A physical layer key generation method based on deep learning

Abstract

The application discloses a physical layer key generation method based on deep learning, which comprises the following steps: channel estimation is performed at an Alice end and a Bob end; the channel estimation result of the Alice end is input into a trained neural network at the Alice end, the output of the trained neural network at the Alice end is quantized by using a key quantization algorithm, the channel estimation result of the Bob end is input into a trained neural network at the Bob end, and the output of the trained neural network at the Bob end is quantized by using the key quantization algorithm, so as to obtain key sequences K A and K B at the Alice end and the Bob end, wherein the trained neural network at the Alice end and the trained neural network at the Bob end are both constructed based on a self-encoder, and the method has the characteristics of high physical layer key generation rate and difficulty in eavesdropping.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of physical layer security, and relates to a physical layer key generation method based on deep learning. BACKGROUND

[0002] With the development of wireless communication technology, the number of mobile terminals increases rapidly, and diversified wireless devices urgently need to establish secure and reliable communication. Meanwhile, due to the natural openness of the transmission medium and the instability of the network structure in the wireless communication system, wireless transmission is easily eavesdropped and interfered by illegal users. Therefore, wireless network security technology is particularly important in wireless communication technology.

[0003] Traditional encryption mechanisms generally assume that the physical layer can achieve error-free transmission, and encrypt and decrypt information through key distribution at the network layer, but this usually relies on sacrificing complexity for security. With the continuous development of wireless networks, this mechanism will no longer be reliable and applicable. Unlike traditional upper-layer encryption technology, the physical layer key generation mechanism utilizes the reciprocity and spatial uniqueness of the wireless fading channel in the time division duplex mode to generate shared keys between legitimate users. The classical physical layer key generation process includes five parts: channel sounding, randomness extraction, quantization, information negotiation and privacy amplification.

[0004] Although the physical layer key generation protocol has been deeply studied, there are still great challenges and unsolved problems in designing a secure and reliable key generation scheme. The key technology to realize physical layer encryption is to ensure that the legitimate transceivers generate consistent keys, which requires that the channels observed by both parties have high reciprocity. However, due to the incomplete reciprocity of the actual communication system, the channels observed by both parties are very similar but not completely identical, and the generated keys have inconsistent bits. In order to reduce the mismatch rate of the keys generated by both parties, the scheme needs to rely on the legitimate users to exchange a large amount of information through the wireless channel for error correction, and frequent information exchange will leak a certain amount of key information. Due to the above disadvantages, the generation rate of the physical layer key in the actual system is low and easy to be eavesdropped. SUMMARY

[0005] The purpose of the present application is to overcome the above-mentioned shortcomings of the prior art, and to provide a physical layer key generation method based on deep learning, which has the characteristics of high physical layer key generation rate and is not easy to be eavesdropped.

[0006] In order to achieve the above purpose, the physical layer key generation method based on deep learning provided by the present application comprises:

[0007] Alice end performs channel estimation, and Bob end performs channel estimation;

[0008] The channel estimation result of the Alice end is input into the trained neural network of the Alice end, the output of the trained neural network of the Alice end is quantized by using the key quantization algorithm, the channel estimation result of the Bob end is input into the trained neural network of the Bob end, the output of the trained neural network of the Bob end is quantized by using the key quantization algorithm, and the key sequence K of the Alice end and the Bob end is obtained A and K B The trained neural network of the Alice end and the trained neural network of the Bob end are both constructed based on a self-encoder.

[0009] The input of the neural network of the Alice end is the channel estimation result of the Alice end The structure adopts a full connection network, and the activation function adopts a Relu function.

[0010] The input of the neural network of the Bob end is the channel estimation result of the Bob end The structure adopts a full connection network, and the activation function adopts a Relu function.

[0011] The neural network of the Alice end is also trained.

[0012] The specific process of training the neural network of the Alice end is as follows:

[0013] The sequence obtained by the maximum likelihood detection module of the Alice end is The cross entropy between the original pilot sequence x B is determined as the loss function of the neural network of the Alice end, and the parameters of the neural network of the Alice end are updated by using a stochastic gradient descent algorithm to minimize the loss function.

[0014] The neural network of the Bob end is also trained.

[0015] The specific process of training the neural network of the Bob end is as follows:

[0016] The output of the neural network of the Alice end is sent to the Bob end through a wireless channel, and after the Bob end receives it, the loss function of the neural network of the Bob end is constructed;

[0017] The parameters of the neural network of the Bob end are updated by using a stochastic gradient descent algorithm to minimize the loss function.

[0018] The loss function of the neural network of the Bob end is:

[0019]

[0020] Wherein, M is the data batch size of the first training, x A [i] is the label value of the i th sample, that is, the original pilot sequence sent by Alice, is the predicted value of the i th sample, L CEL is the cross-entropy loss function, and λ1 is the weight of the cross-entropy loss. MSE is the mean square error loss function, and λ2 is the weight of the mean square error loss.

[0021] The present application has the following beneficial effects:

[0022] The physical layer key generation method based on deep learning provided by the present application designs a neural network based on an autoencoder, generates a key by using the trained neural network, the neural network can efficiently represent channel characteristics and give a channel probe sequence with higher similarity, reduces the dependence of key generation on channel reciprocity, and is beneficial to improve the correlation of channel probe values; in addition, compared with traditional physical layer key generation technology, the present application generates a key based on a neural network, reduces information interaction in the key generation process, reduces the probability of key leakage, and at the same time improves the generation rate of the key; at the same time, the present application uses a neural network to realize a maximum likelihood detection module, and uses it for training of a channel feature learning module, avoiding the influence caused by the non-differentiable characteristics of the maximum likelihood detection. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The present application provides a key generation network design flowchart based on an autoencoder;

[0024] Figure 2 The present application provides a flowchart in the test stage;

[0025] Figure 3 The present application provides a curve graph of the maximum likelihood detection bit error rate under BPSK modulation varying with the signal-to-noise ratio;

[0026] Figure 4 The present application provides a curve graph of the bit error rate of the received signal after equalizing the channel estimation value varying with the signal-to-noise ratio;

[0027] Figure 5 The present application provides a curve graph of the MSE of the channel estimation results of the two communication parties varying with the signal-to-noise ratio;

[0028] Figure 6 The present application provides a curve graph of the key length extracted by the present application and the comparative scheme varying with the signal-to-noise ratio;

[0029] Figure 7 The present application provides a curve graph of the key inconsistency rate extracted by the present application and the comparative scheme varying with the signal-to-noise ratio. DETAILED DESCRIPTION

[0030] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments, and are not intended to limit the scope of the present application. In addition, in the following description, the description of the known structures and technologies is omitted to avoid unnecessary confusion of the concepts disclosed in the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of the present application.

[0031] The structural schematic diagrams according to the disclosed embodiments of the present application are shown in the drawings. These drawings are not drawn to scale, in which some details are enlarged for the purpose of clear expression, and some details can be omitted. The shapes of various regions, layers and their relative size and positional relationship shown in the drawings are only exemplary, and in actuality, there can be deviations due to manufacturing tolerances or technical limitations, and the regions / layers with different shapes, sizes and relative positions can be additionally designed by those skilled in the art according to actual needs.

[0032] Reference Figure 1 and Figure 2 The deep learning-based physical layer key generation method described in the present application includes the following steps:

[0033] 1) Training phase:

[0034] 11) Offline pre-training of a classification network capable of realizing signal detection in an AWGN channel under different signal-to-noise ratio environments. Since the maximum likelihood detection module needs to be called in the neural network training process, it is required that the maximum likelihood detection module is differentiable so that the internal parameters of the neural network can be passed, and therefore it is encapsulated as a module, and the maximum likelihood detection is realized inside by using a classification neural network.

[0035] 12) Alice and Bob respectively send N pilot signals for channel estimation;

[0036] The pilot sequence sent by Alice for the i-th time is x A (i), and the received signal of Bob at this time is represented as y B (i), and the channel estimation value is Similarly, the pilot sequence sent by Bob for the i-th time is x B (i), and the received signal of Alice at this time is represented as y A (i), and the channel estimation value is After sorting the N channel estimations, the data set of Alice is obtained as The label is The dataset of Bob side is The label is

[0037] 13) Initialize the fully connected neural network at Alice side and Bob side respectively, and construct two neural networks, wherein the input of the neural network at Alice side is The output is the channel feature value with higher consistency with Bob side The input of the neural network at Bob side is The output is the channel feature value with higher consistency with Alice side

[0038] 14) At Alice side, change the matrix shape of the output of the neural network to Use the received signal y at Alice side A The product of , realize channel equalization, and process the result of the product through the maximum likelihood detection module to obtain the recovered pilot sequence Bob side is the same, and the recovered pilot sequence is obtained

[0039] 15) Alice side sets the cross-entropy loss function for the pilot sequence The label is the original pilot sequence x B , and the error back propagation is used for training until the number of training iterations reaches the set iteration period.

[0040] 16) Bob side sets the cross-entropy loss function for the pilot sequence The label is the original pilot sequence x A The output obtained by Alice side after training is sent to Bob side, and the MSE loss function is set for the output of Bob side The label is The weights of the cross-entropy loss function of Alice side and the MSE loss function of Bob side are set and added, and the total error is used for back propagation until the number of training iterations reaches the set iteration period.

[0041] 17) The parameters of the neural network at Bob side are sent to Alice side, and Alice side and Bob side respectively calculate the output of the training sample through the network, and the MSE of the output result is counted.

[0042] 2) Test phase:

[0043] Alice side and Bob side respectively input the channel measurement value into the trained neural network, and respectively use the key quantization algorithm to quantize the output of the trained neural network to obtain the key sequence KA and K B , the flow chart of the test phase is shown in Figure 2 .

[0044] The maximum likelihood detection module in step 11) is a pre-trained binary classification neural network, which is a fully connected network. The structure and training results of the classification network are described below.

[0045] The input of the neural network is the signal with added noise at a specific signal-to-noise ratio: A [0] =x+n, x is the original signal, n is Gaussian white noise, the output and the input have the same dimension, both are 2; the structure of the neural network is a fully connected network, the activation function uses the sigmod function, and the loss function is the cross-entropy loss. Pre-train the maximum likelihood detection module in different signal-to-noise ratio environments, and when the neural network converges, the comparison chart of the training results and the theoretical curve is shown in the accompanying Figure 3 .

[0046] In step 13), Alice and Bob establish neural networks respectively, and the network structure of the two terminals is described below.

[0047] The neural network structure of Alice

[0048] 1a) The input of the neural network for channel estimation at Alice is the channel estimation result of Alice The structure adopts a fully connected network, and the activation function uses the Relu function; when the neural network converges, the output of the neural network is the channel feature value with higher similarity to Bob

[0049] 1b) Change the shape of the output of the neural network to Use the product of the received signal y A at Alice and to realize the function of the channel equalization module.

[0050] 1c) Take the output of the channel equalization module as the input of the maximum likelihood detection module for signal detection, and get the recovered pilot sequence When the neural network is trained to the convergence stage, the output of the maximum likelihood detection module increases the consistency between the original pilot sequence x B sent by Bob.

[0051] The neural network structure of Bob

[0052] 2a) The input of the autoencoder network for channel estimation at Bob is the channel estimation result of Bob The structure adopts a full connection network, and the activation function adopts a Relu function; when the neural network converges, the output of the neural network is a channel feature value with higher similarity to the Alice end

[0053] 2b) Change the shape of the output of the neural network to Use the product of the received signal y B of the Bob end to realize the function of the channel equalization module.

[0054] 3c) Use the output of the channel equalization module as the input of the maximum likelihood detection module for signal detection to obtain the recovered pilot sequence When the neural network is trained to the convergence stage, the output of the maximum likelihood detection module is more consistent with the original pilot sequence x A sent by the Alice end.

[0055] Steps 14) and 15) train the neural network of the Alice end, and steps 16) and 17) train the neural network of the Bob end. The training methods of the neural networks of the two ends are described below.

[0056] Training of the neural network of the Alice end

[0057] First, the neural network of the Alice end is trained. The Alice end determines the cross entropy between the sequence obtained by the maximum likelihood detection module and the original pilot sequence x B as the loss function of the Alice end, and uses the stochastic gradient descent algorithm to update the parameters of the neural network of the Alice end to minimize the loss function.

[0058] Training of the neural network of the Bob end

[0059] After the Alice end is trained, the output of the neural network of the Alice end is sent to the Bob end through a wireless channel. After receiving, the Bob end constructs the loss function of the neural network of the Bob end:

[0060]

[0061] Wherein, M is the data batch size of one training, x A [i] is the label value of the i-th sample, i.e., the original pilot sequence sent by the Alice end; is the predicted value of the i-th sample, i.e., the output of the maximum likelihood detection module of the Bob end; L CEL(·) is a cross-entropy loss function, and λ1 is the weight of the cross-entropy loss function, L MSE (·) is a mean square error loss function, and λ2 is the weight of the mean square error loss function.

[0062] Bob adopts a stochastic gradient descent algorithm to update the parameters of the neural network at the Bob end through training, so as to minimize the loss function.

[0063] In the key quantization algorithm in the test phase, a double-threshold quantization algorithm is adopted, and the specific implementation process is as follows:

[0064] Taking the Alice end as an example, the Bob end is the same. The Alice end performs channel sounding and obtains channel characteristics through channel estimation, and the channel characteristics are The quantization threshold is represented as:

[0065]

[0066] Wherein, mean(·) is the mean value of the channel characteristic sequence, σ is the quantization factor, var(·) is the variance of the channel characteristic sequence, and the quantization criterion is represented by a quantization function:

[0067]

[0068] Wherein, 2 is an illegal quantization code word and needs to be deleted, and the deletion process is essentially a negotiation process of the Alice end and the Bob end through a public channel to exchange non-quantized code words.

[0069] Figure 4 The error rate of the recovered transmission signal after the legitimate receiver performs channel equalization on the received signal according to the estimated channel parameters is shown.

[0070] Figure 5 The mean square error between the channel parameters observed by the legitimate parties under the least square method and the neural network algorithm is shown. After the neural network training, the mean square error of the channel parameters observed by the legitimate device is significantly reduced, which indicates that the consistency of the channel observation is improved.

[0071] Figure 6 The curve of the key generation rate extracted by the present application and the comparative scheme with the change of SNR is shown. Compared with the key generation rate directly using the traditional channel estimation value quantization, the key generation rate extracted by the present application is higher.

[0072] Figure 7 The curve of the key inconsistency rate extracted by the present application and the comparative scheme with the change of SNR is shown. Analysis shows that through the present application, the key inconsistency rate extracted by the legitimate parties is reduced, and the lower the signal-to-noise ratio, the greater the degree of reduction of the inconsistency rate, which indicates that the anti-noise ability of the present application is strong.

Claims

1. A deep learning-based physical layer key generation method, characterized by, Comprising: Channel estimation at the Alice end, channel estimation at the Bob end; The channel estimation result of the Alice end is input into the trained neural network of the Alice end, the output of the trained neural network of the Alice end is quantized by using the key quantization algorithm, the channel estimation result of the Bob end is input into the trained neural network of the Bob end, and the output of the trained neural network of the Bob end is quantized by using the key quantization algorithm, thereby obtaining the key sequence of the Alice end and the Bob end and wherein the trained neural network of the Alice end and the trained neural network of the Bob end are both constructed based on an autoencoder. Further comprising: training the neural network at the Bob end; The specific process of training the neural network at the Bob end is: The output of the neural network at the Alice end is sent to the Bob end through a wireless channel, and after being received by the Bob end, a loss function of the neural network at the Bob end is constructed. The Bob end adopts a stochastic gradient descent algorithm to update the parameters of the neural network at the Bob end through training, so as to minimize the loss function; The loss function of the neural network at the Bob end is: in, The size of the data batch for one training session. For the first The label values ​​of each sample, i.e., the original pilot sequence sent by Alice. For the first The predicted value for each sample, Let cross-entropy be the loss function. The weights are the cross-entropy loss values. Let the mean squared error loss function be used. The weights for the mean squared error loss; Further comprising: training the neural network at the Alice end; The specific process of training the neural network at the Alice end is: The sequence obtained by the maximum likelihood detection module at the Alice end The cross entropy between the original pilot sequence is determined as the loss function of the neural network at the Alice end, and the parameters of the neural network at the Alice end are updated by training through the stochastic gradient descent algorithm to minimize the loss function. 2.The deep learning based physical layer key generation method of claim 1, wherein, The input of the neural network at the Alice end is the channel estimation result at the Alice end ; The structure adopts a fully connected network, and the activation function adopts a Relu function. 3.The deep learning based physical layer key generation method of claim 1, wherein, The input of the neural network at the Bob end is the channel estimation result at the Bob end ; The structure adopts a fully connected network, and the activation function adopts a Relu function.

Images