A machine learning-based radio frequency fingerprinting method

Abstract

This invention discloses a machine learning-based radio frequency fingerprint recognition method. The method comprises the following steps: collecting radio frequency fingerprints from existing devices; constructing a neural network structure, importing radio frequency fingerprint data for training, and deriving a recognition model; importing the radio frequency fingerprint of the device under test into the trained recognition model for verification, and obtaining the recognition result. This invention uses a file reading method to read the radio frequency signal sampling data of the device to be identified. The neural network discriminates the read data, extracting constellation point features, carrier offset, etc., from the sampling data for classification, thus achieving contactless device identification and authentication. Whenever the received device radio frequency signal changes, it is processed and input into a convolutional neural network to obtain the recognition result. This eliminates the need for repetitive and tedious calculations, reducing costs, and achieving a superior recognition rate with lower complexity.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, and in particular to a radio frequency fingerprinting method based on machine learning. Background Technology

[0002] The security issues of IoT application systems mainly include three aspects: resistance to signal interference, resistance to malicious intrusion, and communication protection. Due to its flexible networking and ease of deployment, wireless communication has become the preferred solution for solving the "last mile" communication problem in IoT systems. However, due to its openness, wireless communication systems are more vulnerable to malicious attacks compared to traditional wired communication systems. Therefore, device radio frequency fingerprinting technology has emerged.

[0003] Wireless devices possess distinct radio frequency (RF) fingerprints—hardware differences reflected in their communication signals. These differences can be extracted by analyzing the received RF signals. This method of extracting device hardware characteristics based on communication signals is called "RF fingerprint extraction," which allows for the identification of different wireless devices. This method was first proposed by Hall et al. in Canada during their research on the classification and identification of wireless devices such as Bluetooth. It utilizes the transient components of the transmitter signal to uniquely identify the transmitter—this technique is known as RF fingerprinting. They constructed an intrusion detection system for IEEE 802.11b devices based on RF fingerprinting to combat attacks such as MAC address cloning.

[0004] However, existing radio frequency fingerprint recognition technology suffers from problems such as cumbersome calculations, high costs, and low recognition accuracy, which urgently need to be addressed. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based radio frequency fingerprint recognition method that has low computational cost, high recognition rate, and low cost.

[0006] The technical solution to achieve the objective of this invention is: a machine learning-based radio frequency fingerprint recognition method, comprising the following steps:

[0007] Step 1: Collect the radio frequency fingerprint of existing devices;

[0008] Step 2: Build a neural network structure, import RF fingerprint data for training, and derive a recognition model;

[0009] Step 3: Import the RF fingerprint of the device under test into the trained recognition model for verification and obtain the recognition result.

[0010] Compared with the prior art, the present invention has the following significant advantages: (1) By using deep learning, a training set is constructed and a convolutional neural network is trained. After training, whenever the received device radio frequency signal changes, it is processed and input into the convolutional neural network to obtain the recognition result. There is no need to repeat tedious calculations. With low complexity, a better recognition rate is obtained; (2) After the network is trained, as long as the device radio frequency fingerprint database does not change, there is no need to perform new training. It can distinguish different wireless devices without demodulating communication data, and can realize contactless device identification and authentication.

[0011] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0012] Figure 1 This is a flowchart of the radio frequency fingerprint recognition method based on machine learning according to the present invention.

[0013] Figure 2 This is a flowchart of the device's radio frequency signal processing procedure.

[0014] Figure 3 This is the differential constellation diagram output after processing the device's RF fingerprint.

[0015] Figure 4 This is a diagram of the convolutional neural network structure. Detailed Implementation

[0016] This invention aims to provide a machine learning-based method for device radio frequency fingerprinting, which distinguishes different wireless devices without demodulating communication data, enabling contactless device identification and authentication. The method involves processing the collected device radio frequency signals to generate training samples, forming training and testing sets. Labels are generated using the raw data to train a convolutional neural network (CNN), and the accuracy is finally provided through the testing set. After training, the CNN is input into software to obtain the output, which is the device identification result.

[0017] To achieve the above objectives, the technical solution adopted in this invention includes: processing the received device RF fingerprint to obtain a differential constellation diagram, generating training samples, forming a training set and a test set, and generating labels using the original data. A convolutional neural network is trained using the training set and its labels, and the neural network is tested using the test set.

[0018] Furthermore, the specific process includes:

[0019] S1. The received device RF signal is preprocessed, IQ quadrature branch processed, delayed, and differentially divided to generate a differential constellation diagram;

[0020] S2. Process the difference constellation diagram into a two-dimensional matrix to generate training samples, where 90% constitute the training set and 10% constitute the test set, and generate labels corresponding to the training set;

[0021] S3. Using the Keras framework, employing stochastic gradient descent, and introducing the Adam adaptive learning rate algorithm and regularization method, a CNN is built, and the convolutional neural network is trained using the training set and its labels;

[0022] S4. Input the processed device RF signal into the convolutional neural network (CNN) to obtain the output, which is the device identification result. Whenever the received device RF signal changes, it is processed and input into the CNN to obtain the identification result. After the CNN is trained, no new training is required as long as the device RF fingerprint database does not change.

[0023] This invention discloses a machine learning-based radio frequency fingerprint recognition method, comprising the following steps:

[0024] Step 1: Collect the radio frequency fingerprint of existing devices;

[0025] Step 2: Build a neural network structure, import RF fingerprint data for training, and derive a recognition model;

[0026] Step 3: Import the RF fingerprint of the device under test into the trained recognition model for verification and obtain the recognition result.

[0027] As a specific example, step 1, which involves collecting the radio frequency fingerprint of an existing device, is as follows:

[0028] The system collects radio frequency signals emitted by multiple wireless communication devices, processes the raw signals to obtain a differential constellation diagram to extract radio frequency fingerprint information, distinguishes different devices, labels them differently, and classifies the processed data into different folders.

[0029] As a specific example, the construction of the neural network structure described in step 2 is as follows:

[0030] Step S1: Using the Keras framework, stochastic gradient descent is employed, and the Adam adaptive learning rate algorithm and regularization method are introduced to build a DNN neural network.

[0031] Step S2: The DNN neural network has 3 convolutional layers followed by 1 pooling layer, which is looped 4 times and then connected to 3 fully connected layers. The activation function is the softmax function.

[0032] As a specific example, step 2, which involves importing radio frequency fingerprint data to train and derive a recognition model, is detailed as follows:

[0033] The radio frequency fingerprint information of the device to be trained is imported into the neural network. 90% of the data is used as training data and 10% as test data. The recognition model is derived through neural network operation and the accuracy is given.

[0034] As a specific example, step 3 involves importing the radio frequency fingerprint of the device under test into the trained recognition model for verification, and obtaining the recognition result, as follows:

[0035] Collect the radio frequency fingerprint information of the device under test, import it into the trained recognition model, obtain the row vector representing the tag, determine whether it belongs to the already recorded device, and identify which device it is based on the accuracy.

[0036] As a specific example, step 1 is as follows:

[0037] After acquiring the device's radio frequency signal, the signal is first preprocessed and its energy is normalized to eliminate the influence of different transmission power at the transmitting end on the signal.

[0038] The preprocessed baseband signal is divided into two paths. One path is directly sent to the differential module, and the other path is sent to the I / Q delay circuit. After processing by the delay circuit, the signal enters the differential module, which draws the differential constellation diagram on the complex plane.

[0039] After obtaining the differential constellation diagram, the processed data is categorized into different folders according to different devices.

[0040] As a specific example, step 2 is as follows:

[0041] First, the difference constellation diagram is processed into a two-dimensional real-valued matrix, and the training samples are divided into training set and test set in a 9:1 ratio;

[0042] Let the number of devices be N, then the number of generated tags is N. Each device's tag is 0 or 1, denoted in matrix form. When N = 3, the tag of device 1 is [1 0 0].

[0043] Secondly, using the Keras framework, stochastic gradient descent, and introducing the Adam adaptive learning rate algorithm and regularization method, a CNN model is built, and the convolutional neural network is trained using the training set and its labels.

[0044] The first few layers of the CNN model use convolutional and pooling layers, while the last three layers use fully connected layers. The parameters and weights are adjusted using the backpropagation (BP) algorithm, as detailed below:

[0045] (1) Convolutional layer convolution: In a convolutional layer, the feature map is convolved with the output feature map of the previous layer by several different convolutional kernels, and then passed through an activation function. Each output map is a weighted sum of multiple input maps convolved.

[0046] (2) Downsampling layer sampling: The function of the downsampling layer is to reduce the dimension of the output map of the previous layer connected to it by a set ratio, but without changing the number of output feature maps;

[0047] (3) Adjusting parameters in the backpropagation (BP) algorithm:

[0048] (a) Forward propagation: For multi-class problems, there are C classes and N training samples. The cost function is represented by the squared error function.

[0049] (b) Backpropagation: Backpropagation error is the partial derivative of the cost function with respect to the weights w and biases b;

[0050] (4) Recognition rate calculation: The error between the output label value and the ideal label is ε. Find the samples that meet ε≤0.1 and count the number of samples. This gives the recognition rate:

[0051] correct=∑x / M

[0052] Where x is 1 when ε≤0.1 and 0 when ε>0.1; the denominator M is the total number of test data.

[0053] As a specific example, step 3 is as follows:

[0054] The RF fingerprint of the device under test is imported into the trained recognition model for verification, resulting in an N-dimensional row vector M:

[0055] M = [m1m2…m N ]

[0056] in Let max{m1m2…m N}=m k If m k If the value is greater than 0.8, the device is identified as device k; otherwise, the device is not in the database.

[0057] The present invention will be further illustrated below with reference to the accompanying drawings and specific examples. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0058] Combination Figure 1 This embodiment provides a device radio frequency fingerprinting method based on deep learning. Assume there are N wireless communication devices, and N sets of device radio frequency signals of duration t are collected. The received signals are input into a signal processor to generate a differential constellation diagram, which is then input into a convolutional neural network for device identification.

[0059] The device radio frequency fingerprinting method based on deep learning mainly includes the following steps:

[0060] S1: Process the acquired device radio frequency signals to generate a differential constellation diagram.

[0061] The input to the CNN-based radio frequency fingerprinting network of this invention is processed radio frequency image data, using a differential constellation diagram instead of the original data points, which helps to accelerate the convergence of the CNN network.

[0062] After acquiring the device's RF signal, a simple preprocessing step is performed. The primary purpose is to normalize the signal energy and eliminate the influence of varying transmit power at the transmitting end. The preprocessed baseband signal is then split into two paths: one is directly fed into the differential module, and the other is fed into both I / Q delay units. After processing by the delay units, the signal enters the differential module, allowing a stable and clear differential constellation diagram to be plotted on the complex plane. Once the differential constellation diagram is obtained, the processed data is categorized into different folders according to the different devices. The device RF signal processing flow is as follows: Figure 2 As shown, the difference constellation diagram is as follows: Figure 3 As shown.

[0063] S2. Process the difference constellation diagram into a two-dimensional real-valued matrix to generate training samples, forming a training set and a test set.

[0064] First, the difference constellation diagram is processed into a two-dimensional real-valued matrix, and the training samples are divided into training set and test set in a 9:1 ratio;

[0065] Let the number of devices be N, then the number of generated tags is N. Each device's tag is 0 or 1, and it is represented in matrix form. For example, when N = 3, the tag of device 1 is [1 0 0].

[0066] S3. Using the Keras framework, stochastic gradient descent, and introducing the Adam adaptive learning rate algorithm and regularization method, a CNN is built, and the convolutional neural network is trained using the training set and its labels.

[0067] The first few layers of a CNN model primarily use convolutional and pooling layers, while the last three layers employ fully connected layers. Parameter and weight adjustments are performed using the backpropagation (BP) algorithm. The algorithm is described in detail below:

[0068] (1) Convolutional Layer Convolution: In a convolutional layer, the feature map is obtained by convolving several different convolutional kernels with the output feature map of the previous layer, and then passing it through an activation function, f(·). Each output map may be a weighted sum of multiple input maps convolved together.

[0069]

[0070] (2) Downsampling layer sampling: The function of the downsampling layer is to reduce the dimension of the output map of the preceding layer by a certain proportion, without changing the number of output feature maps. The formula is:

[0071]

[0072] Where down(·) represents a downsampling function that reduces the output image by n. 2 The activation function here is the same as that in other layers.

[0073] (3) Adjusting parameters in the backpropagation (BP) algorithm:

[0074] (a) Forward propagation: For multi-class problems, there are C classes and N training samples. The cost function is represented by the squared error function, i.e.:

[0075]

[0076]

[0077] in, This represents the ideal label value of the nth image in the kth class of samples. This represents the network output label value corresponding to the input graph.

[0078] (b) Backpropagation: The backpropagation error is the partial derivative of the cost function with respect to the weights w and biases b. The parameter update is defined as follows, where α is the learning rate.

[0079]

[0080]

[0081] (4) Recognition rate calculation: The error between the output label value and the ideal label can be expressed as: ε=∥OT∥. Where T is the ideal label value.

[0082] The recognition rate can be obtained by finding and counting the samples that meet the condition ε≤0.1.

[0083] correct=∑x / M

[0084] Where x is 1 when ε≤0.1 and 0 when ε>0.1. The denominator is the total number of test data.

[0085] Figure 4 A schematic diagram of the CNN network with multiple hidden layers used in this software is provided. The activation function of each layer is the ReLU function, and the last layer uses the softmax function for output.

[0086] S4. Input the processed device RF signal and the trained model into the convolutional neural network for recognition, and obtain the N-dimensional row vector M:

[0087] M = [m1m2…m N ]

[0088] in Let max{m1m2…m N}=m k If m k If the value is greater than 0.8, the device is identified as device k; otherwise, the device is not in the database.

Claims

1. A machine learning based radio frequency fingerprinting method, characterized in that, The steps are as follows: Step 1, collect the radio frequency fingerprints of existing equipment, as follows: Collect the radio frequency signals emitted by multiple wireless communication devices, process the original signals to obtain differential constellation diagrams to extract radio frequency fingerprint information, distinguish according to different devices, mark different labels, and classify the processed data into different folders; Step 2, build a neural network structure, import radio frequency fingerprint data for training and obtain a recognition model, as follows: First, process the differential constellation diagram into a two-dimensional real value matrix, and divide the training samples into a training set and a test set according to a ratio of 9:1; Let the number of devices be N, then the number of labels generated is N, and the label of each device is 0 or 1, denoted as a matrix. When N=3, the label of device 1 is [1 0 0]; Second, use the Keras framework, use the stochastic gradient descent method, introduce the Adam adaptive learning rate algorithm and the regularization method, build a CNN model, and use the training set and its labels to train the convolutional neural network; The first few layers of the CNN model use convolution and pooling layers, and the last three layers use fully connected methods. The parameter and weight adjustment uses the BP algorithm, as follows: (1) Convolution in the convolution layer: In a convolution layer, the feature map is obtained by convolving several different convolution kernels with the output feature map of the previous layer, and then passing through an activation function. Each output map is the weighted sum of the convolution of multiple input maps; (2) Sampling in the down-sampling layer: The role of the down-sampling layer is to reduce the dimension of the output map of the previous layer connected to it by a certain proportion, but does not change the number of output feature maps; (3) Parameter adjustment by back propagation BP algorithm: (a) Forward propagation: For multiclass problems, there are C classes, and C training samples, the cost function is represented by a squared error function. (b) Backpropagation: The error, i.e. the cost function, is backpropagated through the weights and biases to their partial derivatives. (4) Recognition rate calculation: the error between the output label value and the ideal label is , find the samples that meet and count the number, and then get the recognition rate : ； wherein, when , take 1, when , take 0; the denominator is the total number of test data; Step 3, import the radio frequency fingerprint of the device to be tested into the trained recognition model for testing to obtain the recognition result. 2.The machine learning based radio frequency fingerprinting method of claim 1, wherein, The neural network structure in step 2 is built as follows: Step S1, use the Keras framework, use the stochastic gradient descent method, introduce the Adam adaptive learning rate algorithm and the regularization method, and build a DNN neural network; Step S2, the DNN neural network is provided with 3 convolution layers and 1 pooling layer, and after 4 cycles, 3 fully connected layers are connected, and the activation function uses the softmax function. 3.The machine learning based radio frequency fingerprinting method of claim 1, wherein, The training of the neural network and the obtaining of the recognition model in step 2 are as follows: Import the radio frequency fingerprint information of the device to be trained into the neural network, use 90% of the data as training data and 10% as test data, and output the recognition model after neural network operation and give the accuracy rate. 4.The machine learning based radio frequency fingerprinting method of claim 1, wherein, Step 3 introduces the radio frequency fingerprint of the device to be tested into the trained recognition model for testing to obtain the recognition result, as follows: Collect the radio frequency fingerprint information of the device to be tested, and import it into the trained recognition model to obtain a row vector representing the label, and determine whether it belongs to the device that has been recorded, and identify which device according to the accuracy rate. 5.The machine learning based radio frequency fingerprinting method of claim 1, wherein, The step 1 is as follows: After collecting the radio frequency signals of the device, first, pre-process the signals, normalize the energy of the signals, and exclude the influence of different transmission powers on the signals; The preprocessed baseband signal is divided into two paths, one of which is directly sent to a differential module, and the other is sent to an I / Q two-path delay device, and then enters the differential module after being processed by the delay device, that is, a differential constellation diagram is drawn on the complex plane; After obtaining the differential constellation diagram, according to different equipment, the processed data is classified into different folders. 6.The machine learning based radio frequency fingerprinting method of claim 1, wherein, The step 3 is specifically as follows: The radio frequency fingerprint of the to-be-tested equipment is introduced into the trained identification model for testing, and an N-dimensional row vector M is obtained: The radio frequency fingerprint of the to-be-tested equipment is introduced into the trained identification model for testing, and an N-dimensional row vector M is obtained: ； wherein , record , if then the device is identified as device k, otherwise the device is not in the database.

Images