Training method of WiFi device fingerprint authentication model based on channel confrontation at different positions

Abstract

The invention provides a training method of a WiFi equipment fingerprint authentication model based on confrontation of channels at different positions, and specifically, a constructed twin confrontation authentication model is used for carrying out training authentication on collected channel state information CSI data sets at different positions, so that the fingerprint identification capability under different channels is enhanced, and equipment fingerprints can be efficiently authenticated; the method is simple in implementation step, rapid in operation, low in cost, easy to implement and high in stability.

Description

Technical Field

[0001] The invention belongs to the technical field of information security, and in particular relates to a training method for a WiFi device fingerprint authentication model based on confrontation of channels at different locations. Background Art

[0002] Device fingerprint (Radio Frequency Fingerprint, RFF) technology is a technology that identifies and verifies devices by analyzing hardware features in wireless transmission signals, such as frequency error, clock offset, modulation error, etc. Since the hardware RF characteristics of each transmitter device vary due to manufacturing processes and environmental factors, a unique "fingerprint" is formed. Since RF fingerprints are unique and cannot be cloned, they can help identify potential attackers and unauthorized devices in wireless networks, thereby providing enhanced security in identity authentication and security certification.

[0003] In WiFi radio frequency fingerprint recognition technology, the channel state information (CSI) of the transmitter obtained by the receiver contains hardware fingerprint information and can be extracted for identity recognition and authentication. However, due to the different collection locations of training samples and test samples, CSI will change significantly, resulting in poor stability of the extracted radio frequency fingerprint. Specifically, factors such as the propagation environment, distance, obstacles, interference, etc. of the signal at different locations may cause signal attenuation, distortion and frequency shift, thereby affecting the stability and accuracy of the radio frequency fingerprint. The data in the training phase is usually collected at a specific location, but the data in the test phase may be obtained under different environmental conditions. This difference in location and environment will cause changes in the radio frequency characteristics, making it impossible for the identity authentication technology based on radio frequency fingerprint to achieve effective cross-domain recognition. Therefore, how to maintain the stability and consistency of radio frequency fingerprints under channel conditions at different locations has become an important challenge facing radio frequency fingerprint technology in practical applications.

[0004] In addition, although RF fingerprint recognition technology has a high potential for device identity authentication in theory, it still faces many challenges in practical applications, especially the lack of an effective authentication model. Most existing RF fingerprint recognition methods focus on the extraction of a single feature or simple pattern matching. Therefore, the lack of an authentication model that can adapt to various channel changes and has high robustness has led to insufficient accuracy and reliability of RF fingerprint technology in cross-domain applications. Therefore, how to design an RF fingerprint authentication model that can comprehensively consider environmental adaptability, robustness, and cross-domain generalization capabilities is still a key problem in the development of this technology.

[0005] Adversarial thinking originates from the fields of machine learning and artificial intelligence, and is widely used in adversarial training in particular. It mainly improves the robustness and generalization ability of the model by introducing "adversarial samples" or "adversarial attacks". Under the framework of adversarial thinking, the model not only needs to learn how to identify normal samples, but also needs to be able to resist specially designed disturbances or attacks to avoid the vulnerability of the model when facing malicious inputs or abnormal environments.

[0006] Therefore, it is of great significance to use the adversarial idea to build a neural network model for fingerprint extraction and device authentication based on WiFi CSI. This model can effectively cope with environmental changes and channel interference, improve the stability and robustness of RF fingerprints, and thus solve the limitations of traditional methods in cross-domain identification. Summary of the invention

[0007] Technical solution: In order to solve the above-mentioned technical problems, the present invention provides a WiFi device fingerprint authentication model and training method based on channel confrontation at different locations. Specifically, the constructed twin confrontation authentication model is used to train and authenticate the collected CSI data set of channel state information at different locations, thereby enhancing the fingerprint recognition capability under different channels and efficiently authenticating device fingerprints. The method has simple implementation steps, fast calculation, low cost, easy implementation and high stability.

[0008] The technical solution of a training method for a WiFi device fingerprint authentication model based on confrontation of channels at different locations adopted by the present invention is:

[0009] Step 1: Collect the channel state information (CSI) data of A and B at different positions and construct sample pairs S1 and S2;

[0010] Step 2: Input the data sample into the twin adversarial model;

[0011] Step 3: Based on the feature extractor G of the twin adversarial model f (x,θ f ), and obtain the distance vector D(f 1, f2); where f 1, f2 is the feature vector of different samples, x is the CSI of the input feature extractor, θ f are the parameters in the feature extractor;

[0012] Step 4: Add the distance vectors obtained in step 3 to the device prediction label network G y (x,θ y ) and the position channel prediction label network G c (x,θ c );θ y Parameters,θ,in the device label prediction networkc Predict parameters in the label network for the position channel;

[0013] Step 5: Add a gradient reversal layer GRL to update the network training so that the features it extracts are more focused on the hardware fingerprint characteristics of the device;

[0014] Step 6: Use the trained network model to authenticate device sample pairs and complete device authentication.

[0015] As an improvement, in step 1, a data sample set S1 is collected at position A for all devices within a certain period of time, wherein the same devices are device positive sample pairs and different devices are negative sample pairs, constituting device positive and negative sample pairs; then a data sample set S2 is collected at position B for all devices within a certain period of time, and the data sample pairs S1 and S2 constitute positive and negative sample pairs of channels at different positions.

[0016] As an improvement, in step 2, the structure of the twin adversarial model includes a feature extraction network G f (x,θ f ), device label prediction network G y (x,θ y ) and the position channel label prediction network G c (x,θ c ), and a gradient reversal layer between the extraction network and the prediction network; the structure is to force the device authentication network to learn features related to hardware fingerprints and decouple features related to channel positions through adversarial training.

[0017] As an improvement, the feature extraction network G f (x,θ f ) adopts a CNN neural network architecture, including two convolutional layers and two fully connected layers. The device label prediction network G y (x,θ y ) and the position channel label prediction network G c (x,θ c ) consists of two fully connected layers and outputs a two-dimensional vector as a label; the gradient reversal layer is used in the back-propagation process, and the gradient of the channel label prediction network loss is automatically reversed before being back-propagated to the parameters of the feature extractor.

[0018] As an improvement, the specific steps of obtaining the distance vector in step 3 are as follows:

[0019] Step 3.1: Add the data sample pairs S1 and S2 collected in step 1 to the feature extraction network G f (x,θ f ) Convert the input CSI map into a D-dimensional feature vector f∈R D , expressed as f = Gf (x,θ f ), where θ represents the parameters in the feature extraction network;

[0020] Step 3.2: Input the two feature vectors f1=G obtained from the sample pair f (x1,θ f ), f2=G f (x2,θ f ), calculate the distance vector between two eigenvectors:

[0021] As an improvement, in step 4, according to the sample labels, the distance vectors are added to the device sample pair label prediction network G y (x,θ y ) and channel sample pair label prediction network G c (x,θ c ); through G y (x,θ y ) and G y (d,θ y ) to obtain the device sample pair label y i and channel sample pair label c i , using cross entropy as the loss function of the device label prediction network, respectively expressed as:

[0022]

[0023] Then proceed to obtain the minimum L y The feature extractor parameters θ f ,θ y ; Maximize L y The feature extractor parameters under

[0024] θ f ,θ c , the training ends; the calculation formula 3 of the loss function is:

[0025]

[0026] In formula 1, L y (θ f ,θ y ) is the loss function of the device sample to the label prediction network; where θ f and θ y is the feature extractor parameter; N d represents the number of device sample pairs; i is the index of the device sample pair; Represents the D-dimensional feature vector output by the feature extractor.

[0027] In formula 2, L c (θf ,θ y ) is the loss function of the position channel sample to the label prediction network; where θ f and θ y is the feature extractor parameter; N c represents the number of device sample pairs; j is the index of the device sample pair; Denotes the D-dimensional feature vector representing the output of the feature extractor.

[0028] In formula 3, E(θ f ,θ y ,θ c ) is the loss function of the overall network; L y (θ f ,θ y ) is the loss function of the device sample to the label prediction network; L c (θ f ,θ y ) is the loss function of the position channel sample to the label prediction network; N d Represents the number of device sample pairs; N c The number λ of channel sample pairs is a weight coefficient used to balance the proportion of the loss of device samples to the label prediction network and the loss of channel samples to the label prediction network in the total loss function.

[0029] As an improvement, the gradient reversal layer is added in step 5 to perform network training and updating, which is to achieve the same transformation in the forward propagation process, as shown in Formula 4

[0030]

[0031] During the back propagation process, the gradient of the channel label prediction network loss is automatically negated before being back propagated to the parameters of the feature extractor. The parameter update process of the feature extraction network becomes:

[0032]

[0033] Through this adversarial method, a feature extractor is trained to extract features that are specific to the transmitting device but independent of the channel; where θ represents the parameters in the feature extraction network; x is the input CSI; R λ (x) is the transformation function; I is the identity matrix; μ is the learning rate parameter; L9 is the loss function of the device sample for the label prediction network; L : The loss function of the label prediction network for position channel samples.

[0034] As an improvement, in step 6, after the network training is completed, the channel label prediction network is removed and only the device label prediction network is retained; in the authentication stage, after the sample pair is feature extracted to obtain a distance vector, the distance vector is directly input into the device sample pair label prediction network to complete the authentication.

[0035] Beneficial effects: The technical solution proposed by the present invention is to collect channel state information CSI at different locations, input it into the twin adversarial model for training, and finally obtain an authentication network for identification through feature extraction and label prediction. The method is easy to implement and has good channel robustness, and can realize secure access authentication for wireless terminals. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is the CSI set collected by the same device at location A within a short period of time.

[0037] Figure 2 It is the CSI set collected by the same device at location B within a short period of time.

[0038] Figure 3 This is the model diagram of the twin adversarial model training stage of the present invention.

[0039] Figure 4 This is the model diagram of the twin adversarial model authentication stage of the present invention. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present invention will be described clearly and completely below so that those skilled in the art can better understand the advantages and features of the present invention, thereby making a clearer definition of the protection scope of the present invention. The embodiments described in the present invention are only a part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention.

[0041] In the present invention, a WiFi fingerprint authentication method based on adversarial training of channels at different locations comprises the following steps:

[0042] Step 1: Collect device and location channel CSI sample information and construct sample pairs S1 and S2.

[0043] Step 2: Input the data sample into the twin adversarial model.

[0044] Step 3: Based on the feature extractor G of the twin adversarial model f (x,θ f ), and further obtain the distance vector D(f 1, f2), where f 1,f2 is the feature vector of different samples, x is the CSI of the input feature extractor, θ f are the parameters in the feature extractor.

[0045] Step 4: Add the distance vectors obtained in step 3 to the device prediction label network G y (x,θ y ) and position channel prediction label network G c (x,θ c ).

[0046] Step 5: Add a gradient reversal layer GRL to update the network training so that the features it extracts are more focused on the hardware fingerprint characteristics of the device.

[0047] Step 6: Use the trained network model to authenticate device sample pairs and complete device authentication.

[0048] (1) In the data collection stage of step 1, the present invention collects a data sample set S1 at position 1 for all devices in a short period of time, wherein the same devices are device positive sample pairs and different devices are negative sample pairs, constituting device positive and negative sample pairs; and collects a data sample set S2 at position 2 for all devices in a short period of time, wherein S1 and S2 constitute the position channel positive and negative sample pairs.

[0049] Figure 1 and Figure 2 It is the CSI data collected by the same device at positions A and B in a short period of time; the channel state information (CSI) can be expressed as:

[0050] Y(K)=T(K)H(K)R(K)+Z(K)

[0051] Where R(K) and T(K) are the frequency response characteristics of the receiver and transmitter filters, respectively, H(K) is the frequency domain expression of the channel response, and Z(K) is the signal noise. Figure 1 and Figure 2 It can be seen that H(K) collected at different positions in a short period of time has a high degree of consistency, while different positions have significant differences, so different position channel labels can be given to positions 1 and 2.

[0052] Assuming that the signal-to-noise ratio of the received signal is high enough, the above formula is approximately equal to:

[0053] S(K)=T(K)R(K)H(K)

[0054] In a single receiver identification system, T(K)R(K) can be uniformly regarded as the transmitter fingerprint.

[0055] (2) The collected CSI data samples S1 and S2 are added to the twin adversarial model for pre-training. The twin adversarial model mainly consists of a training process and a certification stage, such as Figure 3 and Figure 4 As shown:

[0056] First, the feature extraction network G f (x,θ f ), and the device label prediction network G y (x,θ y ) and the position channel label prediction network G c (x,θ c ), and finally a gradient reversal layer (GRL) is added between the two. These three parts together constitute a twin adversarial model that can amplify the learning device features while ignoring the channel features.

[0057] (3) Add the data sample pairs S1 and S2 collected in step 1 to the feature extraction network G f (x,θ f ) Convert the input CSI map into a D-dimensional feature vector f∈R ； , expressed as f = G f (x,θ f ), where θ represents the parameters in the feature extraction network.

[0058] The two feature vectors f1=g obtained from the input sample pair f (x1,θ f ), f2=G f (x2,θ f ), calculate the distance vector between two eigenvectors:

[0059] (4) According to the sample label, the distance vector is added to the device sample pair label prediction network G y (x,θ y ) and channel sample pair label prediction network G c (x,θ c ); through G y (x,θ y ) and G y (d,θ y ) to obtain the device sample pair label y i and channel sample pair label c i , using cross entropy as the loss function of the device label prediction network, respectively expressed as:

[0060]

[0061] The following two tasks need to be completed during the training phase. The first one is to find the method that can minimize L y The feature extractor parameters θ f ,θ y , to achieve accurate prediction of device samples for labels and minimize device authentication errors; the second task is to find the one that can maximize L y The feature extractor parameters θ f ,θ c , we need to confuse the prediction results of channel samples on labels, so that the feature extraction layer can extract common features under different channels and maximize the channel identification error. The loss function of the overall network is defined as:

[0062]

[0063] Among them, L y (θ f ,θ y ) is the loss function of the device sample to the label prediction network; where θ f and θ y is the feature extractor parameter; N d represents the number of device sample pairs; i is the index of the device sample pair; Represents the D-dimensional feature vector output by the feature extractor.

[0064] Among them, L c (θ f ,θ y ) is the loss function of the position channel sample to the label prediction network; where θ f and θ y is the feature extractor parameter; N c represents the number of device sample pairs; j is the index of the device sample pair; Denotes the D-dimensional feature vector representing the output of the feature extractor.

[0065] Among them, E(θ f ,θ y ,θ c ) is the loss function of the overall network; L y (θ f ,θ y ) is the loss function of the device sample to the label prediction network; L c (θ f ,θ y ) is the loss function of the position channel sample to the label prediction network; N d Represents the number of device sample pairs; N c The number λ of channel sample pairs is a weight coefficient used to balance the proportion of the loss of device samples to the label prediction network and the loss of channel samples to the label prediction network in the total loss function.

[0066] (5) Adding a gradient reversal layer realizes the training update of the network and realizes the identity transformation in the forward propagation process, as shown in the formula

[0067] R λ (x) = x

[0068]

[0069] During the back propagation process, the gradient of the channel label prediction network loss is automatically negated before being back propagated to the parameters of the feature extractor. The parameter update process of the feature extraction network becomes:

[0070]

[0071] Through this adversarial method, a feature extractor can be trained to extract features that are specific to the transmitting device but independent of the channel. Where θ represents the parameters in the feature extraction network; x is the input CSI; R λ (x) is the transformation function; I is the identity matrix; μ is the learning rate parameter; L9 is the loss function of the device sample for the label prediction network; L : is the loss function of the position channel sample pair label prediction network. Figure 3 As shown, the CSI input is a data set of CSI at different locations.

[0072] (6) After network training is completed, Figure 4 As shown in Figure 1, the channel label prediction network is removed during the authentication phase, and only the device label prediction network is retained. After the sample pair is feature extracted to obtain the distance vector, the distance vector is directly input into the device sample pair label prediction network to complete the authentication.

[0073] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A training method for a WiFi device fingerprint authentication model based on channel confrontation at different locations, characterized in that: The authentication model includes the following steps: Step 1: Collect the channel state information (CSI) data of A and B at different positions and construct sample pairs S1 and S2; Step 2: Input the data sample into the twin adversarial model; Step 3: Based on the feature extractor G of the twin adversarial model f (x,θ f ), and obtain the distance vector D(f 1, f2); where f 1, f2 is the feature vector of different samples, x is the CSI of the input feature extractor, θ f is the parameter in the feature extractor; Step 4: Add the distance vectors obtained in step 3 to the device prediction label network G y (x,θ y ) and the position channel prediction label network G c (x,θ c );θ y Parameters,θ,in the device label prediction network c Predict parameters in the label network for the position channel; Step 5: Add a gradient reversal layer GRL to update the network training so that the features it extracts are more focused on the hardware fingerprint characteristics of the device; Step 6: Use the trained network model to authenticate device sample pairs and complete device authentication.

2. The training method of the WiFi device fingerprint authentication model based on different position channel confrontation according to claim 1 is characterized in that: In the step 1, a data sample set S1 is collected at position A for all devices within a certain period of time, wherein the same devices are device positive sample pairs and different devices are negative sample pairs, constituting device positive and negative sample pairs; then a data sample set S2 is collected at position B for all devices within a certain period of time, and the data sample pairs S1 and S2 constitute positive and negative sample pairs of channels at different positions.

3. The training method of the WiFi device fingerprint authentication model based on different position channel confrontation according to claim 1 is characterized in that: In step 2, the structure of the twin adversarial model includes a feature extraction network G f (x,θ f ), device label prediction network G y (x,θ y ) and the position channel label prediction network G c (x,θ c ), and a gradient reversal layer between the extraction network and the prediction network; the structure is to force the device authentication network to learn features related to hardware fingerprints and decouple features related to channel positions through adversarial training.

4. The training method of the WiFi device fingerprint authentication model based on different position channel confrontation according to claim 3 is characterized in that: The feature extraction network G f (x,θ f ) adopts a CNN neural network architecture, including two convolutional layers and two fully connected layers; the device label prediction network G y (x,θ y ) and the position channel label prediction network G c (x,θ c ) consists of two fully connected layers and outputs a two-dimensional vector as a label; the gradient reversal layer is used in the back-propagation process, and the gradient of the channel label prediction network loss is automatically reversed before being back-propagated to the parameters of the feature extractor.

5. According to the training method of the WiFi device fingerprint authentication model based on different position channel confrontation according to claim 1, Features: The specific steps of obtaining the distance vector in step 3 are as follows: Step 3.1: Add the data sample pairs S1 and S2 collected in step 1 to the feature extraction network G f (x,θ f ) Convert the input CSI map into a D-dimensional feature vector f∈R D , expressed as f = G f (x,θ f ), where θ represents the parameters in the feature extraction network; Step 3.2: Input the two feature vectors f1=G obtained from the sample pair f (x1,θ f ), f2=G f (x2,θ f ), calculate the distance vector between two eigenvectors:

6. The training method of the WiFi device fingerprint authentication model based on different position channel confrontation according to claim 1 is characterized in that: In step 4, according to the sample label, the distance vector is added to the device sample pair label prediction network G y (x,θ y ) and channel sample pair label prediction network G c (x,θ c ); through G y (x,θ y ) and G y (d,θ y ) to obtain the device sample pair label y i and channel sample pair label c i , using cross entropy as the loss function of the device label prediction network, respectively expressed as: Then proceed to obtain the minimum L y The feature extractor parameters θ f ,θ y ; Maximize L y The feature extractor parameters θ under f ,θ c , the training ends; the calculation formula 3 of the loss function is: In formula 1, L y (θ f ,θ y ) is the loss function of the device sample to the label prediction network; where θ f and θ y are the feature extractor parameters; N d represents the number of device sample pairs; i is the index of the device sample pair; Represents the D-dimensional feature vector output by the feature extractor; In formula 2, L c (θ f ,θ y ) is the loss function of the position channel sample to the label prediction network; where θ f and θ y are the feature extractor parameters; N c represents the number of device sample pairs; j is the index of the device sample pair; represents the D-dimensional feature vector output by the feature extractor; In formula 3, E(θ f ,θ y ,θ c ) is the loss function of the overall network; L y (θ f ,θ y ) is the loss function of the device sample to the label prediction network; L c (θ f ,θ y ) is the loss function of the position channel sample to the label prediction network; H d Represents the number of device sample pairs; N c The number λ of channel sample pairs is a weight coefficient used to balance the proportion of the loss of device samples to the label prediction network and the loss of channel samples to the label prediction network in the total loss function.

7. The training method of the WiFi device fingerprint authentication model based on different position channel confrontation according to claim 1 is characterized in that: In step 5, the gradient reversal layer is added to perform network training and updating, which is to achieve the identity transformation in the forward propagation process, as shown in Formula 4 During the back propagation process, the gradient of the channel label prediction network loss is automatically negated before being back propagated to the parameters of the feature extractor. The parameter update process of the feature extraction network becomes: Through this adversarial method, a feature extractor is trained to extract features that are specific to the transmitting device but independent of the channel; where θ represents the parameters in the feature extraction network; x is the input CSI; R λ (x) is the transformation function; I is the unit matrix; μ is the learning rate parameter; L6 is the loss function of the device sample to the label prediction network; L7 is the loss function of the location channel sample to the label prediction network.

8. The training method of the WiFi device fingerprint authentication model based on different position channel confrontation according to claim 1 is characterized in that: In step 6, when the network training is completed, the channel label prediction network is removed and only the device label prediction network is retained; in the authentication stage, after the sample pair is feature extracted to obtain a distance vector, the distance vector is directly input into the device sample pair label prediction network to complete the authentication.

Images