uniform linear array. The spatial domain direction pattern is jointly constructed through the steering vector and the received covariance matrix. The minimum variance distortionless response (MVDR) algorithm is used to enhance the main lobe in the target direction and form a null in the interference direction. The array element weighting coefficients are dynamically adjusted to improve the suppression ability for interfering sources with angles close to each other. The spatial resolution is better than 5°, and the sidelobe suppression ratio reaches 35 dB, effectively isolating the mixed signals of multiple near - angle emitters; the second stage is a Volterra filter containing third - order non - linear terms. The extended Volterra series is used to model non - Gaussian non - linear perturbations. A non - linear kernel function with a memory depth of 10 symbol periods is introduced into the filtering structure to capture the polynomial feature components in frequency - swept and amplitude - modulated interference signals. The coefficients are solved in combination with the least mean square error criterion, and the filtering intensity is adjusted through a step - size adaptive mechanism to ensure the suppression of strong non - linear interference while maintaining the integrity of the target signal; the third stage is a blind source separation module based on the joint diagonalization algorithm. A joint diagonalization model is constructed using multiple groups of time - delay correlation matrices. The JADE (Joint Approximate Diagonalization of Eigen - matrices) algorithm is used to perform the demixing operation of multiple source signals. The separation error is controlled within 0.1 dB, and an orthogonalization reconstruction mechanism is added after the demixing output to enhance the signal restoration degree. This blind source separation module is particularly suitable for typical scenarios such as an uncertain number of emitters in the electromagnetic spectrum, a complex channel response, or the existence of cooperative interference, etc.; the above three - stage filter structure forms a series relationship according to the spatial domain - non - linear - source level, and the online optimization and update of the filter parameters are realized through a unified conjugate gradient algorithm framework. The update period does not exceed 1 ms, ensuring that the entire filtering system has low latency, high robustness, and strong suppression ability, and can operate in a complex and changeable high - frequency radio environment for a long time without human intervention.

[0041] Furthermore, as Figure 1 shown, the filter coefficients are updated online using the conjugate gradient method, and the update period does not exceed 1 millisecond to adapt to the complex electromagnetic interference environment. In the security protection verification, the ECDSA algorithm is used for digital signature verification, the correlation coefficient of signal fingerprint matching is higher than 0.9, and the time - delay difference of the propagation path consistency verification is less than 100 nanoseconds. The FPGA hardware platform integrates a dynamic partial reconfiguration module, supporting fast switching among three processing modes: feature extraction, filtering processing, and signal verification. The reconfiguration time does not exceed 10 milliseconds.

[0042] It should be noted that in this embodiment, the filter coefficients are optimized and updated online by the conjugate gradient method, and quickly converge through the residual minimization strategy, supporting a complete weight adjustment within 1 millisecond to ensure that the system still maintains high-efficiency interference suppression performance when the spectrum environment fluctuates violently. The security protection verification mechanism uses the ECDSA elliptic curve digital signature algorithm to verify the identity of the signal source, and combines the physical layer fingerprint matching technology to extract the carrier frequency offset and error vector magnitude characteristics for comparison, and the correlation coefficient is stably higher than 0.9. The propagation path consistency verification module jointly estimates the signal source position based on the time difference of arrival (TDOA) and Doppler frequency shift of multiple receiving nodes, supports nanosecond-level delay determination, and the path deviation is controlled within 100 nanoseconds. The FPGA hardware platform integrates a partially dynamically reconfigurable logic array, which can achieve rapid switching between three modes of feature extraction, filtering processing and security verification within 10 milliseconds, ensuring that the system has high real-time performance and high reliability in task switching, function reconstruction and dynamic resource allocation.

[0043] Embodiment 1: Malicious signal interception in the high-speed maneuvering environment of an unmanned aerial vehicle Application scenario: When a certain type of reconnaissance unmanned aerial vehicle is performing tasks in the urban area, it encounters a malicious interference signal in the high-frequency band (15 MHz), and the signal power fluctuates in the range of -100 dBm to -50 dBm, accompanied by sweep interference (frequency change rate 2 MHz / μs).

[0044] Implementation details of the technical solution: Signal sensing and capture: The radio frequency front end uses the ADI AD9361 chip (frequency coverage 3 - 30 MHz, instantaneous bandwidth 20 MHz), and generates I / Q signals through dual-channel down-conversion. The sampling rate is dynamically adjusted to 80 MS / s, and the signal is purified through multi-stage filtering (out-of-band rejection ≥ 70 dB).

[0045] The FPGA realizes digital down-conversion, and the decimation factor is set to 8, and the signal-to-noise ratio of the output baseband signal is ≥ 25 dB.

[0046] Malicious signal detection: Extract a 17-dimensional feature vector: time-frequency domain energy entropy (STFT window length 256 points), cyclic spectrum correlation coefficient (frequency offset 3.75 MHz), fourth-order cumulant skewness (sampling 2048 points).

[0047] Classify using an improved ResNet-18 model (adding an SE attention module), and the detection threshold is dynamically set to 0.72. The double-threshold trigger condition: power density > -70 dBm and cyclic stationarity α = 0.65.

[0048] Adaptive filtering: First level: 16 - element virtual beamforming, spatial resolution 3°, suppressing sidelobe interference (≥35dB); Second level: Third - order Volterra filter (memory depth 10 symbols), dynamically reduced to first - order to cope with swept - frequency interference; Third level: Blind source separation (JADE algorithm), separation error 0.08dB.

[0049] Security verification and counter - measure: Signal fingerprint matching (correlation coefficient 0.93), TDOA positioning error 42 meters, injecting co - channel interference (+5dBm, lasting 5s) after ECDSA verification passes.

[0050] Effect comparison (as Figure 2 shown).

[0051] Example 2: Anti - spoofing protection in complex urban electromagnetic environment Application scenario: The base station in the city center is attacked by forged signals. The signals are disguised as legitimate users (carrier frequency offset 9ppm, EVM = 3.5%), accompanied by multipath interference (delay spread 200ns).

[0052] Implementation details of the technical solution: Electromagnetic environment awareness: The database real - time collects spectrum data (occupancy rate > 80%), dynamically adjusts the detection threshold to 0.78 through Q - learning, and compresses the filter update period to 20 seconds.

[0053] DNN - assisted feature extraction, identifying the features of spoofing signals (the detection sensitivity of EVM exceeding the standard reaches 0.5%).

[0054] Hybrid signal separation: Joint spatial - transform domain separation: Beamforming (beam width 4°) cascaded with blind source separation, the separation degree is increased by 10dB, and the signal - to - interference - plus - noise ratio is increased from 15dB to 32dB.

[0055] Anti - spoofing verification: Physical - layer fingerprint verification: Carrier frequency offset detection (8ppm < 10ppm), EVM = 3.8% (exceeding the threshold), triggering blockchain forensics (Hyperledger Fabric platform), timestamp accuracy 0.8μs.

[0056] Dynamically reconfiguring FPGA: Switch to anti - spoofing mode, reconfiguration time 7ms, running fingerprint extraction (delay 3μs) and interference injection module in parallel.

[0057] Cooperative positioning: Three-base-station TDOA (spacing 1.5 km) + Doppler frequency shift correction, positioning error of 38 m after Kalman filtering, path backtracking matching error < 80 ns.

[0058] Effect comparison (as Figure 3 shown).

[0059] In summary, an intelligent processing system integrating sensing, recognition, suppression, and protection has been constructed around the processing chain of "signal perception - multi-dimensional feature extraction - environment modeling - deep recognition - anomaly self-learning - cascaded filtering - source verification". In terms of working principle, the system first receives and down-converts through a reconfigurable radio frequency structure to generate I / Q signals, and then performs multi-dimensional feature extraction including short-time Fourier transform, cyclic spectrum, and high-order statistics analysis. Subsequently, an environment modeling mechanism driven by a graph neural network is introduced to perform distributed alignment and dimensionality reduction on the feature space to improve the robustness of the model in multiple scenarios. In the recognition stage, a deep residual network embedded with an attention module is used to classify the signals and output a malicious probability estimate. If the confidence level of the detection result is low, the original signal is traced back and retrained to achieve online self-learning optimization of the classification model. After signal recognition, interference suppression is performed through a three-stage filtering architecture composed of spatial domain beams, non-linear Volterra filtering, and blind source separation, and the filter parameters are dynamically adjusted online. Finally, digital signature verification, fingerprint comparison, and path consistency analysis are completed on the FPGA platform to perform safe traceability of the signal source. This solution effectively solves problems such as high false alarm rate of fixed threshold detection, difficulty in suppressing non-linear interference by the filtering mechanism, lack of adaptability of the model, and weak source verification means, and realizes precise recognition, deep suppression, and security defense of malicious signals in a complex high-frequency electromagnetic environment.

[0060] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for identifying malicious signals and adaptive filtering in the radio high-frequency band, characterized in that It includes the following steps: S1. Signal sensing and capture: Receive signals through a broadband reconfigurable RF front-end, generate I / Q two-channel baseband signals using a dual-channel quadrature down-conversion structure, and the dynamic adjustment range of the sampling rate is 5 - 100 MS / s; S2. Multi-dimensional feature extraction: Synchronously perform time-frequency domain, modulation domain, and statistical domain analysis on the baseband signals, and extract feature indicators including the energy aggregation degree of the short-time Fourier transform, cyclic spectrum symmetry, etc.; S2a. Feature alignment driven by environment perception: Construct a prior interference environment model using spectral images, historical interference labels, and geographical location information, and perform distributed alignment and dimensionality reduction on the feature vectors through a graph structure-based semantic embedding algorithm to enhance the feature discrimination ability and environmental adaptability; S3. Malicious signal detection: Construct a classification model based on an improved deep residual network, which embeds a channel attention module on the basis of ResNet-18, input the aligned feature vectors, output the probability estimate of malicious signals, and the detection threshold is dynamically set to 0.65 - 0.85; S3a. Abnormality perception and model self-learning mechanism: Establish a confidence evaluation mechanism between the detection results and historical sample labels, use an anomaly detection module to judge the credibility of samples based on feature density and reconstruction error, automatically backtrack the original signal when the credibility is lower than the preset threshold, and trigger an incremental training process to dynamically update the classification model parameters to achieve unsupervised online adaptive optimization; S4. Adaptive filtering processing and security protection verification: Adopt a three-stage cascaded filtering architecture. The first stage is spatial domain beamforming based on a 16-element virtual array, the second stage is a Volterra filter containing third-order nonlinear terms, and the third stage is a blind source separation module based on the joint diagonalization algorithm. The coefficients of the three-stage filters are updated online through the conjugate gradient method. Implement digital signature verification and whitelist filtering mechanisms on the FPGA hardware platform, and perform propagation path backtracking and transmitter feature matching on the filtered signals. The matching includes signal fingerprint matching, propagation path consistency verification, and transmitter certificate verification.

2. The method for identifying malicious signals and adaptive filtering in the radio high-frequency band according to claim 1, wherein The multi-dimensional feature extraction in step S2 includes: The time-frequency domain feature calculates the energy entropy of the time-frequency matrix using the short-time Fourier transform with an 80% overlap rate, the modulation domain feature extracts the symmetry index of the cyclic spectrum density function, and the statistical domain feature calculates the normalized skewness value of the fourth-order cumulant.

3. The method for identifying malicious signals and adaptive filtering in the radio high-frequency band according to claim 1, characterized in that The environment modeling in step S2a uses a multi-source data fusion method including spectral images, interference types, and geographical location information to construct a prior model, and performs alignment and dimensionality reduction processing on the features through graph structure learning.

4. The malicious signal recognition and adaptive filtering method in the radio high-frequency band according to claim 1, characterized in that The improved deep residual network in step S3 uses ResNet-18 as the basic structure and embeds a channel attention module at the backend of its convolutional layer to enhance the high-weight feature expression ability.

5. The method for identifying malicious signals and adaptive filtering in the radio high-frequency band according to claim 1, wherein The abnormality perception mechanism in step S3a calculates the confidence based on the joint analysis of the historical sample distribution density and feature reconstruction error, and the confidence threshold is set to 0.

4.

6. The method for identifying malicious signals and adaptive filtering in the radio high-frequency band according to claim 1, characterized in that When the confidence in step S3a is lower than the preset threshold, the system quickly backtracks the original I / Q signal, re-extracts features, and calls the incremental learning module to adjust the parameters of the classification network.

7. The method for identifying malicious signals and adaptive filtering in the radio high-frequency band according to claim 1, wherein The three - stage cascaded filtering structure in step S4 includes: the spatial resolution of the spatial domain beamforming module is better than 5°, the Volterra filter is a third - order structure, and the blind source separation module uses the joint diagonalization algorithm with a separation error less than 0.1 dB.

8. The method for identifying malicious signals and adaptive filtering in radio high-frequency bands according to claim 1, characterized in that, The filter coefficients are updated online using the conjugate gradient method, and the update period does not exceed 1 millisecond to adapt to the complex electromagnetic interference environment.

9. The method for identifying malicious signals and adaptive filtering in radio high-frequency bands according to claim 1, wherein, In the security protection verification, the digital signature verification uses the ECDSA algorithm, the correlation coefficient of the signal fingerprint matching is higher than 0.9, and the time delay difference of the propagation path consistency verification is less than 100 nanoseconds.

10. The method for identifying malicious signals and adaptive filtering in the radio high-frequency band according to claim 1, characterized in that, The integrated dynamic partial reconfiguration module of the FPGA hardware platform supports the fast switching of three processing modes: feature extraction, filtering processing, and signal verification, and the reconfiguration time does not exceed 10 milliseconds.

Images