Swin-conv-unet-based radio frequency fingerprint identification method and system under low signal-to-noise ratio

Abstract

The application discloses a radio frequency fingerprint identification strategy suitable for a low signal-to-noise ratio environment for the purpose of protecting wireless network security. In an Internet of Things system, how to avoid attacks such as illegal monitoring, tampering and impersonation by intruders is a prerequisite for guaranteeing user security. Therefore, a radio frequency fingerprint identification technology without password authentication is proposed to guarantee the security of a wireless network. However, in a low signal-to-noise ratio environment, the existing radio frequency fingerprint identification technology may not have good performance, and therefore it is an urgent problem to be solved to improve the radio frequency fingerprint identification accuracy in a low signal-to-noise ratio environment. The application converts the denoising problem of a one-dimensional radio frequency signal into a denoising problem of a two-dimensional time-frequency graph, and denoises the Wigner-Ville time-frequency graph of the radio frequency signal through a Swin-Conv-UNet denoising network, so as to reduce the influence of noise on the radio frequency fingerprint and greatly improve the accuracy of device identification.

Description

Technical Field

[0001] This invention relates to the fields of mobile communication technology and wireless network security, and in particular to an identification strategy that ensures the accuracy of radio frequency fingerprint recognition under low signal-to-noise ratio conditions. Background Technology

[0002] With the rapid development of IoT technology, the open nature of its wireless networks has also brought numerous network security vulnerabilities. To ensure wireless network security, access authentication technology has been proposed. It primarily prevents unauthorized intruders by identifying and verifying user identities. However, traditional authentication technologies are susceptible to attacks such as unauthorized eavesdropping, tampering, and impersonation. Intruders can deceive devices by copying keys and identification information such as IP / MAC addresses. Therefore, non-password authentication radio frequency fingerprinting technology has been proposed to ensure wireless network security. However, in low signal-to-noise ratio (SNR) environments, existing radio frequency fingerprinting technologies may not perform well. Therefore, improving the accuracy of radio frequency fingerprinting in low SNR environments has become a key issue. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a radio frequency fingerprint recognition strategy that can ensure a high recognition accuracy in low signal-to-noise ratio environments.

[0004] To address the aforementioned technical problems, this invention combines Wigner-Ville distribution theory and deep learning theory, employing a segmentation convolutional network as its backbone and incorporating a shifted window multi-head self-attention mechanism and a residual convolutional network to create a denoising network (Swin-Conv-UNet denoising network), and a residual neural network (ResNet18 classification network). The process includes the following steps:

[0005] (1) The signals are transformed into Wigner-Ville distribution (WVD) time-frequency diagrams f1 by Wigner-Ville transform;

[0006] (2) Construct the Swin-Conv-UNet denoising network;

[0007] (3) Input the WVD time-frequency graphs f1 under different signal-to-noise ratios into the Swin-Conv-UNet denoising network and train it using deep learning theory;

[0008] (4) The weights obtained from training are used as the optimal weights W of the Swin-Conv-UNet. s ;

[0009] (5) The radio frequency signal to be identified of the known tag is converted into WVD time-frequency diagram f2 by Wigner-Ville transform;

[0010] (6) By loading model W S To denoise f2, we obtain the reconstructed image x;

[0011] (7) Divide the reconstructed images x under different signal-to-noise ratios into training set, validation set and test set according to the ratio (which can be 0.8:0.1:0.1);

[0012] (8) Build a ResNet18 network;

[0013] (9) Input the training set and validation set into the ResNet18 deep learning theory for training and validation, respectively;

[0014] (10) Use the saved weights as the optimal weights W for the ResNet18 network. R ;

[0015] (11) During the testing phase, test sets with different signal-to-noise ratios are sequentially input into the weighted set W. S The Swin-Conv-UNet denoising network and weights W R The ResNet18 classification network;

[0016] (12) Calculate the recognition accuracy under different signal-to-noise ratios based on the label and output.

[0017] Preferably, step (1) includes the following specific steps:

[0018] (11) Input the IQ complex signal s;

[0019] (12) Obtain the WVD time-frequency diagram according to the Wigner-Ville transform formula, i.e.

[0020]

[0021] Where t represents time, f represents frequency, τ represents time delay, and s * (t) is the conjugate of s(t), and e represents the natural base.

[0022] Preferably, step (2) includes the following specific steps:

[0023] (21) The feature map is obtained through a 3*3 convolutional layer;

[0024] (22) Three downsampling operations are achieved by using three convolutional layers with a stride of 2;

[0025] (23) Three upsampling operations are achieved by using three transposed convolutional layers with a stride of 2;

[0026] (24) Build the Swin-Conv(SC) module;

[0027] (25) Insert 4 SC modules before each downsampling;

[0028] (26) Insert 4 SC modules between the last downsampling and the first upsampling;

[0029] (27) Insert 4 SC modules after each upsampling;

[0030] (28) Connect the outputs of convolutional layers and SC modules of the same size using a shortcut;

[0031] (29) Finally, image reconstruction is achieved through a 3*3 convolutional layer.

[0032] Preferably, the specific steps of step (24) include:

[0033] (241) The feature map X is obtained through a 1*1 convolutional layer;

[0034] (242) Divide X into two parts, X1 and X2, according to the number of channels;

[0035] (243) X1 is input into the SwinT module to obtain Y1;

[0036] (244) X2 is input into the 3*3 RConv module to obtain Y2;

[0037] (244) Y1 and Y2 are combined to obtain Y;

[0038] (245) Y achieves feature fusion of the two modules through a 1*1 convolution to obtain C;

[0039] (246) Adding the feature map X to C yields the output Z, i.e.

[0040] Z = Conv 1×1 (concat(SwinT(X1),RConv(X2)))+X

[0041] Preferably, the specific steps of step (243) include:

[0042] (2431) In the l-th layer, X1 first passes through a normalization layer to obtain X. Nl ;

[0043] (2432)X Nl The feature map X is divided into M small feature maps by M non-overlapping windows. il ;

[0044] (2433) For each X il Calculate its multi-head self-attention and concatenate them to obtain the feature map X. Hl ;

[0045] (2434)XHl Adding X1 gives X Wl ,Right now:

[0046] X Wl =WMSA(Norm(X1))+X1

[0047] (2435)X Wl X is obtained through a normalization layer. WNl ;

[0048] (2436)X WNl X is obtained through an MLP layer. WMl ;

[0049] (2437)X WMl Add X Wl Get X l , represented as:

[0050] X l =MLP(Norm(X) Wl ))+X Wl

[0051] (2438) In the (l+1)th layer, X l X is first obtained through a normalization layer. N(l+1) ;

[0052] (2439)X N(l+1) The feature map X is divided into M small feature maps by M non-overlapping windows with shifts of [M / 2, M / 2]. i(l+1) ;

[0053] (24310) For each X i(l+1) Calculate its multi-head self-attention and concatenate them to obtain the feature map X. H(l+1) ;

[0054] (24311) Calculate and output Y1 according to steps (2434)-(2437), which is represented as

[0055] Y1 = MLP(Norm(SWMSA(Norm(X) l ))+X l ))+(SWMSA(Norm(X l ))+X l )

[0056] Preferably, the specific steps of step (244) include:

[0057] (2441)X2 is passed through a 3x3 convolutional layer to obtain X. R ;

[0058] (2442)X RAdding X2 gives the output Y2, which is...

[0059] Y2 = Conv 3×3 (X2)+X2

[0060] Preferably, step (3) includes the following specific steps:

[0061] (31) Adam is selected as the optimizer, the learning rate is set to 0.0001, the batch size is set to 4, and the L1 norm is used as the loss function, expressed as:

[0062]

[0063] Where n represents the number of samples, and for the i-th sample, x i This represents the corresponding clean image, y i This represents the denoised image output by the SCUNet network;

[0064] (32) Initialize minimum loss Loss1 min It is infinitely large;

[0065] (33) Calculate the loss Loss1 between the reconstructed image x and the clean image y output by the network;

[0066] (34) Feed Loss1 back to the Swin-Conv-UNet network;

[0067] (35) If Loss1 < Loss1 min Then update the minimum loss Loss1. min =Loss1 and save the model weights W S ;

[0068] (36) Repeat steps (33)-(35) until Loss1≥Loss1 is maintained for 5 consecutive times. min .

[0069] Preferably, step (9) includes the following specific steps:

[0070] (91) During the training phase, Adam was selected as the optimizer, the learning rate was set to 0.001, the batch size was set to 128, and the cross-entropy loss was used as the loss function, expressed as:

[0071]

[0072] Where N represents the total number of samples, K represents the total number of categories, and y ik Let y represent the label of the i-th sample. If the i-th sample belongs to category k, then y ik =1, otherwise 0; This represents the predicted probability that the i-th sample belongs to category k.

[0073] (92) Initialize minimum loss Loss2 min It is infinitely large;

[0074] (93) The training loss Loss2 is calculated using the training set output and labels. train ;

[0075] (94) Loss2 train Feedback is sent to the ResNet18 network;

[0076] (95) Calculate the validation set loss Loss2 val ;

[0077] (96) If Loss2 val <Loss2 min Then update the minimum loss Loss2. min =Loss2 val And save the model weights;

[0078] (97) Repeat steps (93)-(96) until Loss2 is maintained for 10 consecutive times. val ≥Loss2 min .

[0079] The present invention also provides an radio frequency fingerprint recognition system based on a Swin-Conv-UNet denoising network, comprising:

[0080] (1) Wigner-Ville transform module, used to transform one-dimensional radio frequency signals into WVD time-frequency diagrams;

[0081] (2) The Swin-Conv-UNet denoising network module is used to denoise the WVD time-frequency graph;

[0082] (3) The ResNet18 classification network module is used to identify the denoised WVD time-frequency graph.

[0083] The beneficial effects of this invention are as follows: First, this invention uses the Wigner-Ville distribution theory to transform one-dimensional IQ complex samples into two-dimensional WVD time-frequency maps. Then, combining the Swin-Conv-UNet denoising network and the classic ResNet18 classification network, it proposes an RF fingerprint recognition algorithm suitable for low signal-to-noise ratio (SNR) environments. This algorithm can guarantee high recognition accuracy even in high-noise environments. Numerical results show that the proposed algorithm maintains excellent recognition accuracy in low SNR environments, and compared with existing RF fingerprint recognition designs, the solution of this invention provides significant performance gains. Detailed Implementation

[0084] The invention will be validated using the large-scale WiFi dataset WiSig. In the simulation, Gaussian white noise is chosen to represent the impact of the AWGN channel. To control for the effects of channel variations, data packets measured on the same day are selected to construct the dataset. To avoid the influence of different receivers on the RF fingerprint characteristics, signals received from the same receiver are selected to construct the dataset.

[0085] The identification process is carried out in the following manner, including the following steps:

[0086] (1) First, consider the denoising problem of Swin-Conv-UNet in RF fingerprint recognition. Swin-Conv-UNet is used to denoise the WVD time-frequency image after Wigner-Ville transform, and the denoising network is trained and tested using deep learning theory. The denoising performance of Swin-Conv-UNet is measured using the L1 norm loss function. Through repeated iterations, the optimal weights of the denoising model are obtained until the L1 norm loss curve converges.

[0087] (2) Secondly, consider the recognition problem of radio frequency fingerprint recognition strategy under low signal-to-noise ratio. Radio frequency fingerprint features in low signal-to-noise ratio environment are easily submerged by noise, resulting in incorrect recognition results. ResNet18 is used to classify the WVD time-frequency map after denoising in step (1), and the classification network is trained and tested through deep learning theory to calculate the cross-entropy loss and accuracy. Iterate repeatedly until the cross-entropy loss curve converges to obtain the optimal weight of the classification model.

[0088] (3) Radio frequency fingerprint recognition problem under low signal-to-noise ratio based on Swin-Conv-UNet. Combining the optimal weights of the model obtained in steps (1) and (2), the output results of the two models in series are calculated to obtain the accuracy of radio frequency fingerprint recognition under different signal-to-noise ratios, and realize the radio frequency fingerprint recognition scheme under low signal-to-noise ratio environment.

[0089] In step (1), the denoising of the WVD time-frequency map using Swin-Conv-UNet includes the following steps:

[0090] (11) Construct the Swin-Conv-UNet denoising network;

[0091] (12) Select Adam as the optimizer, set the learning rate to 0.0001, the batch size to 4, and the L1 norm as the loss function. Initialize the minimum loss Loss1. min It is infinitely large;

[0092] (13) Calculate the network output loss Loss1;

[0093] (14) Feed Loss1 back to the Swin-Conv-UNet network;

[0094] (15) If Loss1 < Loss1 min Then update the minimum loss Loss1. min =Loss1 and save the model weights W S ;

[0095] (16) Repeat steps (13)-(15) until Loss1 ≥ Loss1 is maintained for 5 consecutive times. min .

[0096] Step (2), using ResNet18 for radio frequency fingerprint recognition includes the following steps:

[0097] (21) Build a ResNet18 network;

[0098] (22) During the training phase, Adam was selected as the optimizer, the learning rate was set to 0.001, the batch size was set to 128, the cross-entropy loss was used as the loss function, and the minimum loss Loss2 was initialized. min It is infinitely large;

[0099] (23) Calculate the training loss Loss2 using the training set output and labels. train ;

[0100] (24) Loss2 train Feedback is sent to the ResNet18 network;

[0101] (25) Calculate the validation set loss Loss2 val ;

[0102] (26) If Loss2 val <Loss2 min Then update the minimum loss Loss2. min =Loss2 val And save the model weights W R ;

[0103] (27) Repeat steps (23)-(26) until Loss2 is maintained for 10 consecutive times. val ≥Loss2 min .

[0104] Step (3) involves calculating the recognition accuracy of RF fingerprint recognition under low signal-to-noise ratio based on Swin-Conv-UNet, including the following steps:

[0105] (31) Input the test sets with different signal-to-noise ratios into the weight W in sequence. S The Swin-Conv-UNet denoising network and weights W RThe ResNet18 classification network;

[0106] (32) Calculate the recognition accuracy under different signal-to-noise ratios based on the label and output.

[0107] Based on the above method, this embodiment also provides an RF fingerprint recognition system based on a Swin-Conv-UNet denoising network, including:

[0108] (1) Wigner-Ville transform module, used to transform one-dimensional radio frequency signals into WVD time-frequency diagrams;

[0109] (2) The Swin-Conv-UNet denoising network module is used to denoise the WVD time-frequency graph;

[0110] (3) The ResNet18 classification network module is used to identify the denoised WVD time-frequency graph.

[0111] Although the present invention has been illustrated and described with reference to preferred embodiments, those skilled in the art should understand that various changes and modifications can be made to the present invention without departing from the scope defined by the claims.

Claims

1. A radio frequency fingerprint recognition method based on a Swin-Conv-UNet denoising network, characterized in that, Includes the following steps: (1) Convert the radio frequency signal into a WVD time-frequency diagram by Wigner-Ville transform. ; (2) Construct the Swin-Conv-UNet denoising network; (3) WVD time-frequency diagram The Swin-Conv-UNet denoising network is trained using deep learning theory to obtain its optimal weights. ; (4) Convert the radio frequency signal to be identified of a known tag into a WVD time-frequency diagram using the Wigner-Ville transform. ; (5) By loading the model Come to Denoising is performed to obtain the reconstructed image. ; Reconstruct the image It is divided into a training set, a validation set, and a test set; (6) Build a ResNet18 network; (7) Input the training set and validation set into ResNet18 respectively for training and validation using deep learning theory to obtain the optimal weights of the ResNet18 network. ; (8) During the testing phase, the test set is sequentially input with weights of... The Swin-Conv-UNet denoising network and weights are The ResNet18 classification network; (9) Calculate the recognition accuracy based on the labels of the test set and the output of the ResNet18 classification network; The specific steps of step (1) include: (11) Input IQ complex signal ; (12) The WVD time-frequency diagram is obtained according to the Wigner-Ville transform formula, i.e. ； in Indicates time, Indicates frequency, Indicates time delay, yes conjugate, Represents the natural base; The specific steps of step (2) include: (21) Through a The convolutional layers obtain feature maps; (22) Three downsampling operations are achieved using three convolutional layers with a stride of 2; (23) Three upsampling operations are achieved by using three transposed convolutional layers with a stride of 2; (24) Build the Swin-Conv module; (25) Insert 4 Swing-Conv modules before each downsampling; (26) Insert 4 Swing-Conv modules between the last downsampling and the first upsampling; (27) Insert 4 Swing-Conv modules after each upsampling; (28) Connect the outputs of convolutional layers and the Swin-Conv module of the same size using a shortcut; (29) Finally, through a Convolutional layers are used to reconstruct images.

2. The radio frequency fingerprint recognition method based on the Swin-Conv-UNet denoising network according to claim 1, characterized in that, The specific steps of step (24) include: (241) Through a The convolutional layer obtains the feature map ; (242) will Divided into two parts based on the number of channels. and ; (243) Inputting the SwinT module yields ; (244) enter The RConv module is obtained ; and splicing together ; (245) Through a The convolutional function achieves feature fusion between the two modules to obtain... ; (246) Add feature map Get output ,Right now ； express convolution, This indicates a splicing operation. This refers to the Swing Transformer module. This represents the residual convolution module.

3. The radio frequency fingerprint recognition method based on the Swin-Conv-UNet denoising network according to claim 2, characterized in that, The specific steps of step (243) include: (2431) in the layer, First, obtain through a normalization layer ; (2432) quilt Divide the non-overlapping windows into Small feature map ; (2433) For each Calculate its multi-head self-attention and concatenate them to obtain the feature map. ; (2434) Plus get ,Right now: ； This represents a window-based self-attention mechanism. Indicates the normalization layer; (2435) Obtained through a normalization layer ; (2436) Obtained through an MLP layer ; (2437) Plus get , is represented as: ； This represents a multilayer perceptron; (2438) in the layer, First, obtain through a normalization layer ; (2439) quilt A person with The non-overlapping window of the shift is divided into Small feature map ; (24310) for each Calculate its multi-head self-attention and concatenate them to obtain the feature map. ; (24311) Perform the calculation and output according to steps (2434)-(2437). , represented as ； in, This represents a self-attention mechanism based on a shifted window.

4. The radio frequency fingerprint recognition method based on the Swin-Conv-UNet denoising network according to claim 2, characterized in that, The specific steps of step (244) include: (2441) Through a The output of the convolutional layer is obtained ; (2442) Plus Get output ,Right now 。 5. The radio frequency fingerprint recognition method based on the Swin-Conv-UNet denoising network according to claim 1, characterized in that, The specific steps of step (3) include: (31) Adam is selected as the optimizer. The norm, as a loss function, is expressed as: ； in, Represents the sample number, for the th For a single sample This represents the corresponding clean image. This represents the denoised image output by the SCUNet network; (32) Initialize minimum loss It is infinitely large; (33) Calculate the reconstructed image output by the network and clean images Losses between ; (34) will Feedback is sent to the Swin-Conv-UNet network; (35) If Then update the minimum loss. And save the model weights ; (36) Repeat steps (33)-(35) until the result is maintained for a certain number of consecutive times. .

6. The radio frequency fingerprint recognition method based on the Swin-Conv-UNet denoising network according to claim 1, characterized in that, The specific steps of step (7) include: (71) During the training phase, Adam is selected as the optimizer, and cross-entropy loss is used as the loss function, expressed as: ； in, Represents the total number of samples. Indicates the total number of categories. Indicates the first The label of the nth sample, if the nth sample Each sample belongs to category ,but Conversely, it is 0; Indicates the first Each sample belongs to category The predicted probability; (72) Initialize minimum loss It is infinitely large; (73) The training loss is calculated using the training set output and labels. ; (74) will Feedback is sent to the ResNet18 network; (75) Calculate the validation set loss ; (76) If Then update the minimum loss. And save the model weights; (77) Repeat steps (73)-(76) until the result is maintained for a certain number of consecutive times. .

7. A radio frequency fingerprint recognition system based on a Swin-Conv-UNet denoising network according to the method of claim 1, characterized in that, include: (1) Wigner-Ville transform module, used to transform one-dimensional radio frequency signals into WVD time-frequency diagrams; (2) The Swin-Conv-UNet denoising network module is used to denoise the WVD time-frequency graph; (3) The ResNet18 classification network module is used to identify the denoised WVD time-frequency graph.

Images