A Method for Identifying the Radiation Source of IEEE 802.11a Signals with Channel Characteristics Removed

Abstract

The present invention discloses a method for identifying the radiation source of IEEE802.11a signals with channel characteristics removed, comprising the following steps: 1) Collecting IEEE802.11a device signals to be identified; 2) Extracting time-domain training sequences; 3) Training an LMS filter; 4) Performing channel equalization processing; 5) Training a neural network; 6) Identifying device identities. This method has the advantages of strong anti-interference ability, high recognition rate, and low implementation difficulty.

Description

Technical Field

[0001] The present invention relates to the field of radio physical layer security, and specifically to a method for identifying the radiation source of IEEE802.11a signals with channel characteristics removed. Background Art

[0002] In recent years, with the development of the Internet of Things, civilian small unmanned aerial vehicles, and fifth-generation mobile communication technology, the number of wireless devices has shown an explosive growth. The wireless signal transmissions of these devices are mostly based on the IEEE802.11 series protocols. Due to the openness of electromagnetic waves, there are potential safety hazards in the traditional wireless network security based on key security protocols. Radio frequency fingerprint recognition based on the physical layer has the advantage that the features are difficult to forge, can effectively and correctly identify the identity of radio frequency devices, and is one of the research hotspots in the field of wireless network security.

[0003] Existing radio frequency fingerprint recognition technologies mainly use convolutional neural networks to extract and classify radio frequency fingerprint features from the time domain or transform domain of received wireless signals. However, in the practical application of radio frequency fingerprint recognition technology, there is still an important problem. Wireless signals are easily interfered by time-varying channels, resulting in the neural network extracting not only the radio frequency fingerprint features of the signal radiation source but also the time-varying channel characteristics from the signal, ultimately leading to a sharp decline in recognition accuracy.

[0004] Channel equalization is an effective technology for eliminating channel characteristics. The adaptive filter algorithm LMS (Least Mean Square, abbreviated as LMS) can be used as a channel equalizer in a communication system. Using the standard IEEE802.11a time domain training sequence as a reference signal and combining the LMS adaptive filter to perform channel equalization and compensation on the signal to be recognized can reduce the influence of the time-varying channel on radio frequency fingerprint recognition.

[0005] The Convolutional Neural Network structure based on IQ Correlation features (IQCNet) is a convolutional neural network for extracting IQ signal features. IQCNet extracts IQ correlation features and time domain features, retains the phase information of the signal in extracting the radio frequency fingerprint information of IQ signals, obtains the mean value of each channel feature through adaptive average pooling, and uses a single fully connected layer for classification. Compared with the traditional convolutional network structure, using IQCNet to extract radio frequency fingerprint features has the advantages of higher recognition accuracy and smaller computational complexity. Summary of the Invention

[0006] The object of the present invention is to provide a method for identifying the radiation source of IEEE 802.11a signals with channel characteristics removed, aiming at the problem in the existing radio frequency fingerprint recognition technology that when using a convolutional neural network to extract radio frequency fingerprints, it is easily interfered by channel characteristics, resulting in a sharp decline in recognition accuracy. This method can effectively remove the adverse effects of the wireless channel on radio frequency fingerprint recognition in different wireless channel environments, achieving good recognition accuracy, and has the advantages of strong anti-interference ability, high recognition rate, and low implementation difficulty.

[0007] The technical solution for achieving the object of the present invention is as follows:

[0008] A method for identifying the radiation source of IEEE 802.11a signals with channel characteristics removed, comprising the following steps:

[0009] 1) Collect the signals of IEEE 802.11a devices to be identified: IEEE 802.11a adopts the standard OFDM technology. For the time-domain signal sequence of the OFDM signal model, it is shown as formula (1):

[0010]

[0011] Among them, X(k) is the frequency-domain data on the k-th subcarrier in OFDM, N represents the total number of N subcarriers in the OFDM signal. In the radio frequency circuit of a radio frequency device, there are non-ideal manufacturing errors in the devices, and the circuit traces are different, which will cause slight differences in the circuit parameters of different radio frequency devices. This slight difference will be added to the transmitted signal in the form of unintentional modulation, resulting in the actual baseband signal model transmitted by the radio frequency device as shown in formula (2):

[0012]

[0013] Among them, Δa(n) represents the unintentional modulation function of the signal envelope, represents the signal phase error, Δf represents the signal frequency offset. For a radio frequency device, there is always a deviation between the actual transmitted signal and the ideal modulation signal, resulting in the existence of Δa(n), Δf and always existing. These three can reflect the individual differences of radio frequency devices. Use a broadband receiver to separately collect the wireless signals of each device that needs to be identified. Assuming there is no signal blind separation problem, since the wireless channel will affect the transmitted signal, the baseband IQ (In-phase and Quadrature-phase, abbreviated as IQ) signal obtained after being processed by the broadband receiver is shown as formula (3):

[0014] r(n) = h(n) * s(n) + n0(n) (3),

[0015] In formula (3), * represents the convolution operation, r(n) is the signal collected by the wideband receiver, h(n) is the channel impulse response, n0(n) is the additive noise, and h(n) and n0(n) together constitute the channel characteristics;

[0016] 2) Time-domain training sequence extraction: It includes:

[0017] 2-1) Signal frame synchronization: In order to find the starting position of a frame of signal in the collected signal, the delay correlation frame synchronization algorithm is used to perform signal frame synchronization on the collected signal r(n) in step 1). The delay correlation algorithm is shown in formulas (4), (5), and (6):

[0018]

[0019]

[0020]

[0021] Among them, L represents the delay amount, r * (n) represents the conjugate signal of r(n), C(n) is the delay autocorrelation function, P(n) is the power function of the received signal, M(n) is the normalized value for decision statistics, and the starting position of the frame can be found by finding the maximum value of M(n);

[0022] 2-2) Time-domain training sequence extraction: In a time-domain frame of IEEE802.11a signal, the long training sequence contained can be used for channel estimation. After signal frame synchronization, the long training sequence is extracted from the collected signal r(n) in step 1) and denoted as X0(n);

[0023] 3) Training the LMS filter: Complete the iterative training of the LMS filter according to the long training sequence X0(n) of the IEEE802.11a signal frame header extracted in step 2). The iterative process of the LMS filter is as follows:

[0024] 3-1) Initializing the filter coefficients: Initialize the filter coefficients according to formula (7): W(n) = 0 (7),

[0025] 3-2) Updating the tap weights: Update the tap weights according to formulas (8), (9), and (10):

[0026] y(n) = W T (n)X0(n) (8),

[0027] e(n) = d(n) - y(n) = d(n) - W T (n)X0(n) (9),

[0028] W(n + 1) = W(n) + 2μe(n)X0(n) (10),

[0029] W(n) = [w0(n), w1(n),..., w N-1 (n)] T , W(n) is the tap weight vector of the LMS adaptive filter, N represents the order of the filter, x0(n) is the long training sequence extracted in step 2-2), X0(n) = [x0(n), x0(n - 1),..., x0(n - N + 1)] T ,

[0030] d(n) is the long training sequence specified by the IEEE802.11a standard without channel interference, d(n) = [d0(n), d0(n - 1),..., d0(n - N + 1)] T ; e(n) is the error signal, μ is the iteration step size;

[0031] 3-3) Repeat step 3-2) until the number of iterations is reached;

[0032] 4) Channel equalization processing: Perform channel equalization processing on the collected signal r(n) to be recognized (n) to perform channel equalization processing, and the formula (11) holds:

[0033]

[0034] where * represents convolution calculation, s(n) is the baseband signal sequence actually transmitted by the RF device, h(n) is the discrete-time impulse response of the transmission channel, n0(n) is the Gaussian noise in the channel, r(n) is the signal sequence collected by the broadband receiver after passing through the channel transmission, and W(n) is the LMS filter trained in step 3), is the filter output signal, that is, the signal after equalization recovery;

[0035] 5) Train the neural network: including:

[0036] 5-1) Dataset production: Process each signal of the IEEE802.11a device to be recognized collected through step 4) to obtain multiple signals after equalization recovery and make them into a dataset: Use the method in step 2-1) to find the starting position of each signal frame in, then extract the first 256 points of each signal frame, and then normalize the energy of each signal segment. Since the collected signal is divided into two channels of IQ, splice the IQ two-channel signals into a two-dimensional signal segment of 256×2 size, and finally classify and package the signal segments according to the corresponding IEEE802.11a device to be recognized to make a dataset;

[0037] 5-2) Training the neural network: Use the dataset obtained in step 5-1) to train the neural network. The neural network adopts the IQCNet model. The gradient descent optimizer for training the model is the Adam optimizer, and the loss function uses the cross-entropy loss function. After training is completed, save the model. Among them, the IQCNet model consists of an input layer, a feature extraction layer, and a classification layer. The input layer receives data input of 256×2 size. The feature extraction layer consists of 6 convolutional layers, 2 max-pooling layers, 6 BN layers, and 1 adaptive average pooling layer. The first convolutional layer uses a 1×2 convolutional kernel with a stride of (1,1), and the subsequent 5 convolutional layers use 3×1 convolutional kernels with a stride of (1,1). A 4×1 max-pooling layer is added after every 2 convolutional layers, and a BN layer is inserted after each convolutional layer. The last layer of the feature extraction layer is the adaptive average pooling layer. The classification layer uses a fully connected layer for classification;

[0038] 6) Identity recognition: The broadband receiver collects the radio frequency signals of the IEEE802.11a device to be recognized, performs channel equalization processing using the method in step 4), and then obtains a two-dimensional signal segment of 256×2 size using the method in step 5-1). Input the processed signal segment into the IQCNet model trained in step 5). After forward derivation and operation of the IQCNet model, the recognition result can be obtained.

[0039] Compared with the existing radio frequency fingerprint recognition technology for IEEE802.11a devices, in the existing radio frequency fingerprint recognition technology, the problem that the radio frequency fingerprint extraction using a convolutional neural network is easily interfered by channel characteristics and the recognition accuracy drops sharply is solved in this technical solution by combining the radio frequency fingerprint recognition method based on a convolutional neural network with an adaptive channel equalization technology. First, use the LMS adaptive filter to eliminate the channel characteristic information in the signal to be recognized, and then use the IQCNet model to extract the radio frequency fingerprint information. It can accurately identify the identity of the IEEE802.11a signal radiation source in different channel environments, with the advantages of strong anti-interference ability, high recognition rate, and low implementation difficulty. It solves the problem that the recognition rate in the existing radio frequency fingerprint recognition technology is easily interfered by channel characteristics and can be applied to non-cooperative reception fields such as wireless network security and information confrontation.

[0040] This method has the advantages of strong anti-interference ability, high recognition rate, and low implementation difficulty. Brief description of the drawings

[0041] Figure 1 It is a schematic flowchart of the method in the embodiment;

[0042] Figure 2 It is a schematic diagram of the time-domain frame structure of the IEEE802.11a signal in the embodiment;

[0043] Figure 3Schematic diagram of the IQCNet model structure in the embodiment. Detailed implementation manners

[0044] The following further elaborates on the content of the present invention in conjunction with the accompanying drawings and embodiments, but does not limit the present invention.

[0045] Embodiment:

[0046] Referring to Figure 1 , a method for identifying the radiation source of IEEE802.11a signals by removing channel characteristics includes the following steps:

[0047] 1) Collect the signals of IEEE802.11a devices to be identified: IEEE802.11a adopts the standard OFDM technology. For the time-domain signal sequence of the OFDM signal model, it is shown in formula (1):

[0048]

[0049] Among them, X(k) is the frequency-domain data on the kth subcarrier in OFDM, N represents the total number of N subcarriers in the OFDM signal. In the RF circuit of the RF device, there are non-ideal manufacturing errors in the devices, and the circuit traces are different, which will cause slight differences in the circuit parameters of different RF devices. This slight difference will be attached to the transmitted signal in the form of unintentional modulation, resulting in the actual transmitted baseband signal model of the RF device as shown in formula (2):

[0050]

[0051] Among them, Δa(n) represents the unintentional modulation function of the signal envelope, represents the signal phase error, Δf represents the signal frequency offset. For the RF device, there will always be a deviation between the actual transmitted signal and the ideal modulation signal, resulting in the existence of Δa(n), Δf and always existing. These three can reflect the individual differences of the RF device. Use a broadband receiver to separately collect the wireless signals of each device that needs to be identified. Assuming there is no signal blind separation problem, since the wireless channel will affect the transmitted signal, the baseband IQ signal obtained after being processed by the broadband receiver is shown in formula (3):

[0052] r(n) = h(n) * s(n) + n0(n) (3),

[0053] In formula (3), * represents the convolution operation, r(n) is the signal collected by the broadband receiver, h(n) is the channel impulse response, n0(n) is the additive noise, and h(n) and n0(n) together constitute the channel characteristics;

[0054] 2) Extraction of time-domain training sequence: including:

[0055] 2-1) Signal frame synchronization: To find the starting position of a frame of signal in the acquired signal, the delay correlation frame synchronization algorithm is used to perform signal frame synchronization on the acquired signal r(n) in step 1). The delay correlation algorithm is shown in formulas (4), (5), and (6):

[0056]

[0057]

[0058]

[0059] Among them, L represents the delay amount, r * (n) represents the conjugate signal of r(n), C(n) is the delay autocorrelation function, P(n) is the power function of the received signal, M(n) is the normalized value for decision statistics, and the starting position of the frame can be found by finding the maximum value of M(n);

[0060] 2-2) Time-domain training sequence extraction: As Figure 2 shown, in an IEEE802.11a time-domain frame, the long training sequences T1 and T2 can be used for channel estimation. After signal frame synchronization, T1 and T2 are extracted from the acquired signal r(n) in step 1) and denoted as X0(n);

[0061] 3) Training the LMS filter: Complete the iterative training of the LMS filter based on the IEEE802.11a signal frame header long training sequence X0(n) extracted in step 2). The iterative process of the LMS filter is as follows:

[0062] 3-1) Initializing the filter coefficients: Initialize the filter coefficients according to formula (7): W(n) = 0 (7),

[0063] 3-2) Updating the tap weights: Update the tap weights according to formulas (8), (9), and (10):

[0064] y(n) = W T (n)X0(n) (8),

[0065] e(n) = d(n) - y(n) = d(n) - W T (n)X0(n) (9),

[0066] W(n + 1) = W(n) + 2μe(n)X0(n) (10),

[0067] W(n) = [w0(n), w1(n),..., w N-1 (n)] T, \(W(n)\) is the tap weight vector of the LMS adaptive filter, \(N\) represents the order of the filter, \(X_0(n)\) is the long training sequence extracted in step 2-2), and \(X_0(n)=[x_0(n),x_0(n - 1),\cdots,x_0(n - N + 1)]\). T ,

[0068] \(d(n)\) is the long training sequence specified by the IEEE802.11a standard without channel interference, and \(d(n)=[d_0(n),d_0(n - 1),\cdots,d_0(n - N + 1)]\). T ; \(e(n)\) is the error signal, and \(\mu\) is the iteration step size.

[0069] 3-3) Repeat step 3-2) until the number of iterations is reached. In this example, the filter order \(N\) is set to 64, the iteration step size \(\mu\) is set to 0.01, and the number of iterations is set to 100.

[0070] 4) Channel equalization processing: Perform channel equalization processing on the collected signal \(r(n)\) to be recognized. There is an equation (11) established:

[0071]

[0072] where * represents convolution calculation, \(s(n)\) is the baseband signal sequence actually transmitted by the RF device, \(h(n)\) is the discrete-time impulse response of the transmission channel, \(n_0(n)\) is the Gaussian noise in the channel, \(r(n)\) is the signal sequence collected by the broadband receiver after passing through the channel transmission, and \(W(n)\) is the LMS filter trained in step 3). is the filter output signal, that is, the signal after equalization and recovery.

[0073] 5) Train the neural network, including:

[0074] 5-1) Dataset production: Process each signal of the IEEE802.11a device to be recognized collected in step 4) to obtain multiple signals after equalization and recovery. and make them into a dataset: Use the method in step 2-1) to find the starting position of each signal frame in, then extract the first 256 points of each signal frame, and then normalize the energy of each signal segment. Since the collected signal is divided into two channels of I and Q, splice the I and Q channel signals into a two-dimensional signal segment of \(256\times2\) size. Finally, classify and package the signal segments according to the corresponding IEEE802.11a device to be recognized to make a dataset. In this example, a total of \(256\times2\) size two-dimensional signal segments are extracted from 6000 signal frames to make a dataset, including 5000 in the training set and 1000 in the test set.

[0075] 5-2) Training the neural network: Use the dataset obtained in step 5-1) to train the neural network. The neural network adopts the IQCNet model. The gradient descent optimizer for training the model is the Adam optimizer, and the loss function uses the cross-entropy loss function. After training is completed, save the model. Among them, the IQCNet model consists of an input layer, a feature extraction layer, and a classification layer. The input layer receives data input of 256×2 size. The feature extraction layer consists of 6 convolutional layers, 2 max pooling layers, 6 BN layers, and 1 adaptive average pooling layer. The first convolutional layer uses a 1×2 convolutional kernel with a stride of (1,1), and the subsequent 5 convolutional layers use 3×1 convolutional kernels with a stride of (1,1). A 4×1 max pooling layer is added after every 2 convolutional layers, and a BN layer is inserted after each convolutional layer. The last layer of the feature extraction layer is the adaptive average pooling layer. The classification layer uses a fully connected layer for classification. In this example, the structure of the IQCNet model is as Figure 3 shown, set the model learning rate to 0.001, and train the model for 60 rounds;

[0076] 6) Identity recognition: The broadband receiver collects the radio frequency signals of the IEEE802.11a device to be recognized, performs channel equalization processing using the method in step 4), and then obtains a two-dimensional signal segment of 256×2 size using the method in step 5-1). Input the processed signal segment into the IQCNet model trained in step 5). After the forward derivation operation of the IQCNet model, the recognition result can be obtained. In this example, the recognition accuracy rate of 6 IEEE802.11a devices using the method in this example reaches 96%.

Claims

1. A method for identifying the radiation source of IEEE802.11a signals with channel characteristics removed, characterized in that, The steps are as follows: 1) Collect the signals of IEEE802.11a devices to be recognized: IEEE802.11a adopts the standard OFDM technology. For the time-domain signal sequence of the OFDM signal model, it is shown as formula (1): Where X(k) is the frequency-domain data on the k-th subcarrier in OFDM, and N represents that there are N subcarriers in the OFDM signal. In the radio-frequency circuit of the radio-frequency device, the devices have non-ideal manufacturing errors, and the circuit traces are different, which will cause slight differences in the circuit parameters of different radio-frequency devices. This slight difference will be added to the transmitted signal in the form of unintentional modulation, resulting in the actual transmitted baseband signal model of the radio-frequency device as shown in formula (2): Among them, Δa(n) represents the unintentional modulation function of the signal envelope, represents the signal phase error, Δf represents the frequency offset of the signal. For a radio frequency device, there is a deviation between the actual transmitted signal and the ideal modulation signal, resulting in the existence of Δa(n), Δf, and always existing. A broadband receiver is used to separately collect the wireless signals of each device that needs to be identified. Assuming there is no problem of signal blind separation, the baseband IQ signal obtained after processing by the broadband receiver is as shown in formula (3): r(n) = h(n) * s(n) + n0(n) (3), In formula (3), * represents the convolution operation, r(n) is the signal collected by the wideband receiver, h(n) is the channel impulse response, n0(n) is the additive noise, and h(n) and n0(n) together constitute the channel characteristics; 2) Extract the time-domain training sequence: including: 2-1) Signal frame synchronization: Use the delay correlation frame synchronization algorithm to perform signal frame synchronization on the collected signal r(n) in step 1). The delay correlation algorithm is shown as formulas (4), (5), and (6): where L represents the delay amount, r * (n) represents the conjugate signal of r(n), C(n) is the delayed autocorrelation function, P(n) is the power function of the received signal, M(n) is the normalized value for decision statistics, and the starting position of the frame is found by finding the maximum value of M(n); 2-2) Extract the time-domain training sequence: In a time-domain frame of the IEEE802.11a signal, the included long training sequence is used for channel estimation. After signal frame synchronization, the long training sequence is extracted from the collected signal r(n) in step 1), denoted as X0(n); 3) Train the LMS filter: Complete the iterative training of the LMS filter according to the long training sequence X0(n) of the IEEE802.11a signal frame header extracted in step 2). The iterative process of the LMS filter is as follows: 3-1) Initialize the filter coefficients: Initialize the filter coefficients according to formula (7): W(n) = 0 (7), 3-2) Update the tap weights: Update the tap weights according to formulas (8), (9), and (10): y(n) = W T (n)X0(n) (8), e(n) = d(n) - y(n) = d(n) - W T (n)X0(n) (9), W(n + 1) = W(n) + 2μe(n)X0(n) (10), W(n) = [w0(n), w1(n),..., w N-1 (n)] T , where W(n) is the tap weight vector of the LMS adaptive filter, N represents the order of the filter, X0(n) is the long training sequence extracted in step 2-2), and X0(n) = [x0(n), x0(n-1),..., x0(n-N+1)] T , d(n) is the long training sequence without channel interference specified by the IEEE802.11a standard, and d(n) = [d0(n), d0(n - 1),..., d0(n - N + 1)] T ; e(n) is the error signal, and μ is the iteration step size; 3-3) Repeat step 3-2) until the number of iterations is reached; 4) Perform channel equalization processing: Perform channel equalization processing on the collected signal r(n) to be recognized (n) to perform channel equalization processing, and formula (11) holds: where * represents convolution calculation, s(n) is the baseband signal sequence actually transmitted by the RF device, h(n) is the discrete-time impulse response of the transmission channel, n0(n) is the Gaussian noise in the channel, r(n) is the signal sequence collected by the broadband receiver after transmission through the channel, and W(n) is the LMS filter trained in step 3), which is the filter output signal, i.e., the signal after equalization and recovery; 5) Train the neural network: including: 5-1) Dataset production: Each IEEE802.11a device signal to be recognized collected is processed through step 4) to obtain multiple equalized and restored signals and a dataset is produced: Using the method in step 2-1), find the starting position of each signal frame in, then extract the first 256 points of each signal frame, and then perform energy normalization on each signal segment. Since the collected signal is divided into two channels of IQ, splice the IQ two-channel signals into a two-dimensional signal segment of 256×2 size. Finally, classify and package the signal segments according to the corresponding IEEE802.11a device to be recognized to form a dataset; 5-2) Training the neural network: Use the dataset obtained in step 5-1) to train the neural network. The neural network adopts the IQCNet model. The gradient descent optimizer for training the model is the Adam optimizer, and the loss function uses the cross-entropy loss function. After training is completed, save the model. Among them, the IQCNet model consists of an input layer, a feature extraction layer, and a classification layer. The input layer receives data input of 256×2 size. The feature extraction layer consists of 6 convolutional layers, 2 max pooling layers, 6 BN layers, and 1 adaptive average pooling layer. The first convolutional layer uses a 1×2 convolutional kernel with a stride of (1,1), and the subsequent 5 convolutional layers use 3×1 convolutional kernels with a stride of (1,1). A 4×1 max pooling layer is added after every 2 convolutional layers, and a BN layer is inserted after each convolutional layer. The last layer of the feature extraction layer is the adaptive average pooling layer. The classification layer uses a fully connected layer for classification; 6) Identity recognition: The broadband receiver collects the radio frequency signal of the IEEE 802.11a device to be recognized, performs channel equalization processing using the method in step 4), and then obtains a two-dimensional signal segment of 256×2 size using the method in step 5-1). Input the processed signal segment into the IQCNet model trained in step 5). After the forward derivation operation of the IQCNet model, the recognition result can be obtained.

Images