A radio frequency fingerprinting training data generation method based on a generative adversarial network

Abstract

The application belongs to the technical field of Internet of Things security, and particularly relates to a radio frequency fingerprint identification training data generation method based on a generative adversarial network. Although the radio frequency fingerprint identification technology is increasingly mature, when the training data is insufficient, the model obtains less effective information from the original data, and the model will have an overfitting phenomenon. Especially in some special environments, such as communication equipment in a relatively concealed environment of a UAV, a large number of signals cannot be obtained for radio frequency fingerprint identification. The radio frequency fingerprint sample enhancement method based on the generative adversarial network model is used for extracting more effective radio frequency fingerprint features and enhancing data, the discriminant model and the generative model of the generative adversarial network are modified in combination with the characteristics of I / Q signals, and the Wasserstein distance formula is used as a function for calculating the sample feature similarity, so that the model can more distinguish the differences between I / Q signal samples, thereby ensuring that the data generated by the model can match the original data.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things security technology, specifically relating to a method for generating radio frequency fingerprint recognition training data based on generative adversarial networks. Background Technology

[0002] With the widespread use of wireless devices, the field of IoT security has received increasing attention. End-to-end information transmission and inter-device authentication present challenges. Traditional methods use IP addresses or cookies for authentication, which require complex encryption algorithms. For some devices with limited computing power, identifying the device becomes a burden on the processor.

[0003] Currently, using neural networks to identify radio frequency fingerprint signals is very popular. However, some devices have long signal transmission cycles, making data sample acquisition difficult and resulting in insufficient data volume, which hinders the training of neural networks. This raises the issue of imbalanced sample sizes among different signal categories in the acquired signals. Common signal data augmentation methods typically employ deep neural networks to construct signal samples, with the sample distribution referencing the distribution of existing signal features. However, due to limitations in the number of signal samples, the distribution of existing sample features cannot represent the actual feature distribution of the signal category. This leads to augmented data that fails to represent the actual signal features, resulting in overfitting and other issues during signal recognition, preventing the model from achieving generalization. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a method for generating radio frequency fingerprint recognition training data based on generative adversarial networks, which solves the problems of low algorithm accuracy and overfitting when training data is insufficient.

[0005] The technical solution of this invention is as follows:

[0006] A method for generating radio frequency fingerprint recognition training data based on generative adversarial networks includes the following steps:

[0007] S1. The acquired radio equipment I / Q signals are preprocessed to obtain a training dataset. Specifically, the I-channel and Q-channel data are combined to form training data with each data item having a format of 2×1024.

[0008] S2. Feature vectors are obtained by extracting features from the training data through a feature vector extraction network. The feature vector extraction network consists of 9 convolutional layers, 1 max pooling layer and 1 fully connected layer connected in sequence.

[0009] S3. Construct a generative adversarial network, including a generator and a discriminator. The generator's input is a feature vector, and its output is the generated sample data. The discriminator's input consists of the generated sample data and the training data. The generator is composed of three sequentially connected deconvolutional modules. Each deconvolutional module includes a deconvolutional layer and a batch normalization layer. The generator ultimately maps its output to the range [-1, 1] using the tanh function. The discriminator is composed of three convolutional modules. Each convolutional module includes a convolutional layer and an average pooling layer. The discriminator outputs a real value to calculate the Wasserstein distance.

[0010] S4. Train the constructed generative adversarial network. During training, first train the discriminator 5 times, then train the generator 1 time. After each training iteration, backpropagation is used to learn the weight parameters. The objective function is the Wasserstein distance.

[0011]

[0012] Among them, f w It is a discriminator network, where x is the training data, w is the discriminator parameters, and f is the discriminator network. w (x) is the judgment result, G θ It is a generator network, where z is the noise input, θ is the generator parameters, and G is the generator network. θ (z) represents the generated sample data, p r It is the data distribution of the training data, p noise It is the data distribution of the generated samples;

[0013] The RMSProp algorithm is used as the optimizer:

[0014]

[0015] Among them, v t Let g be the exponentially weighted average of the squared gradient of the t-th order. t Let represent the gradient at the t-th iteration, and β be a decay coefficient, typically 0.9.

[0016] Then, update the parameters:

[0017]

[0018] Among them, w t Let represent the weight at the t-th iteration, η represent the learning rate, and ∈ is a constant used to prevent division by zero, typically 10. -8 ;

[0019] The model converges when the discriminator is trained to a recognition rate of 50%, resulting in a well-trained generative adversarial network.

[0020] S5. Use the generator in the trained generative adversarial network to generate training data for radio frequency fingerprint recognition.

[0021] The beneficial effects of this invention are that, addressing the problems of decreased algorithm accuracy and overfitting when training data is insufficient, this invention employs a network model-based radio frequency fingerprint sample enhancement method to generate data for training the recognition model, effectively solving the problem caused by insufficient training data. To address the issue of insufficient training data, this invention applies deep learning-based data augmentation methods to radio frequency fingerprint recognition technology, researching a radio frequency fingerprint sample enhancement method based on a generative adversarial network model. Attached Figure Description

[0022] Figure 1 This is a flowchart of the radio frequency fingerprint sample enhancement method based on generative adversarial networks proposed in this invention.

[0023] Figure 2 This is a framework diagram of the generative adversarial network of the present invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] Example

[0026] The processing flow in this example is as follows: Figure 1 As shown, it includes the following steps:

[0027] S1. Input I / Q data, perform preprocessing, and construct the training dataset; the specific method is as follows:

[0028] The input data format is defined as N×4096, where N is a positive integer; the input data is in complex form, representing I and Q signals, with the real part being the I-channel data and the imaginary part being the Q-channel data; 1024 digital signals are used as one signal sample, with a step size of 128 to divide the digital signals, finally obtaining 24×N signal samples in the form of 2×1024. The signal samples are normalized line by line to obtain s[n]:

[0029]

[0030] Where, x rms It is the root mean square of the signal amplitude, and r[n] can be regarded as the received signal.

[0031] The preprocessed samples and labels form a training set, which is then used as the input to the feature extraction network.

[0032] S2. Construct a CNN feature extraction network and train it to output feature vectors. In this example, the CNN feature extraction network consists of 9 convolutional layers, 1 max-pooling layer, and 1 fully connected layer. The first convolutional layer uses 64 7×7 convolutional kernels with a stride of 2; layers 2 to 5 use 64 3×3 convolutional kernels with a stride of 1; and layers 6 to 9 use 128 3×3 convolutional kernels with a stride of 1. All convolutional layers are activated by the Rectified Linear Array (ReLU) function and padded. Dropout is also used to prevent overfitting. The input to the CNN feature extraction network is the training data, which is an N×2×1024 (N represents the number of signal lines, and 1024 represents that each signal consists of 1024 data points) three-dimensional matrix formed by combining the I-channel and Q-channel data from S1. In this example, the data in the training set is randomly shuffled and input into the feature extraction network in batches to obtain the feature vector output. The advantage of this approach is that the feature vectors can reflect the distribution characteristics of the signal samples to a certain extent. Data generated based on these distribution characteristics is closer to the real samples. This can accelerate the convergence speed of the network and enhance the training effect.

[0033] S3. Construct a Wasserstein generative adversarial network (GAN) and train the generator and discriminator. The GAN consists of a generator and a discriminator. The input layer of the generator network uses fingerprint features extracted by the feature extraction network, which is N×256 data points. Then, three deconvolution modules are used, each containing a deconvolution layer and a batch normalization layer. The first deconvolution layer uses 256 4×4 convolution kernels with a stride of 1; the second deconvolution layer uses 512 4×4 convolution kernels with a stride of 2; and the third deconvolution layer uses 1024 4×4 convolution kernels with a stride of 1. After each deconvolution operation, batch normalization is performed and ReLU activation is applied. Finally, the tanh function is used to map the N×2×1024 sample data to the range [-1, 1]. The discriminator's input layer uses a 3D matrix of N×2×1024, formed by randomly shuffling fake samples provided by the generator network and real samples after preprocessing the original data. Then, it passes through three convolutional modules, each consisting of convolutional layers and average pooling layers. The first convolutional layer uses 16 2×5 kernels with a stride of 1; the second uses 32 2×5 kernels with a stride of 1; and the third uses 64 1×5 kernels with a stride of 1. All convolutional layers are activated by the LeakyReLu function and have zero-padded edges. After each convolution, a 2×2 average pooling operation is performed. The WGAN discriminator directly outputs a real value for calculating the Wasserstein distance. The specific training process is as follows:

[0034] The discriminator is trained 5 times, followed by the generator 1 time. After each training iteration, backpropagation is used to learn the weight parameters, with the objective function being the Wasserstein distance, as shown in the following formula:

[0035]

[0036] Among them, f w It is a discriminator network, where x is the training data, w is the discriminator parameters, and f is the discriminator network. w (x) is the judgment result, G θ It is a generator network, where z is the noise input, θ is the generator parameters, and G is the generator network. θ (z) represents the generated sample data, p r It is the data distribution of the training data, p noise It is the data distribution for generating samples.

[0037] To accelerate convergence and enhance training performance, this invention employs the RMSProp algorithm as the optimizer, as shown in the following formula:

[0038]

[0039] First, calculate the exponentially weighted average of the squared gradient. Where g... t Let t represent the gradient at the t-th time, and β be a decay coefficient, typically 0.9.

[0040] Then, update the parameters of the generator and discriminator respectively, as shown in the following formula:

[0041]

[0042] Among them, w t Let g represent the weight in the t-th iteration. t Let η represent the gradient at the t-th iteration, and let η represent the learning rate, which is set to 10 in this embodiment. -6 ∈ is a constant used to prevent division by zero, and is usually 10. -8 v t Let be the exponentially weighted average of the squared gradient of the t-th time step.

[0043] When the discriminator is trained to a recognition rate of 50%, the model converges, and the trained generative adversarial network is obtained.

[0044] S4. Use the trained generator to generate radio frequency fingerprint sample data to serve the radio frequency fingerprint recognition network.

Claims

1. A method for generating radio frequency fingerprint recognition training data based on generative adversarial networks, characterized in that, Includes the following steps: S1. Preprocess the acquired I / Q signals from the wireless device to obtain the training dataset. Specifically, combine the I-channel and Q-channel data to form training data with each data entry having a format of 2×1024. The method is as follows: The input data format is defined as N×4096, where N is a positive integer; the input data is in complex form, representing I and Q signals, with the real part being the I-channel data and the imaginary part being the Q-channel data; 1024 digital signals are used as one signal sample, and the digital signals are divided with a step size of 128, resulting in 24×N signal samples in the form of 2×1024. These signal samples are then normalized line by line to obtain... : ， in, It is the root mean square of the signal amplitude. The signal is considered to be received. S2. Feature vectors are obtained by extracting features from the training data through a feature vector extraction network. The feature vector extraction network consists of 9 convolutional layers, 1 max pooling layer and 1 fully connected layer connected in sequence. S3. Construct a generative adversarial network, including a generator and a discriminator. The generator's input is a feature vector, and its output is the generated sample data. The discriminator's input consists of the generated sample data and the training data. The generator is composed of three sequentially connected deconvolutional modules. Each deconvolutional module includes a deconvolutional layer and a batch normalization layer. The generator ultimately maps its output to the range [-1, 1] using the tanh function. The discriminator is composed of three convolutional modules. Each convolutional module includes a convolutional layer and an average pooling layer. The discriminator outputs a real value to calculate the Wasserstein distance. S4. Train the constructed generative adversarial network. During training, first train the discriminator 5 times, then train the generator 1 time. After each training iteration, backpropagation is used to learn the weight parameters. The objective function is the Wasserstein distance. ， in, It is a discriminator network. It's training data. These are discriminator parameters. It is the judgment result. It is a generator network. For noise input, These are generator parameters. These are the generated sample data. It is the data distribution of the training data. It is the data distribution of the generated samples; The RMSProp algorithm is used as the optimizer: ， in, For the first Exponentially weighted average of the squared subgradient Indicates the first The gradient of the order of 1. It is an attenuation coefficient; Then, update the update parameters of the generator and discriminator respectively, as shown in the following formula: ， in, Indicates the first The weight of the second, Indicates the learning rate. It is a constant used to prevent division by zero; The model converges when the discriminator is trained to a recognition rate of 50%, resulting in a well-trained generative adversarial network. S5. Use the generator in the trained generative adversarial network to generate training data for radio frequency fingerprint recognition.

Images