End-to-end radio-over-fiber adaptive confidential transmission method based on hybrid drive

Abstract

The invention discloses an end-to-end radio-over-fiber adaptive secrecy transmission method based on hybrid driving, and the method specifically comprises the steps: building a radio-over-fiber transmission link, and carrying out a link transmission experiment; collecting signal data, namely a link transmitting end radio frequency signal and a receiving end receiving radio frequency signal; constructing a channel model based on a neural network, and training the model; establishing a transmitting end modulation model, a receiving end demodulation model and a trained channel model to form an end-to-end confidential transmission system; training an end-to-end confidential transmission system model; after training is completed, the modulation model is an encryption module, and the demodulation model is used as a decryption module of a receiving end; and the radio frequency signal generated by the encryption module is input into the radio-over-fiber link for transmission, and the demodulation model is used for carrying out demodulation test on the received radio frequency signal. According to the invention, the security and reliability of the transmission process are improved.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning and optical communication, and in particular relates to an end-to-end optical-carried radio frequency adaptive secure transmission method based on hybrid drive. Background Art

[0002] The analog radio frequency over fiber (RoF) transmission technology based on intensity modulation direct detection (IM / DD) is highly compatible with the system architecture of mobile fronthaul networks due to its significant advantages such as low transmission loss, large bandwidth, simple structure, low cost and strong anti-electromagnetic interference ability. With its excellent performance, this technology shows broad application prospects in future mobile fronthaul networks. With the growing demand for efficient and low-latency data transmission in 5G and future networks, IM / DD RoF technology will become an important support in the field of mobile communications.

[0003] Although optical fiber transmission itself has certain anti-eavesdropping and anti-interference capabilities, in practical applications, with the increasing demand for data security, relying solely on the natural advantages of physical transmission is no longer sufficient to deal with potential eavesdropping and attack risks. Therefore, traditional physical layer encryption methods are introduced to further enhance data security during transmission. These encryption methods include XOR logic encryption, electrical logic encryption, and chaotic encryption. The core idea is to introduce noise into the transmission signal to achieve the effect of data encryption. The receiving end restores useful information through a key or synchronization mechanism. However, these methods have some inherent defects: on the one hand, the noise signal may be damaged after passing through a complex channel, making it difficult for the receiving end to completely remove the noise; on the other hand, the key synchronization process is prone to errors, which causes the bit error rate of the receiving end to increase, ultimately affecting the system performance. Summary of the invention

[0004] In view of the above shortcomings, the present invention provides an end-to-end optical radio frequency adaptive secure transmission method based on hybrid drive.

[0005] The present invention provides an end-to-end optical radio frequency adaptive secure transmission method based on hybrid drive, comprising the following steps:

[0006] Step 1: Build an optical radio frequency transmission link and conduct a link transmission experiment.

[0007] A simulated optical RF link based on intensity modulation direct detection was built, where the center frequency of the RF signal was 10 GHz, the symbol rate of the signal was 2 GBaud, and the optical fiber transmission distances were 18 km and 20 km respectively.

[0008] Step 2: Collect signal data, namely the RF signal at the link transmitter and the RF signal at the receiver.

[0009] In step 2, the RF signal data of the transmitting end and the RF signal output by the beat frequency of the photoelectric detector at the receiving end are collected, wherein the sampling rate is 40GSa / s. The optical fiber transmission distance is changed, and the corresponding signal data is collected.

[0010] Step 3: Construct a channel model ChannelNN based on a neural network, and use the signal data obtained in step 2 to train the channel model ChannelNN.

[0011] In step 3, a channel model ChannelNN based on a neural network is constructed, and the signal data collected in step 2 is used to train the legal channel model and the illegal channel model.

[0012] Step 4: Establish the transmitter modulation model TransNN, the receiver demodulation model ReceivNN and the trained channel model ChannelNN to form an end-to-end secure transmission system.

[0013] Step 5: Train the end-to-end secure transmission system model. The training process is as follows:

[0014] (1) Generate a random bit sequence as a training set for the end-to-end secure transmission system.

[0015] (2) Input the generated random bit sequence into the end-to-end learning secure transmission system.

[0016] (3) Calculate the mean square error (MSE) between the fitting bits and the training bits output by the legitimate system, as well as the mean square error between the fitting bits and the training bits output by the eavesdropped channel.

[0017] (4) Constructing the loss function: The mean square error is used as the loss function for forward training, and the negative mean square error is used as the loss function for adversarial training. Finally, these two loss functions are weighted and fused to form the total loss function. The weighted loss function is jointly optimized through the optimizer, thereby simultaneously optimizing the performance of the transmitter modulation model TransNN and the receiver demodulation model ReceivNN.

[0018] (5) The secure transmission system model is trained repeatedly until a preset number of training cycles or a performance target is reached.

[0019] Step 6: After training, the modulation model TransNN is used as the encryption module, and the demodulation model ReceivNN is used as the decryption module at the receiving end. The RF signal generated by the encryption module is input into the optical carrier RF link for transmission, and the received RF signal is demodulated and tested using the ReceivNN model.

[0020] Furthermore, the input of the modulation model TransNN is bit information; wherein the input bit needs to be one-hot encoded; the modulation model TransNN is a neural network model composed of a fully connected layer and a convolutional network, and other neural network structures can also be used. It is worth noting that the number of neurons in the input layer of the modulation model is related to the modulation order. For example, when the modulation format is 64-QAM, the number of neurons in the input layer is 64. The number of neurons in the output layer of the modulation model is related to the upsampling coefficient, specifically twice the upsampling coefficient.

[0021] Furthermore, the demodulation model ReceivNN is composed of a neural network model, in which the number of neurons in the input layer is equal to the upsampling coefficient, and the number of neurons in the output layer is equal to the modulation order;

[0022] At the same time, the fitting bit value output by the demodulation model needs to be between 0 and 1; the improved hyperbolic tangent activation function Tanh is used in the output layer of the demodulation model ReceivNN, and its expression is:

[0023]

[0024] In the formula, e x and e -x is a natural exponential function.

[0025] Furthermore, in step 5, the main channel model and the eavesdropping channel model need to be trained separately to obtain data-driven main channel model and eavesdropping channel model.

[0026] Model training is divided into two parts. The first is the forward training part of the main channel: the bit information is input into TransNN (symbol mapping and upsampling), and the output baseband signal is matched filtered and up-converted to generate an RF signal, which is then input into the data-driven main channel model and finally passed to the ReceivNN model to demodulate the received RF signal into the fitted bit information. In this process, we hope that the gap between the fitted bit information and the original bit information is as small as possible, that is, during the main channel transmission, the RF signal generated by TransNN can be perfectly demodulated by the ReceivNN model. Therefore, the normalized mean square error (MSE) is selected as the loss function to minimize the demodulation error.

[0027] Next is the adversarial training process of the eavesdropping channel: the bit information is input into TransNN, the generated baseband signal is processed by matched filtering and up-conversion, the RF signal is generated, and input into the data-driven eavesdropping channel model, and finally passed to the ReceivNN model to demodulate the received RF signal into the fitted bit information. In this process, we hope that the gap between the fitted bit information and the original bit information is as large as possible, that is, during the transmission of the eavesdropping channel, the RF signal generated by TransNN should reduce the demodulation performance of the ReceivNN model. Therefore, the mean square error (MSE) is used as the loss function to maximize the demodulation error.

[0028] During the training process, the loss functions of forward training and adversarial training are weighted and combined into a unified loss function, which is then jointly optimized using an optimizer to achieve efficient training of TransNN and ReceivNN, balancing the transmission accuracy of the main channel and the confidentiality of the eavesdropping channel.

[0029] Furthermore, the loss function is constructed: it is composed of the mean square error MSE_loss between the fitting bit and the transmitted bit output by the legitimate channel and the mean square error mse_loss between the fitting bit and the transmitted bit output by the eavesdropping channel. The two are weighted and summed according to different weight coefficients. The expression is:

[0030] loss=(1-θ)×MSE_loss-θ×mse_loss (2)

[0031] Among them, θ is the weight coefficient, and its value range is (0, 1).

[0032] Furthermore, the weight coefficient θ in the loss function is adjusted according to the size of the loss value during the training process.

[0033] Furthermore, during the training process, disturbance noise is added to both ends of the main channel model, and disturbance noise is added to the front end of the eavesdropping channel model.

[0034] The beneficial technical effects of the present invention compared with the prior art are:

[0035] The present invention can achieve highly secure transmission by adaptively learning and adjusting the signal encoding method and upsampling rules without adding complex physical devices. This method relies on data and model-driven self-supervised learning. After obtaining the channel model, only random bit streams are required to complete efficient training. More importantly, this solution makes full use of the inherent security characteristics of neural networks, making the encryption and decryption models extremely difficult to be cracked by brute force, effectively improving the security and reliability of the transmission process. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The present invention is a flow chart of the end-to-end radio frequency over light security method based on hybrid drive.

[0037] Figure 2 The block diagram of the end-to-end RF over fiber link secure transmission structure based on hybrid drive. (A1) RF over fiber link structure diagram. (A2) Traditional RF over fiber link system structure diagram. (A3) End-to-end RF over fiber secure transmission link based on neural network.

[0038] Figure 3 This is a schematic diagram of the structure of the transmitter's neural network-based confidentiality modulation model TransNN.

[0039] Figure 4 This is the training flow chart of the hybrid-driven end-to-end RF-over-optical secure transmission link model.

[0040] Figure 5 This is a block diagram for training the end-to-end RF-over-optical secure transmission link model based on hybrid drive.

[0041] Figure 6 This is the constellation diagram generated by TransNN0 in the first training.

[0042] Figure 7 Constellation diagrams generated by the neural network-based modulation model under different training rounds (a~f are the constellation diagrams generated by TransNN1 to TransNN6 respectively).

[0043] Figure 8 It is the demodulation performance of the demodulation module at the receiving end in different rounds under white-box attack.

[0044] Fig. 9 Figure 2 shows the demodulation performance of the demodulation module at the receiving end under gray box attack at different fiber distances. DETAILED DESCRIPTION

[0045] The present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0046] The process of the end-to-end optical radio frequency adaptive secure transmission method based on hybrid drive of the present invention is as follows: Figure 1 As shown, the specific steps include:

[0047] Step 1: Construct an optical radio frequency transmission link. The specific link structure is as follows: Figure 2 As shown in (A1): The arbitrary waveform generator generates an RF signal to drive the intensity modulator. The modulated optical signal is transmitted through the optical fiber and received by the photodetector after being processed by the attenuator. The received RF signal is amplified by the electrical amplifier and then collected by the oscilloscope and digitally processed.

[0048] Step 2: Collect the RF signals at the transmitting and receiving ends of the link to obtain a data set for the channel model. Figure 2 As shown in (A1), in the schematic diagram of the link structure, the dotted box represents the channel part. In this step, it is necessary to collect the RF signal of the transmitter and the signal of the receiver collected by the oscilloscope. The transmitted RF signal adopts single carrier quadrature amplitude modulation (SC-QAM). Figure 2 (A2) shows the modulation process of SC-QAM, including symbol mapping, upsampling, matched filtering and upconversion. These processes are completed using traditional algorithms: symbol mapping is standard constellation points, and upsampling methods are zero interpolation or repeated sampling. The collected RF signal at the transmitter is the signal modulated by SC-QAM, while the signal collected at the receiver is the RF signal obtained by the photoelectric detector after the RF signal is transmitted through the optical link. Signal data with optical fiber lengths of 18 km and 20 km were collected to build channel models for different optical fiber distances.

[0049] Step 3: A neural network-based channel model ChannelNN is constructed, and the model is trained using the data collected in step 2, thereby obtaining different data-driven channel models with fiber distances of 18 kilometers and 20 kilometers. Among them, the optical fiber RF link channel with a fiber distance of 20 kilometers is used as the main channel, and the optical fiber RF link channel with the remaining distance (18 kilometers) is used as the eavesdropping channel. In the present invention, in order to achieve adversarial training during the training process, an eavesdropping channel with a distance close to the 20-kilometer main channel is deliberately selected: the 18-kilometer optical fiber RF link channel.

[0050] Step 4: Based on the channel model obtained in step 3, an end-to-end transmission system is constructed. The system architecture is as follows: Figure 2 (A3) In this system, the symbol mapping and upsampling modules are replaced by a neural network model. The traditional demodulation algorithm needs to perform spectrum compensation, down-conversion, down-sampling, matched filtering, phase compensation, equalization and symbol decoding in sequence (see Figure 2 (A2)) to demodulate the received RF signal into bit information. In this system, a neural network model called ReceivNN is used to integrate and simplify these demodulation processes. Figure 2 The receiving end of (A3) is modeled using the ReceivNN model. The system architecture is composed of the TransNN, matched filtering, up-conversion, ChannelNN and ReceivNN modules connected in series, and finally an end-to-end optical radio frequency transmission system based on neural network is constructed.

[0051] In this system, both the symbol encoding rules and upsampling rules are embedded in the neural network model TransNN. Due to the "black box" nature of neural networks, these rules cannot be described by explicit analytical expressions. Neural network models are usually composed of a large number of parameters, and there are complex nonlinear relationships between these parameters. As the depth and width of the network increase, the complexity of the model also increases. Even if the attacker obtains some parameters or model structure, it is still difficult to deduce the complete model behavior. In addition, the training process of the neural network relies on a large amount of random initialization and data perturbations. Even if the training is repeated on the same task, the model will produce different results due to initialization and training perturbations. This randomness further increases the difficulty for eavesdroppers to reproduce the demodulation model.

[0052] In order to increase the physical channel security characteristics of the model, adversarial training for eavesdropping channels is added during the training process of the end-to-end system. The structure of TransNN is shown in the figure below: Figure 3 As shown, its input data is bits (100). The bit information is first one-hot encoded (200) and then merged with the disturbance signal in the innermost dimension. In the present invention, the TransNN model consists of a fully connected layer and a convolutional layer (300, ..., 1000). It is worth noting that the TransNN model can also be other structures. The feature of the present invention is that the neural network model is used to replace the symbol encoding and upsampling functions in traditional communication systems. Therefore, the number of neurons in its output layer depends on the upsampling multiple. For example, if the upsampling multiple is 20, the number of neurons in the TransNN output layer is 40, and the two adjacent values ​​represent the real part and imaginary part of a complex number respectively. Receiv is a receiving end demodulation (decryption) model, which is characterized in that the number of neurons in the input layer is related to the upsampling coefficient. When the upsampling multiple is 20, the number of neurons in the input layer is 20; the output signal is the bit stream of the transmitting end, and the number of neurons is related to the modulation order. When the modulation order is 6, the number of output neurons is 6, and the range of neuron output values ​​is [0,1]. Therefore, an end-to-end confidential transmission link is constructed through neural network models such as TransNN and ReceivNN.

[0053] Step 5: This step is to train the end-to-end confidential transmission link constructed in step 4. Figure 4The figure shows the training flow chart of the end-to-end secure transmission system. As can be seen from the training flow chart, the end-to-end transmission system training includes the training and optimization of multiple models. First, the main channel model and the eavesdropping channel model are trained (B1 and B2). Then, the random bits are input into the TransNN model (B3) and processed by the main channel model and the eavesdropping channel model respectively (B4 and B5). The processed features are input into the ReceivNN model for aggregation to obtain the fitting bits (B6 and B7), and then the mean square error between the fitting bits of the main channel and the real bits and the fitting bits of the eavesdropping channel and the real bits are calculated respectively (B8 and B9). In B10, the two loss functions are combined into a loss function through different weight coefficients, and then the TransNN and ReceivNN models are jointly optimized by the optimizer (B11). Finally, it is determined whether the training cycle is completed. If so, the training is terminated, otherwise the training cycle is repeated (B12).

[0054] Figure 5 The training process of a specific confidential transmission scheme in the present invention is shown. The transmission system consists of a transmitter, a channel model and a receiver, and the learning and optimization of channel characteristics are achieved through adversarial training and forward training. First, the input transmission bit is processed by a neural network model (TransNN) for signal conversion. The processed signal passes through a matching filter and an up-conversion module to generate input signals for the main channel and the wiretap channel. The present invention assumes that there is one wiretap channel, in which the wiretap channel corresponds to an 18-kilometer fiber transmission distance wiretap path, and the main channel is a 20-kilometer fiber transmission distance optical radio frequency legal communication path. After the signal is transmitted to the receiving end, it is demodulated by a receiving neural network module (ReceivNN) and the corresponding received bits are output, including the received bits of the main channel (received bit 1) and the received bits of the wiretap channel (received bit 2). At the same time, the mean square error is used to evaluate the error between the transmitting bit of the transmitting end and the demodulated bit of the receiving end in a normal transmission system. MSE_loss is the mean square error between the transmitting bit and the receiving bit of the main channel link, and mse_loss is the mean square error between the transmitting bit and the receiving bit of the wiretap channel link. Therefore, the loss function of this model is:

[0055] loss=(1-θ)×MSE_loss-θ×mse_loss

[0056] Wherein, θ is a weight coefficient, and its value range is (0, 1). During the training process, the weight coefficient will be adjusted periodically. In the present invention, when the value of loss is less than 0, the weight coefficient θ is 0. Figure 5As shown in the figure, in order to reduce the error between the channel model and the actual channel, disturbance noise is added at both ends of the channel during the training process to make up for the lack of channel modeling accuracy. However, it should be pointed out that in the eavesdropping link, the disturbance noise is not added between the eavesdropping channel model and the ReceivNN model. The core objectives of the training can be summarized in two aspects: one is to maximize the mean square error (MSE) of the eavesdropper, thereby significantly improving the bit error rate of the eavesdropper, enhancing the confidentiality of the main channel information transmission, and effectively suppressing the information leakage of the eavesdropping channel; the second is to minimize the mean square error of the main channel transmission, improve the accuracy of the main channel transmission signal, and ensure that the receiving end can correctly demodulate the information. By continuously updating the weights of TransNN and ReceivNN, the entire system gradually adapts to the channel characteristics, achieving more efficient communication performance and more superior information security.

[0057] In order to evaluate the security of the scheme, the following two typical eavesdropping scenarios are designed:

[0058] 1) White Box Attack

[0059] In the white-box attack scenario, it is assumed that the eavesdropper fully understands the model structure of the end-to-end learning transmission system, including the transmitter neural network (TransNN) and the receiver neural network (ReceivNN). However, the eavesdropper does not obtain the specific weights of the model. In this case, the legitimate user uses its trained model (for example, TransNN0 and ReceivNN0) to communicate, and the eavesdropper retrains a set of end-to-end systems with the known model structure to generate new models (TransNN1 and ReceivNN1). Subsequently, the eavesdropper uses its trained receiving model ReceivNN1 to demodulate the transmission signal generated by the legitimate user's transmitting model TransNN0 and observes the demodulation performance of the ReceivNN1 model.

[0060] 2) Gray Box Attack

[0061] In the gray-box attack scenario, it is assumed that the eavesdropper only obtains the structural information of the receiving end model ReceivNN and also masters its weight parameters. However, the transmission channel through which the signal received by the eavesdropper passes is different from the main channel. For example, the communication channel of the legitimate user corresponds to a fiber optic transmission distance of 20 kilometers, while the fiber optic transmission distance of the eavesdropping channel is not 20 kilometers. Due to changes in channel characteristics (such as signal attenuation, noise increase, or dispersion effects caused by different lengths of optical fiber), the signal received by the eavesdropper is different from the main channel at the physical level, and observe whether the receiving performance of ReceivNN has declined.

[0062] These two scenarios simulate the attack capabilities of eavesdroppers under different levels of information possession and channel conditions, providing a comprehensive testing basis for evaluating the robustness and security of confidential transmission schemes.

[0063] Based on the above eavesdropping scenario, a simulation experiment was conducted to verify the security and reliability of information transmission in the present invention. Figure 2 In (A1), the symbol rate of the RF signal is 2GBaud, the center frequency is 10GHz, and the modulation order is 6; the emission frequency of the laser is 193.1THz, the line width is 1MHz; and the extinction ratio of the intensity modulator is 20. According to the previous description, the fiber transmission distance of the legal transmission is 20 kilometers, and it is assumed that the fiber distance of the eavesdropping channel is 18 kilometers.

[0064] In the white-box attack scenario, the system is first trained using the same channel model, training data, and model framework to generate a pair of modulation model TransNN0 and demodulation model ReceivNN0. The constellation diagram generated by TransNN0 is shown in the figure below. Figure 6 As shown. Subsequently, under the same training conditions (including channel model, training data and model framework), 6 groups of modulation and demodulation models (TransNN1, ReceivNN1; TransNN2, ReceivNN2; TransNN3, ReceivNN3; TransNN4, ReceivNN4; TransNN5, ReceivNN5; TransNN6, ReceivNN6) were further trained. The constellation diagrams generated by TransNN1 to TransNN6 in these models are shown in Figure 7 As shown, it is obvious that there are significant differences in the distribution of constellation points of these models. The reason for this difference is that compared with the simple upsampling methods (such as value replication and zero interpolation) in traditional algorithms, the modulation model based on the neural network generates different values ​​for each upsampling point. In addition, the randomness in the neural network training also leads to significant differences in the constellation diagrams generated by each round of training. These characteristics provide additional concealment and security for the system.

[0065] In order to verify the demodulation performance of different modulation models, the experiment uses the RF signal generated by the TransNN0 modulation model and inputs it into the optical carrier RF link, and then uses different demodulation models (ReceivNN1, ReceivNN2, ReceivNN3, ReceivNN4, ReceivNN5 and ReceivNN6) for demodulation. The experimental results are shown in Figure 2. Figure 8As shown in the figure, when the demodulation model does not match the modulation model, the demodulation performance decreases significantly. This shows that even under exactly the same training conditions, there is still an obvious incompatibility between the modulation model and the demodulation model generated in different training rounds. Only the RF signal generated by the TransNN0 generated by the same training can be correctly demodulated by the corresponding ReceivNN0. The experimental results fully verify the unique binding characteristics and inherent security of the confidential transmission scheme, and provide important support for high-confidential data transmission in optical radio frequency communication.

[0066] In the gray-box attack scenario, the same steps as the white-box experiment are followed to complete the end-to-end learning of the confidential transmission system training, generate the encryption encoding model TransNN7 and the decryption demodulation model ReceivNN7, and apply them to the actual optical radio frequency transmission link. Fig. 9 As shown in the figure, when the transmission channel length is 20 kilometers, the system can work normally and the demodulated bit error rate is close to 0. However, when the transmission channel length deviates from 20 kilometers, the bit error rate demodulated by the ReceivNN7 model increases sharply. This experimental result shows that this method has good physical layer security characteristics when the channel conditions change, can effectively resist potential gray box attack threats, and provides important protection for the security of the communication system.

[0067] In summary, the hybrid drive of the present invention combines the data-driven and communication model-driven methods, builds a system framework with neural networks as the core, and introduces classic models in the communication field (such as matched filtering and up-conversion models) to achieve end-to-end adaptive optimization. This solution utilizes the adaptive feature extraction capabilities and inherent security characteristics of neural networks, which not only significantly improves the transmission performance of the system, but also achieves effective protection of transmitted data.

[0068] In this scheme, the RF over fiber transmission system is modeled as an end-to-end neural network transmission system. The transmitter and receiver models learn the channel characteristics through neural network adaptive learning, thereby dynamically adjusting the encoding rules of the transmitter and the demodulation rules of the receiver. This adaptive capability effectively improves the receiving sensitivity of the RF over fiber transmission link. At the same time, for non-target receivers, the encoding rules and decoding rules become "black boxes" due to their high complexity and are difficult to crack. In addition, since the encoder and demodulator of the neural network are jointly trained, their outputs are bound - only the demodulator that has been trained synchronously can correctly decode the signal generated by the corresponding encoder. This is because randomness is introduced during the training process of the end-to-end system, and the encoders and demodulators generated in different training rounds are different. This means that even if an eavesdropper obtains a pair of encoding rules in the system, it cannot be used to decode the transmission signals generated by other pairs of encoding rules. At the same time, the scheme also has the binding characteristics of the channel, that is, when the channel conditions change, even if the eavesdropper obtains the existing demodulation model, it cannot correctly decode the transmission signal. This is because the encoding and decoding rules of the system are highly coupled with the specific channel environment, and the adaptability of the model is only valid under the current channel conditions. When the channel environment changes, even if the eavesdropper has a demodulation model, its demodulation capability will fail due to the lack of synchronous training, thus further achieving physical layer security.

[0069] In general, this secure transmission scheme uses the nonlinear mapping characteristics of neural networks and their ability to adapt to channel characteristics to provide strong physical layer security for optical radio frequency communications. This innovative method that combines binding characteristics with physical layer encryption provides a new idea for highly secure data transmission.

[0070] The above embodiments help those skilled in the art to better understand the present invention, but do not impose any form of limitation on the present invention. It is worth noting that, without departing from the concept of the present invention, those skilled in the art can still make various changes and improvements, and these changes and improvements are within the scope of protection of the present invention.

Claims

1. An end-to-end optical radio frequency adaptive secure transmission method based on hybrid drive, characterized in that: The following steps are involved: Step 1: Build an optical radio frequency transmission link and conduct a link transmission experiment; Step 2: Collect signal data, i.e., the RF signal at the transmitting end of the link and the RF signal at the receiving end; Step 3: Construct a channel model ChannelNN based on a neural network, and use the signal data obtained in step 2 to train the channel model ChannelNN; Step 4: Establish the transmitter modulation model TransNN, the receiver demodulation model ReceivNN and the trained channel model ChannelNN to form an end-to-end secure transmission system; Step 5: Train the end-to-end secure transmission system model. The training process is as follows: (1) Generate a random bit sequence as a training set for the end-to-end secure transmission system; (2) inputting the generated random bit sequence into the end-to-end learning secure transmission system; (3) Calculate the mean square error (MSE) between the fitting bits and the training bits output by the legitimate system, and the mean square error (MSE) between the fitting bits and the training bits output by the eavesdropped channel; (4) Constructing the loss function: The mean square error is used as the loss function for the forward training, and the negative mean square error is used as the loss function for the adversarial training. Finally, the two loss functions are weighted and fused to form the total loss function. The weighted loss function is jointly optimized by the optimizer, thereby simultaneously optimizing the performance of the transmitter modulation model TransNN and the receiver demodulation model ReceivNN. (5) cyclically training the secure transmission system model until a preset number of training cycles or a performance target is reached; Step 6: After training, the modulation model TransNN is used as the encryption module, and the demodulation model ReceivNN is used as the decryption module at the receiving end. The RF signal generated by the encryption module is input into the optical carrier RF link for transmission, and the received RF signal is demodulated and tested using the ReceivNN model.

2. According to claim 1, a hybrid-driven end-to-end optical radio frequency adaptive secure transmission method is characterized in that: The input of the modulation model TransNN is bit information; wherein the input bit needs to be one-hot encoded; the modulation model TransNN is a neural network model composed of a fully connected layer and a convolutional network; the number of neurons in the input layer of TransNN is related to the modulation order M, which is 2 M The number of neurons in the output layer of TransNN is twice the upsampling factor.

3. The hybrid-driven end-to-end radio frequency over fiber adaptive secure transmission method according to claim 1, characterized in that: The demodulation model ReceivNN is composed of a neural network model, in which the number of neurons in the input layer is equal to the upsampling coefficient, and the number of neurons in the output layer is equal to the modulation order; At the same time, the fitting bit value output by the demodulation model needs to be between 0 and 1; the improved hyperbolic tangent activation function Tanh is used in the output layer of the demodulation model ReceivNN, and its expression is: In the formula, e x and e -x is a natural exponential function.

4. The hybrid-driven end-to-end RF-over-optical adaptive secure transmission method according to claim 1, characterized in that: During the training process, an adversarial training strategy for eavesdropping channels was introduced, and the adversarial training was combined with the forward training of legitimate channels through a loss function to form a weighted loss function. The transmitter modulation model and the receiver demodulation model were jointly optimized through the optimizer to improve the security and transmission performance of the system.

5. The hybrid-driven end-to-end RF-over-optical adaptive secure transmission method according to claim 1, characterized in that: The loss function is constructed by the mean square error MSE_loss between the fitting bit and the transmitted bit output by the legitimate channel and the mean square error mse_loss between the fitting bit and the transmitted bit output by the eavesdropping channel. The two are weighted and summed according to different weight coefficients. The expression is: loss=(1-θ)×MSE_loss-θ×mse_loss (2) Among them, θ is the weight coefficient, and its value range is (0, 1).

6. The hybrid-driven end-to-end radio frequency over fiber adaptive secure transmission method according to claim 5, characterized in that: The weight coefficient θ in the loss function is adjusted according to the size of the loss value during the training process.

7. The hybrid-driven end-to-end radio frequency over fiber adaptive secure transmission method according to claim 4, characterized in that: During the training process, disturbance noise is added to the front and rear ends of the main channel model, and disturbance noise is added to the front end of the eavesdropping channel model.

8. The hybrid-driven end-to-end radio frequency over fiber adaptive secure transmission solution according to claim 1, characterized in that: In the entire end-to-end system, in addition to the neural network model, the matched filter model and up-conversion model of the traditional communication model are also required.

Images