Industrial mobile equipment physical layer security authentication method based on ViT

Abstract

The invention relates to the field of industrial wireless network physical layer security, in particular to an industrial mobile device physical layer security authentication method based on a Vision Transform (ViT). The method is suitable for a physical layer authentication problem of wireless equipment in an industrial scene, and comprises three stages of data preprocessing, offline training and online authentication. Specifically, in the data preprocessing stage, channel impulse response is converted into channel frequency response (CFR), then CFR samples are converted into double-channel CFR images, request positions are spliced into position vectors, and finally the double-channel CFR images, the position vectors and corresponding sample labels form labeled training samples; in the off-line training stage, the labeled training samples constructed in the data preprocessing stage are used for carrying out supervised learning training on the designed deep neural network (DNN) model; and in the online authentication stage, acquiring an authentication request signal in real time, and judging the legality of authentication request equipment according to the trained DNN model.

Description

Technical Field

[0001] The invention relates to the field of physical layer security of industrial wireless networks, and in particular to a physical layer security authentication method for industrial mobile equipment based on Vision Transformer (ViT). Background Art

[0002] With the advent of Industry 4.0, factory automation has become a key factor in smart manufacturing. Wireless networks are extremely attractive in implementing factory automation due to their support for rotation and mobility, high deployment flexibility, and low installation costs. However, due to the open and broadcast characteristics of wireless channels, wireless networks face serious threats in terms of physical layer security, including eavesdropping attacks, interference attacks, and spoofing attacks. Among these security threats, spoofing attacks are the most covert and the most difficult to defend against. Spoofing attacks refer to attackers imitating legitimate devices to gain wireless access, implement device hijacking, and inject false information, which may cause wireless devices to make harmful decisions.

[0003] Authentication is generally considered a promising method to counter spoofing attacks. Compared with authentication methods based on upper-layer encryption, physical layer authentication (PLA) uses the inherent randomness and uniqueness of wireless channels to make it difficult for attackers to imitate legitimate devices. As a lightweight authentication method, PLA is very friendly to resource-constrained devices. Among the existing PLA methods based on channel characteristics, the methods based on channel state information (CSI) perform better. However, in industrial environments, due to the presence of a large number of rotating and mobile devices, coupled with the movement of the authentication request device itself, the CSI of the wireless channel is extremely unstable. At the same time, industrial applications require real-time authentication, which brings challenges to PLA for industrial mobile devices. Some existing PLA methods for industrial mobile devices ignore the spatiotemporal correlation between training samples and do not make full use of auxiliary information such as request location, which limits the performance of PLA. Summary of the invention

[0004] Aiming at wireless devices in industrial scenarios, the present invention proposes a physical layer security authentication method for industrial mobile devices based on ViT. This method is applicable to the physical layer authentication problem of wireless devices in industrial scenarios, and includes three stages: data preprocessing, offline training, and online authentication. Specifically, in the data preprocessing stage, the channel impulse response (CIR) is converted into a channel frequency response (CFR), and then the CFR sample is converted into a dual-channel CFR image, and the request position is spliced ​​into a position vector, and finally the dual-channel CFR image, the position vector and the corresponding sample label are combined into a labeled training sample; in the offline training stage, the labeled training samples constructed in the data preprocessing stage are used to perform supervised learning training on the designed deep neural network (DNN) model; in the online authentication stage, the authentication request signal is obtained in real time, and the legitimacy of the authentication request device is judged according to the trained DNN model.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0006] A physical layer security authentication method for industrial mobile equipment based on ViT, comprising the following steps:

[0007] Data preprocessing stage: convert the channel impulse response CIR into the channel frequency response CFR, then convert the CFR sample into a dual-channel CFR image, and splice the request position into a position vector. Finally, the dual-channel CFR image, position vector and corresponding sample label form a labeled training sample.

[0008] Offline training phase: Use labeled training samples to perform supervised learning training on the deep neural network DNN model;

[0009] Online authentication phase: Acquire the authentication request signal in real time and determine the legitimacy of the authentication request device based on the trained DNN model.

[0010] The CIR describes the distortion of the received signal relative to its original transmitted signal. is the sampling of the transmission pilot sequence in the tth period, T represents the transposition, represents a complex number, M represents the sampling length, X t After propagating through the wireless channel, it is measured by the receiver as Then the CIR of the wireless channel in the tth period calculated by the receiver is Specifically:

[0011] h t =R t ★X t

[0012] Among them, ★ is the cross-correlation operation.

[0013] The CIR is converted into CFR by fast Fourier transform, that is, h t Convert to For F t The calculation formula for the kth element in is:

[0014]

[0015] Among them, h t (m) is h t The mth element of , e is a natural constant, and j is an imaginary unit.

[0016] The CFR sample is converted into a dual-channel CFR image, specifically:

[0017] First, F t Represented as two real vectors re(F t ) and im(F t ), re(F t ) and im(F t ) represent F t The real and imaginary parts of the image are then used to construct a dual-channel CFR image, with one channel being the real CFR image. ×M, the other channel is a virtual CFR image Represents a real number, N is the number of reference CFRs, and the reference CFRs need to come from legitimate devices, RCI t and ICI t Specifically:

[0018]

[0019] The request positions are spliced ​​into position vectors, specifically:

[0020] The location of the authentication requesting device in the tth cycle is Then p(t) and the vector P consisting of N reference positions A (t) is concatenated into [p(t); P A (t)], where A indicates that the reference location comes from a legitimate device, P A (t) Specifically P A (t) = [p A (t-1); p A (t-2);…;p A (tN)].

[0021] The deep neural network DNN model consists of a ViT and two feedforward neural networks FNN1 and FNN2, wherein the input of ViT is a dual-channel CFR image, the input of FNN1 is a position vector, the outputs of ViT and FNN1 are used as the input of FNN2 together, and the discrimination result is output. Both FNN1 and FNN2 include three fully connected layers, namely FC1, FC2 and FC3, and each layer uses a ReLU activation function.

[0022] The offline training stage is specifically as follows:

[0023] The dual-channel CFR image with a size of 2×(N+1)×M is input into ViT. The output of ViT is the feature vector v1, and the position vector [p(t); P A (t)] is input into FNN1, and the output of FNN1 is feature vector v2. Then v1 and v2 are concatenated and input into FNN2 for authentication. The output of FNN2 is v3 = [w θ ,1-w θ ],w θ represents the probability that the authentication request device is identified as an illegal device, θ is the parameter of the DNN model, if w θ ≤0.5, the authentication requesting device is considered to be a legitimate device, otherwise it is considered to be an illegal device;

[0024] The DNN model is trained through supervised learning, and the network parameters are optimized and updated to minimize the cross entropy loss L(θ).

[0025] The cross entropy loss L(θ) is:

[0026] L(θ)=-ylog 10 w θ -(1-y)log 10 (1-w θ )

[0027] Wherein, y is the real label of the authentication requesting device. If y=0, it means that the authentication requesting device is a legal device, and y=1 means that the authentication requesting device is an illegal device.

[0028] The online authentication stage is specifically as follows:

[0029] When the receiver receives a new authentication request signal, it first calculates h t , then h t Convert to F t , and construct a dual-channel CFR image, stitching the requested position into a position vector [p(t); P A (t)], and finally the dual-channel CFR image and position vector [p(t); P A(t)] is input into the trained DNN model to determine the legitimacy of the current authentication request device.

[0030] A physical layer security authentication system for industrial mobile equipment based on ViT, comprising:

[0031] The data preprocessing module is used to convert the channel impulse response CIR into the channel frequency response CFR, then convert the CFR sample into a dual-channel CFR image, and splice the request position into a position vector, and finally combine the dual-channel CFR image, the position vector and the corresponding sample label into a labeled training sample;

[0032] Offline training module, used to perform supervised learning training on deep neural network DNN model using labeled training samples;

[0033] The online authentication module is used to obtain the authentication request signal in real time and determine the legitimacy of the authentication request device based on the trained DNN model.

[0034] The present invention has the following beneficial effects and advantages:

[0035] 1. CIR reflects the multipath fading of the wireless channel. Due to the limited number of transmission paths in the real environment, the original CIR data is usually sparse. Therefore, the generalization ability of the trained DNN model decreases. In contrast, CFR represents the response of the wireless channel at different frequency points. CFR is far less sparse than CIR, making it more suitable for training DNN models.

[0036] 2. Convert the CFR samples into dual-channel CFR images. At the same time, the request position can be introduced as auxiliary information to provide additional spatial features. Through the above operations, the DNN model can learn the spatiotemporal correlation between CFR samples, thereby improving the performance of PLA. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the system model;

[0038] Figure 2 This is a schematic diagram of the splicing of CFR data;

[0039] Figure 3 Schematic diagram of the designed DNN model;

[0040] Figure 4 This is a schematic diagram of the online authentication process. DETAILED DESCRIPTION

[0041] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings.

[0042] like Figure 1As shown, the present invention considers an industrial mobile scenario with a dynamic wireless channel. In this scenario, Alice (legitimate sender) and Bob (authenticator) communicate in the presence of Eve (deceptive attacker), and all three are equipped with a single antenna. Specifically, Alice moves along a predefined trajectory while sending an authentication request to Bob. Eve maintains a constant distance from Alice and moves along the same trajectory with the same movement pattern, so that it can imitate Alice to send deceptive requests in idle time slots. Bob is located in a fixed position and is in the same position as the edge server. The edge server is responsible for PLA, and after Bob receives the authentication request, the PLA is used to determine the legitimacy of the authentication request device.

[0043] A physical layer security authentication method for industrial mobile equipment based on ViT, comprising the following steps:

[0044] Data preprocessing stage: convert CIR into CFR, then convert CFR samples into dual-channel CFR images, and splice the request positions into position vectors. Finally, the dual-channel CFR images, position vectors and corresponding sample labels are combined into labeled training samples.

[0045] Offline training phase: supervised learning training of the designed DNN model is performed using the labeled training samples constructed in the data preprocessing phase;

[0046] Online authentication phase: Acquire the authentication request signal in real time and determine the legitimacy of the authentication request device based on the trained DNN model.

[0047] The CIR describes the distortion of the received signal relative to its original transmitted signal. is the sampling of the transmission pilot sequence in the tth period, T represents the transposition, represents a complex number, M represents the sampling length, X t After propagating through the wireless channel, it is measured by the receiver as Then the receiver can calculate the CIR of the wireless channel in the tth period as The specific calculation formula is as follows:

[0048] h t =R t ★X t (1)

[0049] Among them, ★ is the cross-correlation operation.

[0050] In the data preprocessing stage, CIR is converted into CFR, that is, h is converted into CFR by fast Fourier transform. t Convert to For F t The calculation formula for the kth element in is:

[0051]

[0052] Among them, h t (m) is h t The mth element of , e is a natural constant, and j is an imaginary unit.

[0053] In the data preprocessing stage, the CFR samples are converted into dual-channel CFR images, specifically:

[0054] First, F t Represented as two real vectors re(F t ) and im(F t ), re(F t ) and im(F t ) represent F t The real and imaginary parts of Figure 2 The dual-channel CFR image is constructed in the manner shown, one channel is a real CFR image The other channel is a virtual CFR image Represents a real number, N is the number of reference CFRs, and the reference CFRs need to come from legitimate devices, RCI t and ICI t The details are as follows:

[0055]

[0056] In the data preprocessing stage, the requested positions are spliced ​​into position vectors, specifically:

[0057] The location of the authentication requesting device in the tth cycle is Then p(t) and N reference positions can be concatenated as [p(t); P A (t)], A indicates that the reference position comes from a legitimate device, P A (t) Specifically P A (t) = [p A (t-1); p A (t-2);…;p A (tN)].

[0058] In the offline training stage, the labeled training samples constructed in the data preprocessing stage are used to perform supervised learning training on the designed DNN model, specifically:

[0059] The designed DNN model is as follows Figure 3As shown, it consists of a ViT and two feedforward neural networks (FNNs). The two FNNs are denoted as FNN1 and FNN2. FNN1 and FNN2 follow the basic FNN structure and include three fully connected layers, namely FC1, FC2 and FC3. Each layer uses the ReLU activation function.

[0060] During the training process, a dual-channel CFR image with a size of 2×(N+1)×M is input into ViT. The output of ViT is the feature vector v1, and the position vector [p(t); P A (t)] is input into FNN1, and the output of FNN1 is feature vector v2. Then v1 and v2 are concatenated and input into FNN2 for authentication. The output of FNN2 is v3 = [w θ ,1-w θ ],w θ represents the probability that the authentication request device is identified as an illegal device, θ is the parameter of the DNN model, if w θ ≤0.5, the authentication requesting device is considered to be a legitimate device, otherwise it is considered to be an illegal device;

[0061] In the offline training stage, the designed DNN model is trained by supervised learning, and the network parameters are optimized and updated to minimize the cross entropy loss L(θ), that is,

[0062] L(θ)=-ylog 10 w θ -(1-y)log 10 (1-w θ ) (5)

[0063] Wherein, y is the real label of the authentication requesting device. If y=0, it means that the authentication requesting device is a legal device, and y=1 means that the authentication requesting device is an illegal device.

[0064] In the online authentication phase, the authentication request signal is acquired in real time, and the legitimacy of the authentication request device is determined based on the trained DNN model, specifically:

[0065] Online certification process Figure 4 As shown in the figure, when the receiver receives a new authentication request signal, it first calculates h according to (1). t , then according to (2) h t Convert to F t , and then construct a dual-channel CFR image according to (3) and (4), and splice the requested position into a position vector [p(t); P A (t)], and finally the dual-channel CFR image and position vector [p(t); P A(t)] is input into the trained DNN model to determine the legitimacy of the current authentication request device. If the authentication request device is determined to be a legitimate device, the authentication request will be authorized and h t and p(t) to update the reference CIR matrix and the reference position P A (t) For subsequent authentication, AB indicates that the reference CIRs are all from the Alice-Bob legitimate channel. Specifically Otherwise, the authentication request will be rejected to prevent potential attacks.

Claims

1. A physical layer security authentication method for industrial mobile equipment based on ViT, characterized in that: The following steps are involved: Data preprocessing stage: convert the channel impulse response CIR into the channel frequency response CFR, then convert the CFR sample into a dual-channel CFR image, and splice the request position into a position vector. Finally, the dual-channel CFR image, position vector and corresponding sample label form a labeled training sample. Offline training phase: Use labeled training samples to perform supervised learning training on the deep neural network DNN model; Online authentication phase: Acquire the authentication request signal in real time and determine the legitimacy of the authentication request device based on the trained DNN model.

2. According to claim 1, a physical layer security authentication method for industrial mobile equipment based on ViT is characterized in that: The CIR describes the distortion of the received signal relative to its original transmitted signal. is the sampling of the transmission pilot sequence in the tth period, T represents the transposition, represents a complex number, M represents the sampling length, X t After propagating through the wireless channel, it is measured by the receiver as Then the CIR of the wireless channel in the tth period calculated by the receiver is Specifically: h t =R t ★X t Among them, ★ is the cross-correlation operation.

3. According to claim 1, a physical layer security authentication method for industrial mobile equipment based on ViT is characterized in that: The CIR is converted into CFR by fast Fourier transform, that is, h t Convert to For F t The calculation formula for the kth element in is: Among them, h t (m) is h t The mth element of , e is a natural constant, and j is an imaginary unit.

4. According to claim 1, a physical layer security authentication method for industrial mobile equipment based on ViT is characterized in that: The CFR sample is converted into a dual-channel CFR image, specifically: First, F t Represented as two real vectors re(F t ) and im(F t ), re(F t ) and im(F t ) represent F t The real and imaginary parts of the image are then used to construct a dual-channel CFR image, with one channel being the real CFR image. The other channel is a virtual CFR image Represents a real number, N is the number of reference CFRs, and the reference CFRs need to come from legitimate devices, RCI t and ICI t Specifically:

5. According to claim 1, a physical layer security authentication method for industrial mobile equipment based on ViT is characterized in that: The request positions are spliced ​​into position vectors, specifically: The location of the authentication requesting device in the tth cycle is Then p(t) and the vector P consisting of N reference positions A (t) is concatenated into [p(t); P A (t)], where A indicates that the reference location comes from a legitimate device, P A (t) Specifically P A (t) = [p A (t-1); p A (t-2);…;p A (tN)].

6. The physical layer security authentication method for industrial mobile equipment based on ViT according to claim 1 is characterized in that: The deep neural network DNN model consists of a ViT and two feedforward neural networks FNN1 and FNN2, wherein the input of ViT is a dual-channel CFR image, the input of FNN1 is a position vector, the outputs of ViT and FNN1 are used as the input of FNN2 together, and the discrimination result is output. Both FNN1 and FNN2 include three fully connected layers, namely FC1, FC2 and FC3, and each layer uses a ReLU activation function.

7. The physical layer security authentication method for industrial mobile equipment based on ViT according to claim 1 is characterized in that: The offline training stage is specifically as follows: The dual-channel CFR image with a size of 2×(N+1)×M is input into ViT. The output of ViT is the feature vector v1, and the position vector [p(t); P A (t)] is input into FNN1, and the output of FNN1 is feature vector v2. Then v1 and v2 are concatenated and input into FNN2 for authentication. The output of FNN2 is v3 = [w θ ,1-w θ ],w θ represents the probability that the authentication request device is identified as an illegal device, θ is the parameter of the DNN model, if w θ ≤0.5, the authentication requesting device is considered to be a legitimate device, otherwise it is considered to be an illegal device; The DNN model is trained through supervised learning, and the network parameters are optimized and updated to minimize the cross entropy loss L(θ).

8. The physical layer security authentication method for industrial mobile equipment based on ViT according to claim 7 is characterized in that: The cross entropy loss L(θ) is: L(θ)0-ylog 10 w θ -(1-y)log 10 (1-w θ ) Wherein, y is the real label of the authentication requesting device. If y=0, it means that the authentication requesting device is a legal device, and y=1 means that the authentication requesting device is an illegal device.

9. The physical layer security authentication method for industrial mobile equipment based on ViT according to claim 1, characterized in that: The online authentication stage is specifically as follows: When the receiver receives a new authentication request signal, it first calculates h t , then h t Convert to F t , and construct a dual-channel CFR image, stitching the requested position into a position vector [p(t); P A (t)], and finally the dual-channel CFR image and position vector [p(t); P A (t)] is input into the trained DNN model to determine the legitimacy of the current authentication request device.

10. A physical layer security authentication system for industrial mobile equipment based on ViT, characterized in that: include: The data preprocessing module is used to convert the channel impulse response CIR into the channel frequency response CFR, then convert the CFR sample into a dual-channel CFR image, and splice the request position into a position vector, and finally combine the dual-channel CFR image, the position vector and the corresponding sample label into a labeled training sample; Offline training module, used to perform supervised learning training on deep neural network DNN model using labeled training samples; The online authentication module is used to obtain the authentication request signal in real time and determine the legitimacy of the authentication request device based on the trained DNN model.

Images