A radio frequency fingerprinting method based on a lightweight deep learning model

Abstract

The application belongs to the technical field of Internet of Things security, and particularly relates to a radio frequency fingerprint identification method based on a light-weight deep learning model. The application first converts I / Q signals into a complex matrix, then performs singular value decomposition on the complex matrix Q, restores the obtained features into a complex vector after noise reduction, then constructs a light-weight deep learning model to perform complex feature extraction and classification, and finally realizes radio frequency fingerprint identification. The application has the beneficial effect of well solving the interference of environmental noise on radio frequency fingerprint identification.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things security technology, specifically relating to a radio frequency fingerprinting method based on a lightweight deep learning model. Background Technology

[0002] The rapid development of wireless communication technology has greatly facilitated people's lives. With the increasing prevalence of wireless communication, the number of wireless communication devices has inevitably exploded. Because wireless networks are spatially open, wireless communication devices can communicate as long as they are within data reception range. This, to some extent, increases the possibility of wireless communication being hijacked, tampered with, and attacked. Therefore, wireless communication urgently needs new security mechanisms and device authentication technologies to defend against potential threats from wireless devices in the network.

[0003] As early as 1995, Choe et al. proposed using communication signals for device identification. However, the concept of Radio Frequency Fingerprint (RFF) for wireless equipment was formally proposed in 2003. RRF fingerprints refer to the "fingerprint" features extracted from the radio frequency signals emitted by a signal transmitter. Just like human fingerprints, different transmitters have different RRF fingerprints. By utilizing subtle differences in the radio frequency signals, different transmitters can be identified and distinguished, thus achieving transmitter authentication at the physical layer. Because RRF fingerprint recognition operates at the physical layer, its difficulty in forgery allows for secure and reliable identity authentication, effectively improving its security performance.

[0004] Currently, research on radio frequency fingerprinting (RF fingerprinting) mainly falls into two categories: feature extraction-based RF fingerprinting and deep learning-based RF fingerprinting. The former involves acquiring RF signals, performing channel estimation and spectrum analysis, then using specific algorithms to extract features from the signals. These multi-dimensional features are then combined into feature vectors, which are then used for classification and recognition using traditional machine learning methods. The latter involves directly inputting labeled I / Q data into a model, performing feature extraction and classification during the model's forward propagation. Feature extraction-based RF fingerprinting methods are typically offline and require significant prior knowledge, communication domain expertise, and expert experience; otherwise, reliable feature extraction is difficult. Deep learning-based RF fingerprinting typically eliminates the need for offline analysis, directly using I / Q data as input for model training and recognition. It relies on fewer conditions and has higher versatility. However, deploying RF fingerprinting on edge devices faces challenges such as insufficient computing power, slow recognition speed, limited storage resources, and susceptibility to noise interference. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a radio frequency fingerprinting method based on a lightweight deep learning model, which alleviates to some extent the problems of insufficient computing power, slow recognition speed, limited storage resources, and susceptibility to noise interference faced when deploying radio frequency fingerprinting on edge devices.

[0006] The technical solution of this invention is as follows:

[0007] A radio frequency fingerprint recognition method based on a lightweight deep learning model, characterized by the following steps:

[0008] S1. Convert the acquired I / Q signals into complex matrices, specifically:

[0009] The acquired I / Q signals are written as complex vectors X = (x1, x2, x3, ..., x...). N ), where N is the number of samples, x i Let each element in the vector be i = 1, 2, 3, ..., N, and transform X into a complex matrix Q. n×n Where N = n × n, the transformation method is: from x1, x n+1 x 2n+1 ... x (n-1)*n+1 The first row of the matrix consists of x2, x n+2 x 2n+2 ... x (n-1)*n+2 This forms the second row, and so on, starting with x. n x 2n x 3n ... x n*n This forms the last line;

[0010] S2. Perform singular value decomposition on the complex matrix Q:

[0011]

[0012] Where, U∑ i V represents the i-th feature, ∑ i This indicates that it only contains the i-th eigenvalue σ. i A matrix whose all other elements are zero, with eigenvalues ​​σ. i This represents the characteristics of the Q matrix; the larger the eigenvalue σ, the better. i The corresponding more obvious feature U∑ i V; will obtain all ∑ i It is represented as an eigenvalue matrix ∑, where ∑ has n eigenvalues;

[0013] Set the smallest kn eigenvalues ​​in ∑ to zero to obtain ∑ new Where the threshold k∈[0,1), then matrix multiplication Q is used. new =U∑new V gets Q new The complex matrix Q is converted using the reverse operation of the method for converting to a complex matrix in S1. new The data is restored to a complex vector, and the denoised I / Q data is used as training data.

[0014] S3. Construct a lightweight deep learning model, including a feature extractor and a classifier. The feature extractor consists of complex convolutional layers, complex max pooling layers, and complex Bi-LSTM layers. The feature extraction module processes the input training data as follows: First, preliminary feature extraction is performed through six complex convolutional layers. Each complex convolutional layer consists of complex group convolution, complex batch regularization, and complex max pooling. Then, the features extracted by the complex convolutional layers are input into the complex Bi-LSTM layer for further feature extraction, resulting in an output feature vector. This feature vector is a complex vector, which is then input into the classifier. The classifier consists of fully connected layers, complex vector modulo layers, and a Softmax layer. The complex vector modulo layer converts the output of the fully connected layers into real vectors. Finally, the Softmax function maps the real vectors to the prediction confidence for each category, completing the classification.

[0015] S4. Use the training data obtained in S2 to train the lightweight deep learning model constructed in S3 to obtain the trained lightweight deep learning model.

[0016] S5. Input the acquired I / Q data into the trained lightweight deep learning model to obtain the radio frequency fingerprint recognition result. The beneficial effect of this invention is that it effectively solves the interference of environmental noise on radio frequency fingerprint recognition. Attached Figure Description

[0017] Figure 1 This is a diagram of the overall architecture of the present invention.

[0018] Figure 2 This is a schematic diagram of the vector-to-matrix conversion method of the present invention.

[0019] Figure 3 This is a schematic diagram of complex grouped convolution according to the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] The overall network structure of this invention is as follows: Figure 1 As shown, it mainly consists of two parts: a lightweight noise reduction module and an algorithm module based on deep learning in the complex domain.

[0022] Lightweight noise reduction module:

[0023] Adopting such Figure 2The method shown represents the complex vector X = (x1, x2, x3, ..., x...). N Convert to a complex matrix Q n×n ; where the complex vector X = (x1, x2, x3, ..., x... N ), where N is the length of the complex vector, x i Let represent each element in the vector, where i = 1, 2, 3, ..., N. Complex vectors are obtained through... Figure 2 The method shown transforms it into a matrix Q of size n×n.

[0024] After obtaining the complex matrix Q, eigenvalue decomposition of Q is performed using EVD:

[0025] Q=U∑V

[0026] Here, ∑ is a matrix of size (n, n), and the elements on the main diagonal of the ∑ matrix are all eigenvalues ​​σ. i This represents the i-th (i = 1, 2, ..., n) eigenvalue, with all other elements being zero.

[0027] After eigenvalue decomposition, Q can be viewed as the sum of several features:

[0028]

[0029] Where U∑ i V represents the i-th feature, ∑ i This indicates that it only contains the i-th eigenvalue σ. i A matrix whose all other elements are zero. Eigenvalue σ i This represents the characteristics of the Q matrix; the larger the eigenvalue σ, the better. i The corresponding more obvious feature U∑ i V, since the signal can be represented as:

[0030] f(t) = s(t) + n(t)

[0031] Where s(t) represents the signal component and n(t) represents the environmental noise component. Under normal circumstances, the signal strength will not be lower than the noise strength. Therefore, the characteristics corresponding to the signal component should be more obvious than those of the noise component, and its eigenvalues ​​should also be larger. Based on the above analysis, it can be concluded that retaining larger eigenvalues ​​and removing smaller eigenvalues ​​can achieve noise reduction to a certain extent.

[0032] Therefore, for the eigenvalue matrix ∑, we retain the larger eigenvalues ​​and set the smaller eigenvalues ​​to zero to obtain the eigenvalue matrix ∑. new Then, according to the formula Q=U∑ new V calculates a new complex matrix Q1, and finally rearranges Q1 into a complex vector to obtain the denoised signal.

[0033] Algorithm modules and optimizations based on deep learning in the complex field:

[0034] Compared to ordinary convolution, complex convolution does not increase the computational cost. The analysis is as follows:

[0035] First, let's analyze the calculation process of complex convolution. Assume the input feature map is:

[0036] F = X + Yi

[0037] Where X and Y are the real and imaginary parts of the input feature map, respectively, and i is the imaginary identifier. The convolution kernel is:

[0038] K = A + Bi

[0039] Where A and B are the real part convolution kernel and the imaginary part convolution kernel, respectively. The input features are convolved using these kernels:

[0040] K*F=(A+Bi)*(X+Yi)

[0041] Where * denotes a convolution operation, the final simplified form is:

[0042] K*F=(A*XB*Y)+(A*Y+B*X)i

[0043] From the above formula, we can see that K*F requires convolution operations of A*X, B*Y, A*Y, and B*X. Taking A*X as an example, for the sake of simplicity, we assume that A is a 1*1 kernel. Then, the number of FLOPs required for A*X is:

[0044] B = HWC in C out

[0045] Where C in For the number of input channels, C out Where W is the number of output channels, W is the width of the real part X of the output feature map, and H is the height of the real part X of the output feature map.

[0046] A complex convolution K*F consists of 4 ordinary convolutions such as A*X, so the FLOPs of a complex convolution are 4HWC. in C out Since both the input and output feature maps of a complex convolution consist of real and imaginary parts, the number of channels in the input and output feature maps of K*F are 2C. in 2C out Similarly, with a convolution kernel size of 1*1 and an input feature map channel count of 2C... in The number of channels, width, and height of the output feature map are 2C. out In the case of W and H, the FLOPs required for ordinary convolution are 4HWC. in Cout This is the same as the FLOPs of complex convolution.

[0047] In conclusion, introducing complex convolution does not increase the computational cost of convolution.

[0048] Based on this, the number of input channels and the number of output channels of complex convolutional layers should be kept as equal as possible, which can reduce the number of floating-point calculations and memory accesses.

[0049] Optimize the computational complexity of standard complex max pooling.

[0050] Complex modulo operation requires two floating-point multiplications and one floating-point summation. In the above process, assuming a matrix size of m×n, a complex modulo operation is needed for each element of the matrix, and the network involves multiple max-pooling operations, increasing computational overhead. I / Q data contains both positive and negative numbers, with the largest integer and smallest negative number being the "most salient features." Therefore, a complex max-pooling process for RF fingerprint recognition is proposed, preserving the salient features of I / Q data while having lower computational cost. Its implementation involves traversing the matrix composed of complex elements, using the real part of the element with the largest absolute value of its real part as the real part of the result, and the imaginary part of the element with the largest absolute value of its imaginary part as the imaginary part of the result, thus completing the complex max-pooling.

[0051] Example

[0052] In this example, the I / Q data in the signal consists of several sampling points, each containing phase and amplitude information. These sampling points together form an I / Q data set, where each sampling point contains both I-channel and Q-channel data, and its mathematical form is complex. Based on this characteristic, this invention treats the input I / Q data as a complex vector. The model combining high-performing convolutional neural networks and bidirectional long short-term memory neural networks from the deep learning field better preserves the temporal characteristics of the I / Q data. Furthermore, by replacing some real-number convolutions and operations with complex convolutions and operations, the learning and inference process of the network better aligns with the complex composition of the I / Q data, thus better preserving the coupling characteristics between the I-channel and Q-channel data. The above design is as follows... Figure 3 As shown.

[0053] In this embodiment, on the one hand, from the perspective of data, the data used is the I / Q data transmitted by the wireless device. The I and Q data are combined to form a 1×N (N represents the number of values) one-dimensional complex vector, which is used as input data. On the other hand, from the perspective of model, generally speaking, only complex deep neural networks can support complex inputs, which can better extract the features of the input I / Q data.

[0054] Assuming we take n I / Q data samples of size 1×256 (containing k categories) as input, and process them through the network structure shown in Table 1 below, we will finally obtain the confidence score for each category:

[0055] Table 1 Model Structure Details

[0056]

[0057] The complex convolution employed in this invention utilizes the previously mentioned grouped convolution for computational optimization; the complex max pooling used is the computationally optimized complex max pooling method described earlier; and the activation function employed is Hard-Swish, which exhibits similar performance to other activation functions but requires less computation. Furthermore, in this invention, to accommodate complex inputs, it is extended from the real domain to the complex domain. Through these design features, the computational load required for the RFID fingerprint recognition model is reduced, accelerating the model's inference speed.

Claims

1. A radio frequency fingerprint recognition method based on a lightweight deep learning model, characterized in that, Includes the following steps: S1. Convert the acquired I / Q signals transmitted by the wireless device into a complex matrix, specifically: The acquired I / Q signals are written as complex vectors. Where N is the number of samples, This represents each element in the vector. ,Will Transform into a complex matrix ,in The transformation method is: from , , ... The first row of the matrix is ​​formed by , , ... This forms the second row, and so on, from... , , ... This forms the last line; S2. Perform eigenvalue decomposition on the complex matrix Q: ， in, Indicates the first One characteristic, Indicates only the first eigenvalues A matrix in which all other elements are zero, and its eigenvalues. This represents the characteristics of the Q matrix; the larger the eigenvalue, the better. Corresponding to more obvious features All that will be obtained Represented as an eigenvalue matrix , There are n eigenvalues; Will The smallest Setting each eigenvalue to zero yields... , where the threshold Then use matrix multiplication. get The complex matrix is ​​converted using the reverse operation of the method in S1 to a complex matrix. The data is restored to a complex vector, and the denoised I / Q data is used as training data. S3. Construct a lightweight deep learning model, including a feature extractor and a classifier. The feature extractor consists of complex convolutional layers, complex max pooling layers, and complex Bi-LSTM layers. The feature extraction module processes the input training data as follows: First, it performs preliminary feature extraction through six complex convolutional layers. Each complex convolutional layer consists of complex group convolution, complex batch regularization, and complex max pooling. Then, the features extracted by the complex convolutional layers are input into the complex Bi-LSTM layer for further feature extraction, resulting in an output feature vector. This feature vector is a complex vector, which is then input into the classifier. The classifier consists of fully connected layers, complex vector modulo layers, and a Softmax layer. The complex vector modulo layer converts the output of the fully connected layers into real vectors. Finally, the Softmax function maps the real vectors to the prediction confidence for each category, completing the classification. S4. Use the training data obtained in S2 to train the lightweight deep learning model constructed in S3 to obtain the trained lightweight deep learning model. S5. The I / Q signals transmitted by the acquired wireless device are processed using methods such as S1 and S2 to obtain I / Q data, which are then input into the trained lightweight deep learning model to obtain the radio frequency fingerprint recognition result.

Images