Open Set WIFI Device Identification Method and Device

Abstract

The present invention discloses a method and apparatus for identifying open-set Wi-Fi devices, comprising: obtaining the output signal of an authorized, legitimate Wi-Fi device, preprocessing it and extracting a preamble; extracting artificial features and deep features based on the preamble; and concatenating the two types of features to obtain a fused feature vector; training a BP neural network model with a custom loss function as a discriminator; obtaining the output signal of the Wi-Fi device to be identified, querying whether a corresponding discriminator exists; if so, extracting the fused feature vector of the output signal and inputting it into the discriminator for Wi-Fi device identification; if not, deeming the Wi-Fi device an unknown device. The present invention integrates artificial and deep features, enriching the feature set and improving the accuracy of Wi-Fi device identification. Furthermore, a BP neural network with single-classification capability is introduced as a discriminator to reject any unknown Wi-Fi device disguised as a legitimate Wi-Fi device without prior information.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and information security, and in particular to an open set WIFI device identification method and device. Background Art

[0002] As the number of wirelessly connected devices grows, securing wireless communications becomes increasingly challenging. Identity authentication is a crucial aspect of wireless security. While many cryptographic-based authentication methods exist, many devices are limited by computational and power budgets, making them unsuitable for highly complex security algorithms or additional security modules.

[0003] Physical layer authentication enhances wireless communication security by combining channel state information and transmitter hardware fingerprints to authenticate devices. Transmitter fingerprints are caused by the non-idealities of their radio frequency components. The interaction between these non-idealities causes signals from different transmitters to exhibit unique characteristics. Common RF fingerprints include carrier frequency offset.

[0004] While many device identification methods based on RF fingerprinting have been able to classify and identify different devices and perform well, these approaches have a limitation when it comes to identifying unknown disguised devices. Most of the identification algorithms used are supervised learning algorithms, requiring labeled samples with prior information for model training. In applications, it is impossible to obtain the fingerprint characteristics of unknown disguised devices in advance. If any unknown disguised device is identified without prior information, its signal will be misclassified into the closest category, leading to security vulnerabilities. Summary of the Invention

[0005] Purpose of the Invention: To address the problems existing in the prior art, the present invention provides an open set WIFI device identification method and apparatus. While ensuring high accuracy in classifying and identifying authorized and legitimate WIFI devices, the method and apparatus can also reject any unknown devices disguised as legitimate WIFI devices, thereby achieving open set identification. Furthermore, the method and apparatus utilize a fusion of artificial and deep features for classification and identification, effectively improving the recognition accuracy and reducing the misjudgment rate caused by using either artificial or deep features alone.

[0006] Technical solution: On the one hand, the present invention provides an open set WIFI device identification method, comprising the following steps:

[0007] S1: Obtain the output signal of an authorized and legal Wi-Fi device, preprocess the output signal and extract the preamble, extract the manual features and deep features of the output signal based on the preamble, concatenate the two types of features to obtain a fused feature vector, and create a fused feature vector set of preset size from the fused feature vector.

[0008] S2: Dividing the fused feature vector set into a training set, a validation set, and a test set, training a BP neural network model with a custom loss function based on the training set, adjusting the parameters of the BP neural network model using the validation set, selecting the BP neural network model that meets the first preset condition as a discriminator, and setting a label corresponding to the legitimate WIFI device for the discriminator;

[0009] S3: Obtain the output signal of the WIFI device to be identified, and check whether there is a corresponding discriminator. If so, extract the fused feature vector of the output signal and input it into the discriminator to identify the WIFI device. If not, the WIFI device is considered to be an unknown device.

[0010] Furthermore, the process of extracting the preamble code from the output signal includes: presetting a standard preamble code sequence, intercepting a continuous sequence of the same length as the standard preamble code sequence from the output signal, calculating the correlation between the intercepted signal and the standard preamble code sequence by conjugate multiplication, and extracting the intercepted signal whose correlation meets a second preset condition as the preamble code.

[0011] Furthermore, the artificial features include: carrier frequency offset, channel estimation features, and singular values based on frequency response.

[0012] Furthermore, the specific process of extracting the depth feature is:

[0013] Pre-training the autoencoder model based on the preamble set: The autoencoder network consists of two parts: an encoder layer and a decoder layer. During training, the encoder layer uses a nonlinear activation function to learn a compressed representation of the input data; the decoder layer reconstructs the original input data and calculates the reconstruction error. The autoencoder network uses a backpropagation algorithm and optimization method to select features with the information content that meets the third preset condition to form a compressed representation of the input data;

[0014] Copy the encoder layer of a pre-trained autoencoder model and set it as a deep feature extractor, using the compressed representation of the input data as deep features.

[0015] Furthermore, the specific process of S2 is:

[0016] S2.1: Use the non-repeated sampling technique to divide the fused feature vector set of each legitimate Wi-Fi device into a training set, a validation set, and a test set in proportion;

[0017] S2.2: Using random sampling without replacement, extract a preset number of samples from the training set to form a training subset, and repeat the sampling a preset number of times to obtain a training subset, where the number of training subsets is the same as the preset number of sampling times;

[0018] S2.3: Parallel training of BP neural network models with customized loss functions based on the training subsets, where the number of BP neural network models is the same as the number of training subsets;

[0019] S2.4: Adjust parameters and verify performance of the BP neural network model based on the validation set, save the BP neural network model whose prediction error meets the first preset condition as a discriminator, and set a label for the discriminator.

[0020] Furthermore, the BP neural network model uses the optimization target of the single classification algorithm as the loss function.

[0021] Furthermore, the BP neural network model consists of an input layer, a hidden layer and an output layer. The specific process of training the BP neural network model includes: customizing the loss function and defining the optimization target of the single-classification algorithm as the BP neural network loss function; calculating the prediction error based on the custom loss function, and iteratively updating the BP neural network parameters using the back propagation algorithm until the convergence requirements are met; and using the decision function in the output layer to mark the training sample points as normal or abnormal, so that the BP neural network has single-classification capabilities.

[0022] Furthermore, the specific process of S3 is as follows:

[0023] S3.1: Obtain the output signal of the WIFI device to be identified;

[0024] S3.2: Extract tag information from the output signal of the WIFI device to be identified, and select a corresponding discriminator based on the tag information. If there is no corresponding discriminator for the tag information, the WIFI device to be identified is an unknown device and is not disguised as an authorized and legitimate device;

[0025] S3.3: If the label information has a corresponding discriminator, extract the fused feature vector from the output signal of the WiFi device to be identified and use it as the input of the discriminator;

[0026] S3.4: The discriminator outputs a discrimination result. The first discrimination result is a legitimate Wi-Fi device, and the second discrimination result is an unknown illegal Wi-Fi device disguised as a legitimate Wi-Fi device.

[0027] On the other hand, the present invention also provides an open set WIFI device identification device, comprising the following units:

[0028] A signal processing unit is used to obtain the output signal of an authorized and legal Wi-Fi device, pre-process the output signal and extract the preamble, extract the manual features and deep features of the output signal based on the preamble, concatenate the two types of features to obtain a fused feature vector, and compile the fused feature vector into a fused feature vector set of a preset size;

[0029] a model training unit, configured to divide the fused feature vector set into a training set, a validation set, and a test set, train a BP neural network model with a custom loss function based on the training set, adjust parameters of the BP neural network model using the validation set, select a BP neural network model that meets a first preset condition as a discriminator, and set a label corresponding to the legitimate WIFI device for the discriminator;

[0030] The device identification unit obtains the output signal of the WIFI device to be identified and queries whether a corresponding discriminator exists. If so, the fused feature vector of the output signal is extracted and input into the discriminator for WIFI device identification. If not, the WIFI device is considered to be an unknown device.

[0031] Existing technologies typically use supervised learning algorithms, which require labeled samples with prior information for model training. This invention combines artificial features with deep features, enriching the feature set and improving the accuracy of Wi-Fi device recognition. It also introduces a BP neural network with single-classification capabilities as a discriminator. While ensuring high-accuracy classification and identification of authorized and legitimate Wi-Fi devices, it can also reject any unknown Wi-Fi devices disguised as legitimate Wi-Fi devices without prior information, thus achieving open-set Wi-Fi device recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flowchart of the steps of the open set WIFI device identification method in Example 1 of the present invention;

[0033] Figure 2 This is a working framework diagram of the open set WIFI device identification method in Example 1 of the present invention;

[0034] Figure 3 Schematic diagram of the BP neural network topology structure in Example 1 of the present invention;

[0035] Figure 4 This is a flow chart of the BP neural network training process in Example 1 of the present invention;

[0036] Figure 5 This is a schematic diagram of the structure of the open set WIFI device identification device in Example 2 of the present invention. DETAILED DESCRIPTION

[0037] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings in the embodiments of the present invention, and the solution of the present invention will be better described through embodiments.

[0038] Example 1

[0039] Reference Figure 1 This embodiment provides an open set WIFI device identification method, the specific steps are as follows:

[0040] S1: Obtain the output signal of an authorized and legal Wi-Fi device, preprocess the output signal and extract the preamble, extract the manual features and deep features of the output signal based on the preamble, concatenate the two types of features to obtain a fused feature vector, and create a fused feature vector set of preset size from the fused feature vector.

[0041] Specifically, in this embodiment, 10 Wi-Fi devices are selected as target wireless devices and numbered. Wi-Fi devices numbered 1 to 8 are used as authorized and legitimate Wi-Fi devices for training the discriminator; Wi-Fi devices numbered 9 to 10 are used as unknown devices disguised as legitimate Wi-Fi devices for testing the rejection rate of the discriminator for unknown disguised devices.

[0042] The universal software radio peripheral (USRP) is used to collect the output signals of authorized and legal Wi-Fi devices. 10,000 frames of data are collected for each device for model training. Signals from different Wi-Fi devices are collected in different time periods to avoid mutual interference between different Wi-Fi devices.

[0043] The collected output signal is preprocessed, and the preprocessing process includes: down-conversion, oversampling, signal detection and interception, energy normalization, frequency offset estimation and compensation. Among them, the collected wireless signal is directly down-converted to obtain the baseband signal: x(t) represents the complex baseband representation of the transmitted signal, Δf represents the carrier frequency offset, Indicates phase deviation; after down-conversion, the complex baseband signal is sampled at a 20Mbps sampling rate, and then the start and end of each frame of the signal are estimated using the change point principle, and the signal is intercepted; finally, the intercepted signal is energy normalized and stored.

[0044] The preamble code is extracted from the preprocessed signal using a preset standard preamble code sequence and a sliding window method: this embodiment uses a 20Mbps sampling rate and a short preamble code length of 160. A sliding window is used to intercept a continuous sequence of 160 in length from the preprocessed signal, which is conjugated and multiplied with the preset standard short preamble code to calculate the correlation. When the second preset condition is met, the intercepted signal is the preamble code. The second preset condition referred to in this embodiment is that when the correlation is maximum, the sliding window subscript is the starting point of the preamble code.

[0045] Three artificial features are extracted based on the preamble complex signal: carrier frequency offset, channel estimation features, and singular values based on frequency response. The specific calculation method is as follows:

[0046] Calculate the carrier frequency offset: The PLCP preamble consists of a short preamble and a long preamble. The short preamble consists of 10 repeated short training sequences, and the long preamble consists of 2 repeated long training sequences. The frequency offset is: Where N is the length of the preamble sequence, D is the distance between the two repeated sequences, arg() is the angle calculation function, and R is the delay correlation of the repeated sequence: L is the length of the repeated sequence, and r represents the preamble sequence. In this embodiment, the short preamble is first used to perform a coarse frequency offset estimate, and the average value of 10 segments of the repeated sequence is used to obtain the coarse frequency offset estimate f1; the long preamble is compensated: y(t) = x(t)e -j2πf1t , the precise frequency offset estimate f2 is obtained using the repeated sequence of the compensated long preamble code, and the total frequency offset estimate is f1+f2.

[0047] Calculate the channel estimation characteristics: At a 20Mbps sampling rate, the long preamble contains two 64-bit repeating sequences. Calculate the preset standard long preamble repeating sequence frequency response H1 and the actual received long preamble repeating sequence frequency responses H2 and H3 to obtain the channel estimation characteristics:

[0048] Calculate the singular values based on the frequency response: Calculate the frequency response H of the long preamble code, and calculate the singular values: s=svd(H), where svd() is a singular value calculation function in Matlab.

[0049] Extract deep features based on the preamble complex signal using the encoder layer of the pretrained autoencoder network:

[0050] An autoencoder model is pre-trained for each legitimate Wi-Fi device based on the preamble set. In this embodiment, a preamble complex signal of length 320 is used as the autoencoder input to pre-train the autoencoder network. The encoder layer of the autoencoder network consists of three hidden layers, using a sigmoid activation function, with dimensions of 160, 80, and 10, respectively. This layer learns a compressed representation of the input data. Using a backpropagation algorithm and optimization methods, the autoencoder network is forced to select features that meet a third pre-condition as the compressed representation of the input data. In this embodiment, the third pre-condition is to select the feature with the highest information content as the compressed representation of the input data. The number of encoder layers and their dimensions are determined based on actual needs and are not limited thereto.

[0051] The encoder layer of the pre-trained autoencoder network is copied as a deep feature extractor. In this embodiment, the preamble complex signal with a length of 320 is compressed into a real signal with a length of 10 as a deep feature.

[0052] The three extracted artificial features and deep features are spliced into a one-dimensional feature vector to form a fused feature vector, and the fused feature vectors are formed into a fused feature vector set. In this embodiment, the size of the fused feature vector set of each legal WIFI device is 10,000.

[0053] S2: Dividing the fused feature vector set into a training set, a validation set, and a test set, training a BP neural network model with a custom loss function based on the training set, adjusting the parameters of the BP neural network model using the validation set, selecting the BP neural network model that meets the first preset condition as a discriminator, and setting a label corresponding to the legitimate WIFI device for the discriminator;

[0054] S2.1: Use the non-repeated random sampling technique to divide the fused feature vector set of each legal Wi-Fi device, which is 10,000 in size, into training, validation, and test sets in a ratio of 7:2:1. The non-repeated random sampling technique is implemented using the train_test_split auxiliary function in scikit-learn.

[0055] S2.2: The training set size is 7000. Use random sampling without replacement to extract 5000 samples from the training set to form a training subset. Repeat 10 times to obtain 10 training subsets. The number of samples extracted and the number of training subsets are determined according to actual needs and are not limited to this.

[0056] S2.3: Based on 10 training subsets, 10 BP neural network models with customized loss functions are trained in parallel: the BP neural network topology is as follows: Figure 3 As shown in Figure 1, it consists of an input layer, a hidden layer, and an output layer. The number of neurons in the output layer is 1, and the decision function is used to mark the training sample points as normal or abnormal.

[0057] Reference Figure 4 , the BP neural network training process is as follows:

[0058] S2.3.1: Input the fused feature vector training subset;

[0059] S2.3.2: Network initialization: Determine the BP neural network topology, initialize the weights between each layer and the threshold of each neuron;

[0060] S2.3.3: Compute hidden layer outputs based on connection weights and input fusion feature vectors;

[0061] S2.3.4: Calculate the output layer prediction output based on the connection weights and hidden layer outputs;

[0062] S2.3.5: Use a custom loss function to calculate the error between the target output and the predicted output. The custom loss function is designed based on the OC-SVM single-class optimization target.

[0063] Specifically, the OC-SVM algorithm works similarly to the SVM algorithm, training a support vector machine by using zero as a negative sample and all other data as positive samples. The strategy is to map the data into a feature space corresponding to the kernel and construct a hyperplane between the data and the origin, which has the maximum distance from the origin.

[0064] Assume the hyperplane is: w T Φ(X)-ρ=0, where w is the norm perpendicular to the hyperplane, Φ(X) is the RKHS mapping function from the input space to the feature space F, and ρ is the deviation of the hyperplane. The goal is to maximize the distance between the hyperplane and the origin while ensuring correct classification. After adding the slack variable ξ, the OC-SVM algorithm optimizes the objective function as follows:

[0065]

[0066] st(w T Φ(X i ))≥ρ-ξ i ,ξ i ≥0

[0067] Among them, v∈(0,1) is used to adjust the degree of relaxation, and N is the sample dimension.

[0068] Apply OC-SVM to BP neural network to drive the training of neural network. The most important thing is to use the transformation w T g(VX n ) instead of w in the OC-SVM algorithm T Φ(X n ) transformation, where w is the scalar output obtained from the hidden layer to the output layer, g() represents the activation function, and V represents the weight matrix from the network input to the hidden unit, so that the optimization goal can be achieved in the network. Finally, the loss function of the network is:

[0069]

[0070] st(w T g(VX i )≥ρ-ξ i ,ξ i ≥0

[0071] S2.3.6: Use the backpropagation algorithm to calculate the output error of each layer and iteratively update the connection weights between layers until the convergence requirements are met, so that the network has single classification capabilities.

[0072] S2.4: Based on the validation set, the parameters of the 10 trained BP neural network models are adjusted and the performance is verified. The BP neural network model that meets the first preset condition is saved as the discriminator. The first preset condition referred to in this embodiment is to select the BP neural network model with the smallest prediction error as the discriminator, and set the MAC address of the corresponding WIFI device as the label of the discriminator.

[0073] S3: Obtain the output signal of the WIFI device to be identified, and check whether there is a corresponding discriminator. If so, extract the fused feature vector of the output signal and input it into the discriminator to identify the WIFI device. If not, the WIFI device is considered to be an unknown device.

[0074] S3.1: Obtain the output signal of the WIFI device to be identified;

[0075] S3.2: Extracting a MAC address from the output signal of the WIFI device to be identified, and selecting a corresponding discriminator based on the MAC address. If there is no corresponding discriminator for the MAC address, the WIFI device to be identified is an unknown device and is not disguised as an authorized legitimate device.

[0076] S3.3: If a corresponding discriminator exists for the MAC address, extract a fused feature vector from the output signal of the Wi-Fi device to be identified and use it as the input of the discriminator;

[0077] S3.4: The discriminator outputs the discrimination result. If the discriminator outputs a 1, it indicates that the device to be identified is a legitimate Wi-Fi device and belongs to the legal Wi-Fi device category corresponding to the discriminator. If the discriminator outputs a 0, it indicates that the Wi-Fi device to be identified is an unknown illegal Wi-Fi device disguised as a legitimate Wi-Fi device. Repeatedly use different legitimate Wi-Fi device samples (authorized Wi-Fi devices numbered 1 to 8) and unknown disguised Wi-Fi device samples (unknown Wi-Fi devices numbered 9 to 10) to verify the model's classification accuracy for legitimate Wi-Fi devices and its rejection rate for unknown disguised Wi-Fi devices.

[0078] This method achieves the fusion of artificial features and deep features, enriches the feature set, and further improves the recognition accuracy of Wi-Fi devices. In addition, the present invention introduces a BP neural network with a single classification function as a discriminator. On the basis of ensuring the classification and identification of authorized and legitimate Wi-Fi devices with high accuracy, it can also reject any unknown devices without prior information disguised as legitimate Wi-Fi devices, thus realizing the recognition of open-set Wi-Fi devices.

[0079] Example 2

[0080] Reference Figure 4This embodiment provides an open set WIFI device identification device, which includes the following units:

[0081] The signal processing unit 101 is configured to obtain the output signal of an authorized and legitimate Wi-Fi device, preprocess the output signal and extract the preamble, extract the artificial features and deep features of the output signal based on the preamble, concatenate the two types of features to obtain a fused feature vector, and compile the fused feature vector into a fused feature vector set of a preset size;

[0082] a model training unit 102, configured to divide the fused feature vector set into a training set, a validation set, and a test set, train a BP neural network model with a custom loss function based on the training set, adjust parameters of the BP neural network model using the validation set, select a BP neural network model that meets a first preset condition as a discriminator, and set a label corresponding to the legitimate WIFI device for the discriminator;

[0083] The device identification unit 103 is used to obtain the output signal of the WIFI device to be identified, query whether there is a corresponding discriminator, and if so, extract the fused feature vector of the output signal and input it into the discriminator for WIFI device identification; if not, the WIFI device is considered to be an unknown device.

[0084] In addition, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium may store a program, and when the program is executed, it includes some or all steps of any open set WIFI device identification method described in the above method embodiment.

[0085] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0086] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a memory and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned memory includes: various media that can store program codes, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0087] A person skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable memory, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0088] It should be noted that the above is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of equivalent changes or replacements within the technical scope disclosed by the present invention, which still fall within the scope covered by the present invention.

Claims

1. A method for identifying an open set of WIFI devices, characterized in that: The following steps are involved: S1: Obtain the output signal of an authorized and legal Wi-Fi device, preprocess the output signal and extract the preamble, extract the manual features and deep features of the output signal based on the preamble, concatenate the two types of features to obtain a fused feature vector, and create a fused feature vector set of preset size from the fused feature vector. S2.1: Use the non-repeated sampling technique to divide the fused feature vector set of each legitimate Wi-Fi device into a training set, a validation set, and a test set in proportion; S2.2: Using random sampling without replacement, extract a preset number of samples from the training set to form a training subset, and repeat the sampling a preset number of times to obtain a training subset, where the number of training subsets is the same as the preset number of sampling times; S2.3: Parallel training of BP neural network models with customized loss functions based on the training subsets, where the number of BP neural network models is the same as the number of training subsets; S2.4: Parameter adjustment and performance verification are performed on the BP neural network model based on the validation set. The BP neural network model whose prediction error satisfies the first preset condition is saved as a discriminator, and a label is set for the discriminator. The discriminator is a BP neural network with a single classification function, so that unknown Wi-Fi devices without prior information that are disguised as legitimate Wi-Fi devices are rejected. S3: Obtain the output signal of the WIFI device to be identified, and check whether there is a corresponding discriminator. If so, extract the fused feature vector of the output signal and input it into the discriminator to identify the WIFI device. If not, the WIFI device is considered to be an unknown device.

2. The open set WIFI device identification method according to claim 1, characterized in that: The process of pre-processing the output signal includes: down-conversion, over-sampling, signal detection and interception, energy normalization, and frequency offset estimation and compensation.

3. The open set WIFI device identification method according to claim 1, characterized in that: The process of extracting the preamble code from the output signal includes: presetting a standard preamble code sequence, intercepting a continuous sequence of the same length as the standard preamble code sequence from the output signal, calculating the correlation between the intercepted signal and the standard preamble code sequence by conjugate multiplication, and extracting the intercepted signal whose correlation meets a second preset condition as the preamble code.

4. The open set WIFI device identification method according to claim 1, characterized in that: The artificial features include: carrier frequency offset, channel estimation features and singular values based on frequency response.

5. The open set WIFI device identification method according to claim 1, characterized in that: The specific process of extracting depth features is: Pre-train the autoencoder model based on the preamble set: The autoencoder network consists of two parts: the encoder layer and the decoder layer. During pre-training, the encoder layer uses a nonlinear activation function to learn a compressed representation of the input data; The decoder layer reconstructs the original input data and calculates the reconstruction error. The autoencoder network selects features whose information content meets the third preset condition based on the backpropagation algorithm and optimization method to form a compressed representation of the input data; Copy the encoder layer of a pre-trained autoencoder model and set it as a deep feature extractor, using the compressed representation of the input data as deep features.

6. The open set WIFI device identification method according to claim 1, characterized in that: The BP neural network model uses the single classification algorithm optimization target as the loss function.

7. The open set WIFI device identification method according to claim 6, characterized in that: The BP neural network model consists of an input layer, a hidden layer, and an output layer. The specific process of training the BP neural network model includes: customizing the loss function and defining the single-classification algorithm optimization target as the BP neural network loss function; calculating the prediction error based on the custom loss function and iteratively updating the BP neural network parameters using the backpropagation algorithm until the convergence requirements are met; and using the decision function in the output layer to mark the training sample points as normal or abnormal, so that the BP neural network has the ability to classify single classes.

8. The open set WIFI device identification method according to claim 1, characterized in that: The specific process of S3 is: S3.1: Obtain the output signal of the WIFI device to be identified; S3.2: Extract tag information from the output signal of the WIFI device to be identified, and select a corresponding discriminator based on the tag information. If there is no corresponding discriminator for the tag information, the WIFI device to be identified is an unknown device and is not disguised as an authorized and legitimate device; S3.3: If the label information has a corresponding discriminator, extract the fused feature vector from the output signal of the WiFi device to be identified and use it as the input of the discriminator; S3.4: The discriminator outputs a discrimination result. The first discrimination result is a legitimate Wi-Fi device, and the second discrimination result is an unknown illegal Wi-Fi device disguised as a legitimate Wi-Fi device.

9. An open set WIFI device identification device, characterized in that, include: A signal processing unit is used to obtain the output signal of an authorized and legal Wi-Fi device, pre-process the output signal and extract the preamble, extract the manual features and deep features of the output signal based on the preamble, concatenate the two types of features to obtain a fused feature vector, and compile the fused feature vector into a fused feature vector set of a preset size; A model training unit is used to divide the fused feature vector set of each legitimate Wi-Fi device into a training set, a validation set, and a test set in proportion using a non-repeated sampling technique; A preset number of samples are extracted from the training set using a random sampling method without replacement to form a training subset, and the sampling is repeated a preset number of times to obtain a training subset, wherein the number of training subsets is the same as the preset number of samplings; a BP neural network model with a custom loss function is trained in parallel based on the training subset, wherein the number of BP neural network models is the same as the number of training subsets; parameter adjustment and performance verification of the BP neural network model are performed based on a validation set, and the BP neural network model whose prediction error meets a first preset condition is saved as a discriminator, and a label is set for the discriminator; A BP neural network with single-classification function is introduced as a discriminator to reject any unknown WIFI device without prior information that disguises itself as a legitimate WIFI device; The device identification unit obtains the output signal of the WIFI device to be identified and queries whether a corresponding discriminator exists. If so, the fused feature vector of the output signal is extracted and input into the discriminator for WIFI device identification. If not, the WIFI device is considered to be an unknown device.

10. A computer-readable storage medium, characterized in that The storage medium contains the open set WIFI device identification method according to any one of claims 1 to 8.

Images