Electromagnetic information leakage signal denoising enhancement method based on improved DnCNN

Abstract

The application discloses an electromagnetic information leakage signal denoising enhancement method based on an improved DnCNN, and comprises the following steps: constructing a data set; constructing an initial deep denoising residual network model; inputting the data set into the initial deep denoising residual network model and training to obtain a deep denoising residual network model; and using the deep denoising residual network model to obtain denoising information and realize display leakage electromagnetic signal denoising. The application uses the improved DnCNN to improve the network denoising performance by fusing the convolution attention mechanism prompt network feature mapping capability, uses LeakyReLU as the network activation function to enhance the recovery ability of the denoising model after training to the weak effective signal, the denoising ability is more significant than that of the prior art, and the details in the effective signal can be better maintained while denoising, and good generalization is shown for complex signal-to-noise ratio data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electromagnetic information security, and particularly relates to an electromagnetic information leakage signal denoising and enhancing method based on an improved DnCNN. BACKGROUND

[0002] Electronic devices such as computers will unintentionally leak electromagnetic signals containing information when working. After the leaked electromagnetic signals are intercepted, the information in the electromagnetic leakage signals is restored through the information reconstruction technology in the TEMPEST technology, which threatens information security. By analyzing the existing electromagnetic information leakage detection methods, it can be found that the existing detection methods all have some shortcomings. The most important one is that the signal-to-noise ratio of the electromagnetic leakage signal is low, and under the interference of environmental noise, the information leakage characteristics are easy to deform, shift or even be suppressed. In the actual electromagnetic environment, the leaked electromagnetic signals will not only be disturbed by the noise in the environment, but also be disturbed by the electromagnetic waves radiated by other electronic devices or themselves. When the leaked electromagnetic signals are disturbed by a large amount of noise, the useful information will be covered by the noise, and it is difficult to detect the electromagnetic leakage and restore the effective information through the traditional image reproduction means. The existing detection methods have limited adaptability, and still need to carry out detection work on the premise of clearly defining the leakage characteristics. Therefore, it is difficult to standardize the definition and search of information leakage characteristics, which becomes a key problem faced by electromagnetic information leakage detection. In order to better detect the risk and more effectively check the electromagnetic leakage in the relatively complex electromagnetic environment, it is very important to remove the noise disturbance in order to realize the restoration and reconstruction of the information in the electromagnetic leakage signal.

[0003] The leakage mode of electromagnetic signals includes radiation leakage and conduction leakage, and the video information on the computer display can occur conduction leakage through the video cable. However, when electromagnetic information leaks, the electromagnetic signals radiated contain a large amount of noise, which has randomness and is difficult to determine the noise type through prior knowledge, and the electromagnetic signals taken in the actual environment have different noise in different environments, and the signal-to-noise ratio of the signals collected in different environments changes greatly, which brings difficulties to the further analysis and extraction of the signals. It is an urgent problem to suppress the noise in the electromagnetic signals and improve the signal-to-noise ratio. SUMMARY

[0004] The purpose of the present application is to provide an electromagnetic information leakage signal denoising and enhancing method based on an improved DnCNN. The electromagnetic information leakage signal denoising and enhancing method based on the improved DnCNN uses a deep convolution residual neural network to train a data set, and through the use of residual skip connection, the potential clean image is implicitly removed through the operation in the hidden layer, so as to train the noise data to restore the clean data.

[0005] To achieve the above object, the application provides an electromagnetic information leakage signal denoising enhancement method based on an improved DnCNN, which specifically comprises the following steps:

[0006] constructing a data set;

[0007] constructing an initial deep-layer denoising residual network model;

[0008] inputting the data set into the initial deep-layer denoising residual network model and training the same to obtain a deep-layer denoising residual network model;

[0009] using the deep-layer denoising residual network model to obtain denoising information and realize display leakage electromagnetic signal denoising.

[0010] Optionally, the data set is constructed by:

[0011] acquiring electromagnetic signals based on double-channel synchronization;

[0012] acquiring a label data set and a feature data set based on the electromagnetic signals;

[0013] constructing the data set based on the label data set and the feature data set.

[0014] Optionally, the electromagnetic signals include pure signals and noisy signals.

[0015] Optionally, the initial deep-layer denoising residual network model is constructed by:

[0016] constructing a one-dimensional selective convolution kernel network;

[0017] inputting the noisy signals into the one-dimensional selective convolution kernel network to obtain feature mapping information;

[0018] extracting features from the feature mapping information to obtain low-dimensional feature information;

[0019] reconstructing based on the low-dimensional feature information to obtain denoised signals and construct the initial deep-layer denoising residual network model.

[0020] Optionally, the one-dimensional selective convolution kernel network is constructed by:

[0021] obtaining feature mapping data based on a one-time convolution operation and an activation operation;

[0022] performing grouped convolution operation on the feature mapping data to obtain a plurality of convolutional feature maps;

[0023] summing a plurality of the convolutional feature maps to obtain a convolutional layer;

[0024] A global average pooling operation and a full connection operation are performed on the convolution layer, and a plurality of weight channel feature layers are obtained by using a softmax function;

[0025] A one-dimensional selective convolution kernel is selected from the plurality of weight channel feature layers to obtain a feature map processed by the one-dimensional selective convolution kernel, and the one-dimensional selective convolution kernel network is constructed.

[0026] Optionally, in the deep denoising residual network model, the activation function adopts a leaky rectified linear unit.

[0027] Optionally, the deep denoising residual network model is used to obtain the denoising information, including:

[0028] The deep denoising residual network model is predicted to obtain the noise of the noisy signal.

[0029] The noise of the noisy signal is calculated to obtain the denoising information.

[0030] The present application has the following advantages:

[0031] In the deep convolution residual neural network, the local attention mechanism is used to better map the noisy electromagnetic signal, and the Leaky RELU is introduced as the activation function of the residual denoising network. For the low signal-to-noise ratio electromagnetic signal unintentionally leaked, the recovery ability of the weak effective signal is enhanced, the performance of the denoising is improved, and the details in the effective signal are better maintained. The improved DnCNN is used to improve the network denoising performance by fusing the convolution attention mechanism to prompt the network feature mapping ability, and the Leaky ReLU is used as the network activation function to enhance the recovery ability of the weak effective signal of the denoised model after training. The denoising ability is more significant than the existing method, and the details in the effective signal are better maintained while denoising. The data with complex signal-to-noise ratio shows good generalization. BRIEF DESCRIPTION OF DRAWINGS

[0032] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application and their description are used to explain the application without limiting the application. In the drawings:

[0033] Figure 1 The flow chart of the improved DnCNN-based electromagnetic information leakage signal denoising enhancement method of the embodiment of the present application;

[0034] Figure 2 The improved DnCNN network structure diagram proposed in the embodiment of the present application;

[0035] Figure 3This is a structural diagram of the SKConv1d proposed in Embodiment 1 of the present invention;

[0036] Figure 4 This is a comparison diagram of the performance of the SKConv1d structure proposed in Embodiment 1 of the present invention in the network;

[0037] Figure 5 The original measured data one-dimensional signal and two-dimensional reconstructed image proposed in Embodiment 2 of the present invention are shown, wherein (a) is the waveform diagram and reconstructed image of the measured noisy data, and (b) is the waveform diagram and reconstructed image of the measured clean data;

[0038] Figure 6 The images show a comparison of the results of the original measured data collected in Embodiment 2 of the present invention after being denoised by the algorithm. (a) shows the denoising result of the measured data from the DnCNN network, (b) shows the denoising result of the measured data from the DnCNNs network, and (c) shows the denoising result of the measured data from the S-DnCNN network. Detailed Implementation

[0039] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0040] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0041] Example 1

[0042] like Figure 1 As shown, this embodiment provides a method for denoising and enhancing electromagnetic information leakage signals based on an improved DnCNN, specifically including the following steps:

[0043] Step 1: Construct a training set by simultaneously acquiring clean electromagnetic signals and noisy electromagnetic signals through dual channels as the label dataset and feature dataset for training, respectively.

[0044] Step 2: Construct a deep denoising residual network. In the network, a one-dimensional SKConv1d convolutional attention module is constructed to improve the feature mapping capability. The activation function in the network is Leaky ReLU to improve the network's ability to recover weak signals. Finally, the residual skip connection method is used to implicitly remove the potential clean image through the operation in the hidden layer, thereby training noisy data to recover clean data.

[0045] like Figure 2As shown, the deep denoising residual network is constructed. First, the electromagnetic acquisition device is used to collect the noisy signal and the pure signal in double channels. The noisy signal is input into the network. First, the feature mapping data is obtained based on the convolution operation and the activation operation. After the one-dimensional selective convolution operation is performed on the feature mapping data, the feature mapping is performed through the activation function. Then, the features are further extracted through the repeated convolution, activation and normalization operations. Finally, the data is restored to the initial size through the convolution operation. Then, the original input noisy signal is subtracted from the network output to train the noise data. For a new noisy signal data, the noise obtained through the network feature extraction is subtracted from the data to obtain the denoised signal.

[0046] A one-dimensional SKConv1d convolution attention module is constructed in the network to improve the feature mapping capability. Figure 4 As shown, the noisy data is input into the denoising network constructed by the selective convolution and the network without the selective convolution to compare the performance of the two. As shown in Figure 4 Finally, the labeling results show that the denoising network after adding the one-dimensional SKConv1d has better performance and better feature mapping capability.

[0047] As shown in Figure 3 The SKconv1d construction specifically includes:

[0048] S1, decomposition: first, the feature mapping data is obtained through the convolution operation and the activation operation. Then, the one-dimensional convolution operation with the kernel size of 3, 5 and 7 is performed on the input feature mapping to obtain and wherein and are the feature maps obtained by performing the convolution operation with three different sizes.

[0049] S2, fusion: first, the is obtained by performing the element-wise sum operation on the

[0050]

[0051] After obtaining the convolution layer U, the global average pooling operation is used to obtain s c :

[0052]

[0053] wherein, s c is the statistical information of each channel obtained by using the global average pooling, F gp is the global average pooling operation, U c is the feature map of the channel c, H is the height of the feature map U c , and W is the width of the feature map U cThe width of the convolutional layer is i, j is the subscript of the two-dimensional vector.

[0054] The information carried by each channel of the statistical convolutional layer can be counted through the global average pooling layer, and then s c The feature output full connection layer obtains z, which reduces the feature dimension and improves the network efficiency, and z is:

[0055] z=F fc (s) = δ(β(W s )) (3)

[0056] where z is the reduced dimension data of s, F fc is a full connection operation, s is the feature output of the previous step, δ is a Leaky ReLU activation function, β is a BN layer, W s is the weight matrix of s, W s ∈R d×c , R d×c represents that the numbers in the set belong to real numbers, and the size is d x c, r is the reduction ratio, and L is the minimum value of d.

[0057] S3, selection: after the fusion operation, different spatial scale weight information can be adaptively obtained, then the softmax function is performed on z to obtain the α, β and γ weight channel feature layers, respectively:

[0058]

[0059] The Leaky ReLU activation function and all the activation functions in the entire network are Leaky ReLU, and the calculation formula of Leaky ReLU is as follows:

[0060] g(x) = max(0, x) + ξ x min(0, x) (5)

[0061] where g(x) is the Leaky ReLU calculation function, x is the value in the input activation function, and ξ represents a constant, which is 0.02 in the embodiment.

[0062] Step three, training the improved DnCNN denoising network: the double-channel data of step one is put into the denoising model of step two for training. The data passes through three parts in the network: the first part generates feature mapping, the second part extracts low-dimensional features, and the third part reconstructs output. Finally, the noise obtained by training is output through the residual output. When new data needs to be detected, the noise can be removed through the trained model to obtain denoised data.

[0063] Electromagnetic leakage signal denoising principle construction:

[0064] Noisy electromagnetic leakage signals can be considered as a linear superposition of effective signals and noise:

[0065] y = e + n (6)

[0066] Where y, e, and n represent noisy electromagnetic leakage signal, pure electromagnetic signal, and noise, respectively.

[0067] The residual learning strategy is adopted to fit the optimal mapping relationship between noisy data and noise using an improved DnCNN. The specific process is shown in formula (7):

[0068]

[0069] Where RES represents residual learning, θ represents network parameters, mainly including weights and biases, and n is the noise used to improve the predictions of DnCNN. To make n equal to... By approaching infinity, the MSE denoising loss function in formula (8) is used to optimize the network parameter θ.

[0070]

[0071] Where L(θ) is the denoising loss function, N is the data size of the network batch, and n i For the i-th noise slice, y i This represents a noisy feature patch in the training set; The Frobenious norm is used. Since the improved DnCNN is a supervised deep learning method, it requires a large amount of training data to optimize network parameters. This embodiment uses mini-batch gradient descent to train the network, with a batch size of 64, a learning rate of 0.0001, and 400 training epochs. In each iteration, only a portion of the noisy feature patches and the clean signal label patches are used to calculate the gradient, using local gradients instead of global gradients. This improves computational efficiency during training and reduces network complexity. Finally, the trained improved DnCNN is used to predict noise in noisy electromagnetic signal data, and the denoised record is obtained through subtraction.

[0072] This embodiment provides a method for denoising electromagnetic signals leaking from displays based on an improved DnCNN. For supervised deep learning methods, the authenticity and sufficiency of the training dataset directly determine the performance of the trained model. The training dataset for S-DNCNN (improved DnCNN network) in multiple signal-to-noise ratio simulation mainly includes a signal dataset and a noisy dataset. These two datasets are related to the effective signal recovery capability and noise suppression capability of the trained CNN model.

[0073] Since there is no training data set suitable for electromagnetic signal data under complex signal-to-noise ratio at present, it is difficult to realize the measurement of complex electromagnetic signals in different regions and different environments, so a multiple signal-to-noise ratio test data set based on signal-to-noise ratio matrix (SNRM) is constructed, and the steps are as follows:

[0074] Step one, electromagnetic signal sample set: the sample data of the experiment comes from a display with a resolution of 1024x768 and a refresh rate of 60Hz, and the electromagnetic radiation leakage signals of two images of cat and airplane are collected respectively, and the sampling rate is 250M S / s (i.e. 250M Samples / s, 250M sampling points per second). The display is a PHILIPS HWE9220F liquid crystal display, and 3 samples are collected for each of the 2 states.

[0075] Step two, since the data is large and the data volume is small under the sampling rate of 250MS / s, the network burden is too heavy, so it is down-sampled to 1MS / s (i.e. 1M Samples / s, 1M sampling points per second), and 15000 signal patches with a length of 2000 are cut from these pure records with a sliding length of 2000 respectively; these signal patches together constitute the signal data set of S-DnCNN.

[0076] Step three, in order to construct a multiple signal-to-noise ratio signal data set, a signal-to-noise ratio matrix (SNRM) is constructed before the data is put into the network for training, since the batchsize is set to 64 in the experiment designed in this embodiment, the size of the matrix is set to 4000*64, and the elements in the matrix are random floating-point numbers between-0.01 and 10, and the signal data set slices in the network are cut through Gaussian noise to construct the signal-to-noise ratio SNRM, wherein the size of a random row is the noisy feature data set.

[0077] Specifically, in one iteration process, 64 effective signal pieces {x1, x2, x3, …, x 64}, 64 noise pieces {n1, n2, n3, …, n 64}, a random row in SNRM as {r1, r2, r3, …, r 64} are randomly selected. 64} as the feature data set of S-DNCNN, wherein the signal-to-noise ratio of d i is r i , and {x1, x2, x3, …, x 64} as the label data set of S-DNCNN.

[0078] In order to verify the performance of Leaky ReLU in electromagnetic leakage signal data denoising, the network with ReLU activation function replaced by Leaky ReLU is called DnCNN.

[0079] Table 1 shows the SNR and RMSE of five denoising methods: DnCNN, DnCNNS, S-DNCNN, BRDNet, and wavelet denoising. The initial SNR and RMSE are -5.06 dB and 23.06 dB, respectively.

[0080] Table 1

[0081]

[0082] Example 2

[0083] This example uses a digital signal acquisition device to collect dual-channel electromagnetic leakage information for measured electromagnetic signals. Channel 0 uses calipers to acquire the measured noisy electromagnetic signal as a feature, while channel 1 uses a probe to acquire a clean electromagnetic signal label. A total of 20 frames of data from each channel were collected, totaling 5,000,000 data points. The experimental sample data came from a monitor with a resolution and refresh rate of 1024x768@60Hz, with a sampling rate of 250M S / s (i.e., 250MSamples / s, 250M sampling points per second). Due to the small sample size and large data volume, this implementation performed a slicing process with a sliding step size of 20,000, resulting in 5,000 feature and 5,000 label data points each. Except for the dataset, the network parameters and structure remained consistent between the experimental and simulation experiments.

[0084] Table 2 shows the SNR and RMSE of the five noise reduction methods (DnCNN, DnCNNS, S-DNCNN, BRDNet, and Wavelet) in the experimental data. Noise represents the initial state, and the SNR and RMSE of the initial state are -0.4dB and 5.3, respectively.

[0085] Table 2

[0086]

[0087] Experiments were conducted to verify the denoising results by calculating the signal-to-noise ratio (SNR) and RMSE. S-DnCNN outperformed other algorithms in both SNR and RMSE after denoising. The use of Leaky ReLU allowed the model to retain more negative components, thus recovering the effective signal more completely. The effectiveness of Leaky ReLU was further verified by the denoising effect of DnCNNS. Integrating the constructed one-dimensional selective convolution kernel (SKConv1d) into the feature extraction network formed by local paths further enhanced the model's performance. The final results show that S-DnCNN achieved the highest SNR and best noise suppression effect after denoising compared to other models. It also had the smallest root mean square error, preserving higher signal-to-noise ratio accuracy. The original noisy data is shown below. Figure 5 As shown in (a), the original measured purity data are as follows:Figure 5 (b) the effect comparison chart of DnCNN network, DnCNNs network and S-DnCNN network after algorithmic denoising is shown in FIG. 3. Figure 6 Figure 6 (a) is the denoising result of DnCNN network for measured data, Figure 6 (b) is the denoising result of DnCNNs network for measured data, Figure 6 (c) is the denoising result of S-DnCNN network for measured data.

[0088] The above merely provides the preferred specific embodiments of the present application, but the protection scope of the present application is not limited to this. Any changes or replacements within the technical range disclosed by the present application, which can be easily thought by those skilled in the art, should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.​

Claims

1. An electromagnetic information leakage signal denoising enhancement method based on improved DnCNN, characterized in that, Specifically comprising the following steps: constructing a data set; constructing an initial deep denoising residual network model, comprising: constructing a one-dimensional selective convolution kernel network; inputting a noisy signal into the one-dimensional selective convolution kernel network to obtain feature mapping information; performing feature extraction on the feature mapping information to obtain low-dimensional feature information; reconstructing based on the low-dimensional feature information to obtain a denoised signal, and constructing the initial deep denoising residual network model; inputting the data set into the initial deep denoising residual network model and training to obtain a deep denoising residual network model; using the deep denoising residual network model to obtain denoising information and realize denoising of a display leakage electromagnetic signal.

2. The improved DnCNN-based electromagnetic information leakage signal denoising enhancement method of claim 1, wherein, The data set is constructed by: acquiring an electromagnetic signal based on dual-channel synchronization; acquiring a label data set and a feature data set based on the electromagnetic signal; constructing the data set based on the label data set and the feature data set.

3. The improved DnCNN-based electromagnetic information leakage signal denoising enhancement method of claim 2, wherein, The electromagnetic signal includes a pure signal and a noisy signal.

4. The improved DnCNN-based electromagnetic information leakage signal denoising enhancement method of claim 3, wherein, The one-dimensional selective convolution kernel network is constructed by: obtaining feature mapping data based on a one-time convolution operation and an activation operation; performing grouped convolution operation on the feature mapping data to obtain a plurality of convolutional feature maps; summing a plurality of the convolutional feature maps to obtain a convolution layer; performing global average pooling operation and full connection operation on the convolution layer, and using a softmax function to obtain a plurality of weight channel feature layers; selecting a plurality of the weight channel feature layers to obtain a one-dimensional selective convolution kernel processed feature map, and constructing the one-dimensional selective convolution kernel network.

5. The improved DnCNN-based electromagnetic information leakage signal denoising enhancement method of claim 1, wherein, The activation function in the deep denoising residual network model adopts a leaky rectified linear unit.

6. The improved DnCNN-based electromagnetic information leakage signal denoising enhancement method of claim 3, wherein, Using the deep denoising residual network model to obtain the denoising information comprises: predicting the deep denoising residual network model to obtain noise of the noisy signal; calculating based on the noise of the noisy signal to obtain the denoising information.

Images