Poisoning signal detection method based on signal variation mode decomposition and reconstruction and electronic equipment

Abstract

The invention relates to a poisoning signal detection method based on signal variation mode decomposition reconstruction and electronic equipment, and belongs to the technical field of signal processing and computers.The method comprises the steps that wireless signals to be detected are input into a signal classification model, first prediction probability distribution is obtained, and the signal category is determined according to the first prediction probability distribution; performing variational mode decomposition on the first clean reference signals of the same type according to a preset mode total number; performing variational mode decomposition on the wireless signal according to the total number of modes; performing signal reconstruction according to the low-frequency intrinsic mode component of the wireless signal and the high-frequency intrinsic mode component of the first clean reference signal to obtain a reconstructed wireless signal; and inputting the reconstructed wireless signal into a signal classification model to obtain second prediction probability distribution of the reconstructed wireless signal, and determining whether the wireless signal is poisoned or not according to the difference between the first prediction probability distribution and the second prediction probability distribution. According to the method, efficient defense in a black box scene is realized, and the sensitivity of hidden attack detection and the adaptability of poisoning detection to dynamic attacks are improved.

Description

Technical Field

[0001] This invention relates to the fields of signal processing and computer technology, and in particular to a method and electronic device for detecting poisoning signals based on signal variational mode decomposition and reconstruction. Background Technology

[0002] Deep learning models, by automatically extracting task-relevant features, have demonstrated excellent performance in solving complex problems and have been widely applied in various fields in recent years, such as cybersecurity, computer vision, autonomous driving, language modeling, and signal processing. The rapid advancement and widespread application of deep learning technology have made its inherent security issues a critical challenge that urgently needs to be addressed. Early security problems with deep learning models appeared in malicious email detection systems, where attackers attempted to alter the model's decision boundaries by poisoning the data. In recent years, the security problems faced by deep learning models have become increasingly prominent, with various new types of attacks emerging, including poisoning attacks.

[0003] Deep learning models are vulnerable to poisoning attacks. In a poisoning attack, attackers inject specific perturbations into the model's training data, attempting to implant a backdoor into the deep learning model and thereby tamper with its internal weight parameters to make malicious predictions about specific inputs. Deploying these poisoned models on devices and providing them to users poses a significant security risk. Because the deployed model is a complete black box to the user, who can only query and obtain the model's predicted probabilities for a given input, existing poisoning defense methods typically require prior knowledge such as the poisoned training data and the target model's parameters, which is impractical in real-world applications. Therefore, designing a method suitable for black-box scenarios and capable of effectively detecting poisoning in wireless signals is urgent and necessary.

[0004] In wireless signal modulation classification tasks, Davaslioglu et al. used minute phase shifts in wireless modulation signals as poisoning triggers to achieve effective poisoning attacks. This attack method cleverly utilizes the phase characteristics of wireless signals without significantly altering the signal's appearance. Furthermore, Li et al. proposed a more covert poisoning attack strategy for automatic modulation recognition. They poisoned the in-phase and quadrature (I / Q) data of signals at random locations with low-power Gaussian noise, which slightly altered the amplitude of some sample points. This method is a more covert poisoning attack implementation, making it more difficult to detect and defend against. When the dataset and model parameters cannot be modified after model deployment, it is necessary to detect the input during the model's inference phase to reject malicious inputs. Gao et al. proposed a detection method based on strong perturbation addition. By adding various perturbations to the input samples and calculating the information entropy of these perturbation inputs after model prediction, they determine whether the input sample is a poisoned sample. In addition, Chou et al. used Gradient Weighted Class Activation Mapping (Grad-CAM) with an anomaly detection algorithm to identify triggers contained in samples, accurately identifying poisoned samples. Li et al. utilized the poor mobility of poisoned samples to obtain feature thresholds for the samples through a feature extractor, and then used these thresholds to detect poisoning signals.

[0005] Despite the achievements of the aforementioned methods, several challenges remain, particularly the following technical difficulties: 1) Insufficient detection sensitivity for covert attacks: Existing defense methods (such as STRIP and Grad-CAM) rely on response analysis to input perturbations or detection of local feature anomalies. However, attacks by Davaslioglu and Li implant triggers through small phase shifts or low-power noise, and their signal distortion may be lower than the detection threshold of the defense model. For example, I / Q data contaminated by Gaussian noise exhibits slight amplitude variations, making it difficult for detection methods based on information entropy or gradient activation to distinguish between normal signals and poisoned samples, especially in low signal-to-noise ratio environments. 2) Adaptability deficiencies of dynamic attack strategies: Existing defense mechanisms (such as non-migratory detection) assume that poisoned samples have fixed feature patterns (such as specific trigger locations or migratory differences). However, attackers can design dynamic triggers (such as random location noise or time-varying phase perturbations), rendering the static thresholds or feature extractors relied upon by defenses ineffective. For example, Li's attack randomly contaminates I / Q sampling points, causing methods based on spatial consistency (such as Grad-CAM) or migratory criteria to miss detections due to a lack of stable patterns. Summary of the Invention

[0006] Based on the above analysis, the embodiments of the present invention aim to provide a poisoning signal detection method and electronic device based on signal variational mode decomposition and reconstruction, so as to solve the problems that existing poisoning signal detection methods cannot achieve poisoning detection in black box scenarios, have insufficient detection sensitivity for covert attacks, and have low adaptability to dynamic attack strategies.

[0007] On one hand, embodiments of the present invention provide a method for detecting poisoned signals based on signal variational mode decomposition and reconstruction. The method includes: inputting a wireless signal to be detected into a signal classification model to obtain a first predicted probability distribution of the wireless signal; determining the signal category of the wireless signal based on the first predicted probability distribution; performing variational mode decomposition on a first clean reference signal of the same category according to a preset total number of modes, to obtain high-frequency intrinsic mode components and low-frequency intrinsic mode components of the first clean reference signal; performing variational mode decomposition on the wireless signal according to the total number of modes, to obtain high-frequency intrinsic mode components and low-frequency intrinsic mode components of the wireless signal; reconstructing the signal based on the low-frequency intrinsic mode components of the wireless signal and the high-frequency intrinsic mode components of the first clean reference signal, to obtain a reconstructed wireless signal; inputting the reconstructed wireless signal into the signal classification model to obtain a second predicted probability distribution of the reconstructed wireless signal; and determining whether the wireless signal is poisoned based on the difference between the first predicted probability distribution and the second predicted probability distribution.

[0008] A further improvement to the above method, the step of determining whether the wireless signal is poisoned based on the difference between the first predicted probability distribution and the second predicted probability distribution, includes: determining the similarity between the first predicted probability distribution and the second predicted probability distribution; determining the wireless signal as a poisoned signal in response to the similarity being greater than a similarity threshold corresponding to the signal category; and determining the wireless signal as a clean signal in response to the similarity being less than or equal to the similarity threshold.

[0009] Based on a further improvement of the above method, determining the similarity between the first predicted probability distribution and the second predicted probability distribution includes: determining the similarity between the first predicted probability distribution and the second predicted probability distribution by calculating the Euclidean distance between the first predicted probability distribution and the second predicted probability distribution.

[0010] A further improvement to the above method includes: acquiring a first number of second clean reference signals corresponding to the signal category; inputting the second clean reference signals into the signal classification model to obtain a third predicted probability distribution of the second clean reference signals; performing variational mode decomposition on the second clean reference signals according to the total number of modes to obtain high-frequency intrinsic mode components and low-frequency intrinsic mode components of the second clean reference signals; for each second clean reference signal, reconstructing the signal based on the low-frequency intrinsic mode components of the second clean reference signal and the high-frequency intrinsic mode components of another second clean reference signal to obtain a reconstructed second clean reference signal; inputting the reconstructed second clean reference signal into the signal classification model to obtain a fourth predicted probability distribution of the reconstructed second clean reference signal; and obtaining the similarity threshold based on the third probability distribution and the fourth probability distribution of the second clean reference signals before and after reconstruction.

[0011] A further improvement to the above method, wherein obtaining the similarity threshold based on the third and fourth probability distributions of the second clean reference signal before and after reconstruction, includes: calculating the Euclidean distance between the third and fourth probability distributions of the second clean reference signal before and after reconstruction; determining the standard normal distribution of the first number of Euclidean distances; determining the percentage function based on the inverse function of the standard normal distribution; and determining the similarity threshold by calculating the function value of the percentage function when the percentile is FRR; wherein FRR represents the probability that a clean signal is judged as a poisoned signal.

[0012] Based on a further improvement of the above method, the method further includes: obtaining a second number of third clean reference signals; using the third clean reference signals as input to a neural network model, and setting output labels according to the signal categories of the third clean reference signals, performing supervised training on the neural network model to obtain the signal classification model.

[0013] Based on a further improvement of the above method, the variational mode decomposition includes: performing variational mode decomposition on the I-channel signal and the Q-channel signal respectively to obtain the intrinsic mode components of the I-channel of the total number of modes and the intrinsic mode components of the Q-channel of the total number of modes.

[0014] Based on a further improvement of the above method, the signal reconstruction includes: for the I-channel signal, accumulating the low-frequency intrinsic mode components and high-frequency intrinsic mode components on which the signal reconstruction depends to obtain the reconstructed I-channel signal; for the Q-channel signal, accumulating the low-frequency intrinsic mode components and high-frequency intrinsic mode components on which the signal reconstruction depends to obtain the reconstructed Q-channel signal; and obtaining the reconstructed signal based on the reconstructed I-channel signal and the reconstructed Q-channel signal.

[0015] Based on a further improvement of the above method, the high-frequency intrinsic mode components are a predetermined number of intrinsic mode components with higher center frequencies among the total number of intrinsic mode components, and the other intrinsic mode components are low-frequency intrinsic mode components.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the poisoning signal detection method based on signal variational mode decomposition and reconstruction as described above.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the poisoning signal detection method based on signal variational mode decomposition and reconstruction as described above.

[0018] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the poisoning signal detection method based on signal variational mode decomposition and reconstruction as described above.

[0019] The poisoning signal detection method and electronic device based on signal variational mode decomposition and reconstruction provided by this invention effectively separates and destroys potential poisoning trigger features through the multimodal decomposition and reconstruction mechanism of the signal, while retaining the core features of the signal. Poisoning signal detection only requires the prediction probability distribution in the model inference stage, without accessing training data or modifying model parameters, thus achieving efficient defense in black-box scenarios. By replacing high-frequency modes to effectively destroy triggers, the sensitivity to covert attacks is improved. The poisoning detection process does not require the poisoned sample to be a fixed feature pattern, improving the adaptability of poisoning detection to dynamic attacks and providing a reliable poisoning sample detection scheme for wireless communication security.

[0020] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0021] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0022] Figure 1 This is a flowchart illustrating the poisoning signal detection method based on signal variational mode decomposition and reconstruction provided by the present invention.

[0023] Figure 2 This is a schematic diagram of the poisoning signal detection device based on signal variational mode decomposition and reconstruction provided by the present invention.

[0024] Figure 3 A schematic diagram of the physical structure of an electronic device is provided. Detailed Implementation

[0025] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0026] This invention uses a wireless signal modulation classification task as an example to illustrate the implementation process of the poisoned signal detection method based on signal variational mode decomposition and reconstruction. However, it should be noted that the poisoned signal detection method based on signal variational mode decomposition and reconstruction provided by this invention is applicable to poisoned signal detection in various signal classification scenarios.

[0027] Figure 1 This is a flowchart illustrating the poisoning signal detection method based on signal variational mode decomposition and reconstruction provided by the present invention. Figure 1 As shown, the method includes:

[0028] Step S1: Input the wireless signal to be detected into the signal classification model to obtain the first predicted probability distribution of the wireless signal, and determine the signal category of the wireless signal based on the first predicted probability distribution.

[0029] A signal classification model is a pre-trained model designed to classify signals into a predefined set of categories. For example, for modulated signals, there are 11 modulation categories: 8BPSK, QPSK, 8PSK, 16QAM, 64QAM, BFSK, CPFSK, and PAM4, as well as three analog modulation categories: WB-FM, AM-SSB, and AM-DSB. The classification set can be set to these 11 modulation categories to meet the needs of signal modulation category classification.

[0030] The wireless signal to be detected is input into a signal classification model to obtain a first predicted probability distribution of the wireless signal. This first predicted probability distribution can include the probability values ​​of the wireless signal being classified into different categories within a set of categories. The signal category of the wireless signal is then determined based on this first predicted probability distribution; for example, the category with the highest probability value in the first predicted probability distribution can be taken as the signal category of the wireless signal.

[0031] Step S2: According to the signal category, perform variational mode decomposition on the first clean reference signal of the same category based on the preset total number of modes to obtain the high-frequency intrinsic mode components and low-frequency intrinsic mode components of the first clean reference signal.

[0032] Based on the identified signal category of the wireless signal, Variational Mode Decomposition (VMD) is performed on the known first clean reference signal of the same category. The clean reference signal represents a signal that has been determined not to have been poisoned. The VMD is performed according to a preset total number of modes; that is, the number of Intrinsic Mode Functions (IMFs) obtained from the signal decomposition is equal to the total number of modes. The VMD of the first clean reference signal yields IMFs of different frequencies, including high-frequency IMFs and low-frequency IMFs. The center frequency of the high-frequency IMFs is higher than that of the low-frequency IMFs. The division between high-frequency and low-frequency IMFs can be determined according to preset rules, such as the proportion of high-frequency IMFs in the total number of modes, the number of high-frequency IMFs, or a preset center frequency threshold. Specifically, when determining the high-frequency intrinsic mode components and low-frequency intrinsic mode components based on a preset center frequency threshold, intrinsic mode components that are greater than or equal to the center frequency threshold are taken as high-frequency intrinsic mode components, and intrinsic mode components that are less than the center frequency threshold are taken as low-frequency intrinsic mode components.

[0033] Variational mode decomposition (VMD) achieves high-precision signal decomposition by constructing a constrained optimization model with a sparse frequency domain representation. Its core lies in iteratively solving for physically meaningful intrinsic mode components (EMCs). This method, after pre-setting the total number of modes K, aims to minimize the overall bandwidth of all EMCs. It uses an alternating direction multiplier method to dynamically adjust the center frequency and bandwidth parameters of each mode. The bandwidth calculation is based on an adaptive constraint using a quadratic penalty factor, ensuring strict frequency band division from low to high frequencies. The decomposition process applies linear constraints to completely reconstruct the original signal using the modal components. By introducing Lagrange multipliers to construct an augmented optimization function, it achieves accurate extraction of modal components with a pre-set convergence threshold while ensuring signal energy conservation.

[0034] Variational mode decomposition (VMD) suppresses mode aliasing through an adaptive non-recursive mechanism, ensuring accurate separation of high-frequency and low-frequency components. After decomposition, the low-frequency modes (short for low-frequency intrinsic mode components) of the signal mainly carry the core features, while the high-frequency modes (short for high-frequency intrinsic mode components) contain details or noise components.

[0035] Step S3: Based on the total number of modes, perform variational mode decomposition on the wireless signal to obtain the high-frequency intrinsic mode components and low-frequency intrinsic mode components of the wireless signal.

[0036] Based on the variational mode decomposition method consistent with the first clean reference signal, the wireless signal is subjected to variational mode decomposition according to the total number of modes to obtain the high-frequency intrinsic mode components and low-frequency intrinsic mode components of the wireless signal.

[0037] Step S4: Reconstruct the signal based on the low-frequency intrinsic mode components of the wireless signal and the high-frequency intrinsic mode components of the first clean reference signal to obtain the reconstructed wireless signal.

[0038] Since poisoning attacks typically rely on specific triggers (such as phase shifts or noise), replacing high-frequency components will compromise the integrity of the triggers, but will have a relatively small impact on the main characteristics of the clean signal. Based on the differences between the sub-modes after decomposition of the poisoned and clean signals, and the characteristic that the low-frequency components of a signal generally contain its main features, while the high-frequency components mainly contain its details, VMD can be used to take the lower-frequency modes of the intrinsic mode components of the wireless signal under test as the main modes characterizing the signal. Then, the high-frequency intrinsic mode components of the wireless signal are replaced with the high-frequency intrinsic mode components of the first clean reference signal. The signal is then reconstructed based on the low-frequency intrinsic mode components of the wireless signal and the high-frequency intrinsic mode components of the first clean reference signal to obtain the reconstructed wireless signal.

[0039] Step S5: Input the reconstructed wireless signal into the signal classification model to obtain the second predicted probability distribution of the reconstructed wireless signal, and determine whether the wireless signal is infected based on the difference between the first predicted probability distribution and the second predicted probability distribution.

[0040] For a clean signal to be tested, the main features of the signal remain unchanged after reconstruction, and the prediction result of the input model will not change significantly. However, if the signal to be tested is a poisoned signal with triggers, its recognition pattern depends on the triggers in the signal. After reconstruction, the triggers in the signal to be tested will no longer be complete, and the model will be unable to effectively identify the trigger features in the signal. The model's prediction result for this reconstructed signal will change significantly. This deviation can be used to calculate the difference between the probability distributions of the two signals to detect whether the signal to be tested is poisoned.

[0041] Clean signals, because their core features remain intact, exhibit high prediction similarity before and after reconstruction; that is, the difference between the first predicted probability distribution of the wireless signal before reconstruction and the second predicted probability distribution after reconstruction is small. In contrast, poisoned signals, due to the destruction of their triggers, experience significant shifts in prediction results, leading to decreased similarity; that is, the difference between the first predicted probability distribution of the wireless signal before reconstruction and the second predicted probability distribution after reconstruction is large. Therefore, the difference between the first predicted probability distribution of the wireless signal before reconstruction and the second predicted probability distribution after reconstruction can determine whether a wireless signal is poisoned. This method relies solely on the prediction probabilities during the model inference phase, without requiring modification of model parameters or access to training data, making it suitable for wireless signal security defense in black-box scenarios.

[0042] The poisoning signal detection method based on signal variational mode decomposition and reconstruction provided by this invention effectively separates and destroys potential poisoning trigger features through the multimodal decomposition and reconstruction mechanism of the signal, while retaining the core features of the signal. It only needs to use the predicted probability distribution in the model inference stage to detect poisoning signals, without accessing training data or modifying model parameters, thus achieving efficient defense in black-box scenarios. By replacing high-frequency modes to effectively destroy triggers, it improves the sensitivity to covert attack detection. The poisoning detection process does not require the poisoned sample to be a fixed feature pattern, which improves the adaptability of poisoning detection to dynamic attacks and provides a reliable poisoning sample detection scheme for wireless communication security.

[0043] According to the present invention, a method for detecting poisoned signals based on signal variational mode decomposition and reconstruction is provided. The step of determining whether a wireless signal is poisoned based on the difference between a first predicted probability distribution and a second predicted probability distribution includes: determining the similarity between the first predicted probability distribution and the second predicted probability distribution; determining the wireless signal as a poisoned signal if the similarity is greater than a similarity threshold corresponding to the signal category; and determining the wireless signal as a clean signal if the similarity is less than or equal to the similarity threshold.

[0044] Given that the distribution of predicted probability similarity differs before and after reconstruction for clean signals of different categories, this invention does not use a uniform similarity threshold for detection. Instead, it calculates a similarity threshold for each category of signal through fine-grained segmentation. In actual detection, the corresponding similarity threshold is adopted according to the classification result of the signal to be tested, thereby effectively improving the effectiveness and accuracy of the detection method.

[0045] When determining whether a wireless signal is poisoned based on the difference between the first and second predicted probability distributions, the similarity between the first and second predicted probability distributions is determined. This similarity is then compared with a similarity threshold corresponding to the signal category of the wireless signal. If the similarity is greater than the similarity threshold corresponding to the signal category, the wireless signal is determined to be a poisoned signal. If the similarity is less than or equal to the similarity threshold, the wireless signal is determined to be a clean signal.

[0046] The poisoning signal detection method based on signal variational mode decomposition and reconstruction provided by this invention determines whether a wireless signal is poisoned by comparing the similarity between the first and second predicted probability distributions with the similarity threshold corresponding to the category of the wireless signal, thereby improving the accuracy of poisoning detection.

[0047] According to the present invention, a method for detecting poisoning signals based on signal variational mode decomposition and reconstruction is provided. The step of determining the similarity between the first predicted probability distribution and the second predicted probability distribution includes: determining the similarity between the first predicted probability distribution and the second predicted probability distribution by calculating the Euclidean distance between the first predicted probability distribution and the second predicted probability distribution.

[0048] To accurately measure the similarity between the first and second predicted probability distributions, the difference between them is quantified by calculating the Euclidean distance between the two probability distributions.

[0049] The Euclidean distance, also known as the L2 norm, is used to measure the difference between two probability distributions. The larger the result, the lower the similarity between the two distributions. The specific formula is as follows:

[0050] The Euclidean distance between the first and second predicted probability distributions is expressed as follows:

[0051]

[0052] Where, dis represents the Euclidean distance between the first and second predicted probability distributions, N represents the number of signal classification categories, and y i This represents the probability value of predicting the unreconstructed wireless signal as category i. This represents the probability value of predicting the reconstructed wireless signal as category i.

[0053] The poisoning signal detection method based on signal variational mode decomposition and reconstruction provided by this invention quantifies the difference between the first and second predicted probability distributions by calculating the Euclidean distance between the two probability distributions, thereby improving the accuracy and efficiency of obtaining the difference between the first and second predicted probability distributions.

[0054] According to the present invention, a method for detecting poisoned signals based on signal variational mode decomposition and reconstruction is provided. The method further includes: acquiring a first number of second clean reference signals corresponding to the signal category; inputting the second clean reference signals into the signal classification model to obtain a third predicted probability distribution of the second clean reference signals; performing variational mode decomposition on the second clean reference signals according to the total number of modes to obtain high-frequency intrinsic mode components and low-frequency intrinsic mode components of the second clean reference signals; for each second clean reference signal, reconstructing the signal based on the low-frequency intrinsic mode components of the second clean reference signal and the high-frequency intrinsic mode components of another second clean reference signal to obtain a reconstructed second clean reference signal; inputting the reconstructed second clean reference signal into the signal classification model to obtain a fourth predicted probability distribution of the reconstructed second clean reference signal; and obtaining the similarity threshold based on the third probability distribution and the fourth probability distribution of the second clean reference signals before and after reconstruction.

[0055] Obtain a first number of second clean reference signals corresponding to the signal category, and obtain the similarity threshold of the corresponding signal category by statistically analyzing the similarity of the predicted probability distributions before and after reconstruction of the first number of second clean reference signals.

[0056] The second clean reference signal is input into the signal classification model to obtain the third predicted probability distribution of the second clean reference signal. Variational mode decomposition is then performed on the second clean reference signal based on the total number of modes to obtain the high-frequency intrinsic mode components and low-frequency intrinsic mode components of the second clean reference signal. For each second clean reference signal, signal reconstruction is performed based on the low-frequency intrinsic mode components of the second clean reference signal and the high-frequency intrinsic mode components of another second clean reference signal to obtain the reconstructed second clean reference signal. Since all second reference signals are clean signals, the reconstructed second clean reference signal is also a clean signal. The reconstructed second clean reference signal is input into the signal classification model to obtain the fourth predicted probability distribution of the reconstructed second clean reference signal. A similarity threshold is obtained based on the third and fourth probability distributions of each second clean reference signal before and after reconstruction.

[0057] For example, select n signals for each class from a clean reference signal. i represents the i-th signal category, 0 ≤ i ≤ N-1, and N is the total number of categories. For all clean reference signals Replace the modes with them and reconstruct them as follows Calculate the similarity between the model prediction probability distributions before and after reconstruction of all reference signals, and set an appropriate similarity threshold.

[0058] The poisoning signal detection method based on signal variational mode decomposition and reconstruction provided by this invention improves the accuracy of the similarity threshold by obtaining the similarity threshold according to the third probability distribution and the fourth probability distribution of the first number of second clean reference signals.

[0059] According to the present invention, a method for detecting poisoned signals based on signal variational mode decomposition reconstruction is provided. The step of obtaining the similarity threshold based on the third and fourth probability distributions of the second clean reference signal before and after reconstruction includes: calculating the Euclidean distance between the third and fourth probability distributions of the second clean reference signal before and after reconstruction; determining the standard normal distribution of the first number of Euclidean distances; determining the percentage function based on the inverse function of the standard normal distribution; and determining the similarity threshold by calculating the function value of the percentage function when the percentile is FRR; wherein FRR represents the probability that a clean signal is judged as a poisoned signal.

[0060] To calculate the similarity threshold, the Euclidean distance of the second clean reference signal for each category is first calculated to form an empirical distribution. Experience distribution The first Euclidean distance is the standard normal distribution, which can be called the cumulative distribution function.

[0061] For a random variable X C X C The Euclidean distance between the third and fourth probability distributions of the second clean reference signal can be defined by its cumulative distribution function as:

[0062]

[0063] in, This represents the pass rate of clean samples being judged as normal when the similarity threshold is 'a' in the second clean reference signal of category C. Then, the percentile of its normal distribution is calculated using the inverse function of the cumulative distribution function, which is the percentile function.

[0064]

[0065] Here, μ is the specified probability distribution, with the same value as FRR. FRR represents the probability that a clean signal is classified as a poisoned signal, also known as the false rejection rate. Considering the usability and effectiveness of the detection method, FRR can be set to 1% when obtaining the similarity threshold, reflecting the detection effect with a low error value. PPF(.) is the percentage point function, corresponding to the similarity threshold at the specified probability distribution.

[0066] Because of the standard normal distribution It depends on the signal category and the sample, so the entire similarity threshold setting process is dynamically adaptive.

[0067] The poisoning signal detection method based on signal variational mode decomposition and reconstruction provided by this invention determines the standard normal distribution of a first number of Euclidean distances and dynamically determines the similarity threshold using the false rejection rate and percentage point function. It can adaptively set detection standards for different signal categories, which significantly improves the accuracy and robustness of detection.

[0068] According to the present invention, a method for detecting poisoned signals based on signal variational mode decomposition and reconstruction is provided. The method further includes: acquiring a second number of third clean reference signals; using the third clean reference signals as input to a neural network model, and setting output labels according to the signal categories of the third clean reference signals, performing supervised training on the neural network model to obtain the signal classification model.

[0069] Obtain a second number of third clean reference signals. Use these third clean reference signals as input to the neural network model, and set output labels based on the signal categories of the third clean reference signals to perform supervised training of the neural network model. Specifically, when setting output labels, categories in the output vector that match the signal category of the third clean reference signals are set to 1, and other categories are set to 0. The network parameters can be optimized by minimizing the cross-entropy loss function, ultimately obtaining the signal classification model used in the deployment phase. This signal classification model is treated as a black box model in subsequent poisoning detection processes; only its predicted probability output is accessible, without access to model parameters or internal structure.

[0070] It should be noted that the terms "first," "second," and "third" in the terms "first clean reference signal," "second clean reference signal," and "third clean reference signal" are used to distinguish clean reference signals for different application purposes, and do not necessarily mean that they are different clean reference signals.

[0071] The poisoning signal detection method based on signal variational mode decomposition and reconstruction provided by this invention improves the accuracy of the signal classification model by using a third clean reference signal for supervised training to obtain the signal classification model.

[0072] According to the present invention, a method for detecting poisoning signals based on variational mode decomposition and reconstruction is provided. The variational mode decomposition includes performing variational mode decomposition on the I-channel signal and the Q-channel signal respectively to obtain the intrinsic mode components of the I-channel of the total number of modes and the intrinsic mode components of the Q-channel of the total number of modes.

[0073] The wireless signal to be detected is IQ wireless signal data, which consists of I-channel and Q-channel. During VMD decomposition, the I-channel and Q-channel can be decomposed separately to obtain the intrinsic mode components of the I-channel and the intrinsic mode components of the Q-channel, which represent the total number of modes.

[0074] The poisoning signal detection method based on signal variational mode decomposition and reconstruction provided by this invention performs variational mode decomposition by channel, which is beneficial to extract the features of different channels more accurately and improves the accuracy of signal detection.

[0075] According to the present invention, a method for detecting poisoning signals based on signal variational mode decomposition and reconstruction is provided. The signal reconstruction includes: for the I-channel signal, accumulating the low-frequency intrinsic mode components and high-frequency intrinsic mode components on which the signal reconstruction depends to obtain the reconstructed I-channel signal; for the Q-channel signal, accumulating the low-frequency intrinsic mode components and high-frequency intrinsic mode components on which the signal reconstruction depends to obtain the reconstructed Q-channel signal; and obtaining the reconstructed signal based on the reconstructed I-channel signal and the reconstructed Q-channel signal.

[0076] The signal reconstruction process can be represented as follows:

[0077]

[0078] Where, x I and x Q These represent the I-channel data and Q-channel data of the wireless signal x to be detected, respectively. Indicates with x I I-channel data of a clean reference signal of the same type, Indicates with x I Q-channel data of a clean reference signal of the same type, x represents I The k-th intrinsic mode component, express The k-th intrinsic mode component, x represents Q The k-th intrinsic mode component, express The k-th intrinsic mode component, where K represents the total number of modes and Z represents the number of high-frequency intrinsic mode components that need to be replaced. This represents the reconstructed I-channel signal. This represents the reconstructed Q-channel signal. This indicates the reconstructed wireless signal.

[0079] The poisoning signal detection method based on signal variational mode decomposition and reconstruction provided by this invention obtains the reconstructed I-channel signal and Q-channel signal by accumulating the low-frequency intrinsic mode component and the replaced high-frequency intrinsic mode component, respectively, thereby obtaining the reconstructed signal and improving the accuracy of the reconstructed signal.

[0080] According to the present invention, a poisoning signal detection method based on signal variational mode decomposition and reconstruction is provided, wherein the high-frequency intrinsic mode components are a predetermined number of intrinsic mode components with higher center frequencies among the total number of intrinsic mode components, and the other intrinsic mode components are low-frequency intrinsic mode components.

[0081] The poisoning signal detection method based on signal variational mode decomposition and reconstruction provided by this invention achieves rapid differentiation between high-frequency intrinsic mode components and low-frequency intrinsic mode components by setting the high-frequency intrinsic mode components as a preset number of intrinsic mode components with higher center frequencies among the total number of intrinsic mode components.

[0082] This invention proposes a poisoned sample detection method based on signal variational mode decomposition and reconstruction. This method utilizes the predictive differences between the reconstructed models of poisoned and clean signals to detect poisoned samples. Specifically, this method decomposes the signal using VMD and reconstructs the signal by replacing the high-frequency modes of the decomposed signal. After reconstruction, the low-frequency dominant mode characteristics of the clean signal remain unchanged, and its model prediction probability does not change significantly. However, for poisoned signals, the poisoning model classifies them based on trigger features. Since trigger features are distributed in high-frequency components, when they are replaced by the high-frequency modes of similar clean signals, the trigger features are destroyed, and the model prediction probability changes significantly. This change is quantified as reconstruction classification similarity, and a similarity threshold is set. By comparing the similarity threshold, it is determined whether the signal to be tested is a poisoned signal.

[0083] The poisoning signal detection device based on signal variational mode decomposition and reconstruction provided by the present invention will be described below. The poisoning signal detection device based on signal variational mode decomposition and reconstruction described below can be referred to in correspondence with the poisoning signal detection method based on signal variational mode decomposition and reconstruction described above.

[0084] Figure 2 This is a schematic diagram of the poisoning signal detection device based on signal variational mode decomposition and reconstruction provided by the present invention. Figure 2As shown, the device includes a first predicted probability distribution and signal category determination module 10, a first variational mode decomposition module 20, a second variational mode decomposition module 30, a signal reconstruction module 40, and a signal detection module 50. Specifically: the first predicted probability distribution and signal category determination module 10 is used to: input the wireless signal to be detected into a signal classification model to obtain a first predicted probability distribution of the wireless signal, and determine the signal category of the wireless signal based on the first predicted probability distribution; the first variational mode decomposition module 20 is used to: perform variational mode decomposition on a first clean reference signal of the same category according to the signal category and a preset total number of modes to obtain the high-frequency intrinsic modes of the first clean reference signal. The signal consists of a high-frequency intrinsic mode component and a low-frequency intrinsic mode component; a second variational mode decomposition module 30 is used to: perform variational mode decomposition on the wireless signal according to the total number of modes to obtain the high-frequency intrinsic mode component and the low-frequency intrinsic mode component of the wireless signal; a signal reconstruction module 40 is used to: reconstruct the signal according to the low-frequency intrinsic mode component of the wireless signal and the high-frequency intrinsic mode component of the first clean reference signal to obtain the reconstructed wireless signal; a signal detection module 50 is used to: input the reconstructed wireless signal into the signal classification model to obtain the second predicted probability distribution of the reconstructed wireless signal, and determine whether the wireless signal is poisoned based on the difference between the first predicted probability distribution and the second predicted probability distribution.

[0085] The poisoning signal detection device based on signal variational mode decomposition and reconstruction provided by this invention effectively separates and destroys potential poisoning trigger features through the multimodal decomposition and reconstruction mechanism of the signal, while retaining the core features of the signal. It only needs to use the predicted probability distribution in the model inference stage to detect poisoning signals, without accessing training data or modifying model parameters, thus achieving efficient defense in black-box scenarios. By replacing high-frequency modes to effectively destroy triggers, it improves the sensitivity to covert attack detection. The poisoning detection process does not require the poisoned sample to be a fixed feature pattern, which improves the adaptability of poisoning detection to dynamic attacks and provides a reliable poisoning sample detection scheme for wireless communication security.

[0086] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3As shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other through the communication bus 340. The processor 310 can call logic instructions in the memory 330 to execute a poisoned signal detection method based on signal variational mode decomposition and reconstruction. The method includes: inputting the wireless signal to be detected into a signal classification model to obtain a first predicted probability distribution of the wireless signal; determining the signal category of the wireless signal based on the first predicted probability distribution; performing variational mode decomposition on a first clean reference signal of the same category according to a preset total number of modes, obtaining high-frequency intrinsic mode components and low-frequency intrinsic mode components of the first clean reference signal; performing variational mode decomposition on the wireless signal according to the total number of modes, obtaining high-frequency intrinsic mode components and low-frequency intrinsic mode components of the wireless signal; reconstructing the signal based on the low-frequency intrinsic mode components of the wireless signal and the high-frequency intrinsic mode components of the first clean reference signal, obtaining a reconstructed wireless signal; inputting the reconstructed wireless signal into the signal classification model to obtain a second predicted probability distribution of the reconstructed wireless signal; and determining whether the wireless signal is poisoned based on the difference between the first predicted probability distribution and the second predicted probability distribution.

[0087] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0088] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the poisoned signal detection method based on signal variational mode decomposition and reconstruction provided by the above methods. The method includes: inputting the wireless signal to be detected into a signal classification model to obtain a first predicted probability distribution of the wireless signal; determining the signal category of the wireless signal according to the first predicted probability distribution; and performing variational mode decomposition on a first clean reference signal of the same category according to a preset total number of modes based on the signal category. The wireless signal is decomposed to obtain the high-frequency intrinsic mode components and low-frequency intrinsic mode components of the first clean reference signal; variational mode decomposition is performed on the wireless signal according to the total number of modes to obtain the high-frequency intrinsic mode components and low-frequency intrinsic mode components of the wireless signal; signal reconstruction is performed based on the low-frequency intrinsic mode components of the wireless signal and the high-frequency intrinsic mode components of the first clean reference signal to obtain the reconstructed wireless signal; the reconstructed wireless signal is input into the signal classification model to obtain the second predicted probability distribution of the reconstructed wireless signal; and the difference between the first predicted probability distribution and the second predicted probability distribution is used to determine whether the wireless signal is poisoned.

[0089] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for detecting poisoned signals based on signal variational mode decomposition and reconstruction provided by the above methods. This method includes: inputting a wireless signal to be detected into a signal classification model to obtain a first predicted probability distribution of the wireless signal; determining the signal category of the wireless signal based on the first predicted probability distribution; performing variational mode decomposition on a first clean reference signal of the same category according to the signal category and a preset total number of modes to obtain high-frequency intrinsic mode components and low-frequency intrinsic mode components of the first clean reference signal; performing variational mode decomposition on the wireless signal according to the total number of modes to obtain high-frequency intrinsic mode components and low-frequency intrinsic mode components of the wireless signal; reconstructing the signal based on the low-frequency intrinsic mode components of the wireless signal and the high-frequency intrinsic mode components of the first clean reference signal to obtain a reconstructed wireless signal; inputting the reconstructed wireless signal into the signal classification model to obtain a second predicted probability distribution of the reconstructed wireless signal; and determining whether the wireless signal is poisoned based on the difference between the first predicted probability distribution and the second predicted probability distribution.

[0090] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0091] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0092] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0094] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting poisoning signals based on signal variational mode decomposition and reconstruction, characterized in that, include: The wireless signal to be detected is input into the signal classification model to obtain the first predicted probability distribution of the wireless signal, and the signal category of the wireless signal is determined based on the first predicted probability distribution. According to the signal category, variational mode decomposition is performed on the first clean reference signal of the same category based on the preset total number of modes to obtain the high-frequency intrinsic mode components and low-frequency intrinsic mode components of the first clean reference signal. Based on the total number of modes, variational mode decomposition is performed on the wireless signal to obtain the high-frequency intrinsic mode components and low-frequency intrinsic mode components of the wireless signal. The reconstructed wireless signal is obtained by reconstructing the signal based on the low-frequency intrinsic mode components of the wireless signal and the high-frequency intrinsic mode components of the first clean reference signal. The reconstructed wireless signal is input into the signal classification model to obtain the second predicted probability distribution of the reconstructed wireless signal. The difference between the first predicted probability distribution and the second predicted probability distribution is used to determine whether the wireless signal is infected.

2. The method for detecting poisoning signals based on signal variational mode decomposition and reconstruction according to claim 1, characterized in that, The step of determining whether the wireless signal is infected based on the difference between the first predicted probability distribution and the second predicted probability distribution includes: Determine the similarity between the first predicted probability distribution and the second predicted probability distribution; If the similarity is greater than a similarity threshold corresponding to the signal category, then the wireless signal is determined to be a poisoned signal. If the similarity is less than or equal to the similarity threshold, the wireless signal is determined to be a clean signal.

3. The poisoning signal detection method based on signal variational mode decomposition reconstruction according to claim 2, wherein determining the similarity between the first predicted probability distribution and the second predicted probability distribution includes: The similarity between the first and second predicted probability distributions is determined by calculating the Euclidean distance between them.

4. The method for detecting poisoning signals based on signal variational mode decomposition and reconstruction according to claim 3, characterized in that, The method further includes: Acquire a first number of second clean reference signals corresponding to the signal category; The second clean reference signal is input into the signal classification model to obtain the third predicted probability distribution of the second clean reference signal; The second clean reference signal is subjected to variational mode decomposition based on the total number of modes to obtain the high-frequency intrinsic mode components and low-frequency intrinsic mode components of the second clean reference signal; For each of the second clean reference signals, the signal is reconstructed based on the low-frequency intrinsic mode components of the second clean reference signal and the high-frequency intrinsic mode components of the other second clean reference signals to obtain the reconstructed second clean reference signal. The reconstructed second clean reference signal is input into the signal classification model to obtain the fourth predicted probability distribution of the reconstructed second clean reference signal; The similarity threshold is obtained based on the third and fourth probability distributions of the second clean reference signal before and after reconstruction.

5. The method for detecting poisoning signals based on signal variational mode decomposition and reconstruction according to claim 4, characterized in that, The step of obtaining the similarity threshold based on the third and fourth probability distributions of the second clean reference signal before and after reconstruction includes: Calculate the Euclidean distance between the third probability distribution and the fourth probability distribution of the second clean reference signal before and after reconstruction; Determine the standard normal distribution of the first number of Euclidean distances; The percentage point function is determined based on the inverse function of the standard normal distribution; The similarity threshold is determined by calculating the function value of the percentile function when the percentile is FRR; where FRR represents the probability that a clean signal is judged as a poisoned signal.

6. The method for detecting poisoning signals based on signal variational mode decomposition and reconstruction according to claim 1, characterized in that, The method further includes: Obtain the second number of third clean reference signals; The neural network model is trained in a supervised manner by using the third clean reference signal as input and setting output labels according to the signal category of the third clean reference signal to obtain the signal classification model.

7. The method for detecting poisoning signals based on signal variational mode decomposition and reconstruction according to claim 1, characterized in that, The variational mode decomposition includes: Variational mode decomposition is performed on the I-channel signal and the Q-channel signal respectively to obtain the intrinsic mode components of the I-channel and the intrinsic mode components of the Q-channel of the total number of modes.

8. The method for detecting poisoning signals based on signal variational mode decomposition and reconstruction according to claim 7, characterized in that, The signal reconstruction includes: For the I-channel signal, the low-frequency intrinsic mode components and high-frequency intrinsic mode components on which the signal reconstruction depends are accumulated to obtain the reconstructed I-channel signal; For the Q-channel signal, the low-frequency intrinsic mode components and high-frequency intrinsic mode components on which the signal reconstruction depends are accumulated to obtain the reconstructed Q-channel signal; The reconstructed signal is obtained from the reconstructed I-channel signal and the reconstructed Q-channel signal.

9. The method for detecting poisoning signals based on signal variational mode decomposition and reconstruction according to claim 1, characterized in that, The high-frequency intrinsic modal components are a predetermined number of intrinsic modal components with higher center frequencies among the total number of intrinsic modal components, and the other intrinsic modal components are low-frequency intrinsic modal components.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the poisoning signal detection method based on signal variational mode decomposition and reconstruction as described in any one of claims 1 to 9.

Images