WDM-RoF secure transmission method based on end-to-end learning
Through the deep learning channel model and neural network model based on end-to-end learning, the deep coupling of the modem module is built, which solves the security problem of the WDM-RoF system and realizes high security and reliability optical communication transmission.
Patent Information
- Application Number
- CN202510708983.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-29
- Publication Date
- 2025-08-15
AI Technical Summary
The existing WDM-RoF system has risks of optical power leakage and cross-channel data leakage in terms of security. The traditional passive defense strategy is insufficient and cannot meet the security requirements of high security demand scenarios.
Using an end-to-end learning method, an end-to-end secure transmission system is built through deep learning channel model and neural network model, deep coupling of modem modules is realized, and the modulation and demodulation model is optimized using dynamic weight joint loss function to form channel binding and dynamic interlocking relationships, and system security is enhanced.
Without adding additional physical devices, the security and reliability of the WDM-RoF system is significantly improved, effectively preventing inter-channel crosstalk and data leakage, and improving the system's anti-reverse engineering capabilities.
Smart Images
Figure CN120498524A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of optical communications, and in particular relates to a WDM-RoF secure transmission method based on end-to-end learning. Background Art
[0002] Radio over fiber (RoF) transmission technology, with its technical advantages such as low transmission loss, wide spectrum support, and simple architecture, demonstrates excellent electromagnetic interference suppression capabilities while being cost-effective to build. It has now become a core supporting technology for 5G / 6G mobile communication fronthaul networks and future wireless access systems. In particular, the WDM-RoF system, constructed in conjunction with wavelength division multiplexing (WDM) technology, transmits multiplexed light waves carrying different modulation information over a single optical fiber. While maintaining the advantages of traditional RoF, it significantly increases system capacity compared to single-wavelength solutions. This technology's multi-channel independent optical modulation mechanism not only flexibly adapts to diverse business scenarios, but also establishes the physical layer foundation for achieving terabit-level ultra-large-capacity optical and wireless converged transmission. It is the core driving force behind the current evolution of access network technology and the construction of a new generation of broadband optical and wireless access networks.
[0003] Existing WDM-RoF systems suffer from fundamental security flaws: due to the non-ideal isolation characteristics of the demultiplexing components at the receiving end, optical power leakage occurs when the optical wavelengths of each channel are demodulated and separated. This inter-channel crosstalk effect poses a dual security risk: first, the leakage signal superimposes with the target channel signal, causing in-band interference, resulting in a degradation of the system's bit error rate; more importantly, when transmitting sensitive information, this phenomenon can induce the risk of cross-channel data leakage. In high-security scenarios such as intersatellite and tactical communications, traditional passive defense strategies that rely on component optimization are no longer sufficient, and active protection mechanisms at the system architecture level are urgently needed. Summary of the Invention
[0004] In response to the above technical bottlenecks, the present invention provides a WDM-RoF secure transmission method based on end-to-end learning.
[0005] A WDM-RoF secure transmission method based on end-to-end learning of the present invention comprises the following steps:
[0006] Step 1: Build a WDM-RoF system and conduct transmission experiments.
[0007] Step 2: Systematically collect the RF signals of the transmit and receive ends of each wavelength link to construct a wavelength-characterized channel modeling training dataset.
[0008] Step 3: Establish a deep learning channel model ChannelNN and complete differentiated training of each wavelength channel model based on the data obtained in step 2.
[0009] Step 4: Build an end-to-end secure transmission system consisting of the transmit modulation model TransNN, the receive demodulation model ReceivNN, and the pre-trained channel model ChannelNN. Perform target wavelength training according to the following process (the channel transmitting the target wavelength is called the target channel, and the channel transmitting the non-target wavelength is called the non-target channel):
[0010] (1) Generate a random bit stream as a training dataset.
[0011] (2) Import the training data into the system model.
[0012] (3) A dual evaluation metric is established by calculating the mean square error of the target system output (forward training loss) and the mean square error of the non-target channel output (adversarial training loss).
[0013] (4) Construct a dynamic weight joint loss function and use the optimizer to perform collaborative optimization of TransNN and ReceivNN, where the weight factor is adaptively adjusted according to the training process.
[0014] (5) Iterate the training until the preset termination condition is reached.
[0015] Step 5: After completing the single wavelength link training, repeat the process in step 4 until the working wavelength is covered.
[0016] Step 6: Deploy the trained encryption and demodulation modules for each wavelength to the WDM-RoF system and run the corresponding ReceivNN model to perform multi-channel demodulation testing.
[0017] Furthermore, the network structure of the transmit modulation model TransNN is characterized by an input layer dimension equal to the modulation order, and an output layer dimension equal to twice the upsampling coefficient. The hidden layer can be a fully connected layer or other neural network model. In the present invention, the hidden layer is a convolutional neural network layer.
[0018] Furthermore, the structural characteristics of the receiving demodulation model ReceivNN are as follows:
[0019] a) The input dimensions match the upsampling coefficients.
[0020] b) The output dimension is the base-2 logarithm of the modulation order.
[0021] c) The fitted bit value output by the demodulation model is in the interval [0,1].
[0022] Furthermore, the training process is divided into two stages:
[0023] (1) Target channel performance optimization stage.
[0024] 1) The original bit stream is injected into the TransNN module (with integrated symbol mapping-upsampling function), and the RF signal is generated through baseband shaping filtering and RF up-conversion.
[0025] 2) The RF signal is passed to the data-driven target channel model, and finally the ReceivNN performs waveform demodulation.
[0026] 3) Optimize the target channel transmission performance by minimizing the mean square error loss between the demodulated bits and the transmitting bits. In this stage, gradient reverse propagation is used to achieve the best match between the waveform characteristics of the transmitting end and the nonlinear mapping of the receiving end to the target channel transmission characteristics.
[0027] (2) Non-target channel security enhancement stage.
[0028] 1) The baseband signal generated by TransNN is converted into a RF waveform through the same modulation process;
[0029] 2) Inputting the RF signal into the pre-trained non-target channel simulation model for transmission;
[0030] 3) Create a signal aliasing effect by maximizing the mean square error loss between the demodulated bits and the original bits. This MSE loss maximization strategy significantly degrades the signal demodulation capability of the illegal channel.
[0031] Furthermore, during the training process, disturbance noise is added to both the front-end and back-end of the target channel model, and disturbance noise is added to the front-end of the non-target channel model.
[0032] Furthermore, in order to achieve a dynamic balance between security and reliability, the present invention constructs a composite loss function. The loss function is constructed by the mean square error (target_mesloss) between the fitting bits output by the target channel and the transmitted bits, and the mean square error (nontarget_mseloss) between the fitting bits output by the non-target channel and the transmitted bits. The two are weighted and summed according to different weight coefficients. The expression is:
[0033] loss=ω×target_mseloss-(1-ω)×nontarget_mseloss
[0034] Among them, ω is the weight coefficient, and its value range is (0, 1); the weight coefficient ω will be dynamically modulated according to the size of the loss value during the training process.
[0035] Furthermore, it is necessary to complete the construction of the end-to-end transmission system model for all working wavelength channels in parallel and to train in parallel to obtain the transmission model group with bound wavelength characteristics.
[0036] Compared with the prior art, the present invention has the following beneficial technical effects:
[0037] The present invention innovatively proposes a collaborative optimization method based on adaptive symbol coding, dynamic upsampling rules and demodulation, and successfully realizes the physical layer security transmission of wavelength division multiplexing-optical radio frequency link without adding additional physical devices. It is particularly noteworthy that this method is essentially different from traditional encryption schemes: through a deep coupling encoding and decoding mechanism, the modulation module and the encryption module are combined into one, and the demodulation module synchronously implements the decryption function, thereby avoiding the complex computational process of key distribution and management in traditional encryption schemes. On the technical level, by fully exploring the inherent security of the nonlinear feature space of the neural network, a coding and decoding model with extremely strong anti-reverse engineering capabilities is constructed. This invention can effectively enhance the security and reliability of the transmission process. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 This is a flow chart of the wavelength division multiplexing-optical radio frequency secure transmission method based on end-to-end learning of the present invention.
[0039] Figure 2 Block diagram of the secure WDM-RF over fiber transmission architecture based on end-to-end learning. (B1) WDM-RF over fiber link architecture. (B2) Traditional WDM-RF over fiber link architecture. (B3) WDM-RF over fiber secure transmission link based on end-to-end learning.
[0040] Figure 3 This is a structural diagram of the transmitter encryption-modulation model TransNN.
[0041] Figure 4 This is the training flow chart of the WDM-RF over Optical Fiber secure transmission model based on end-to-end learning.
[0042] Figure 5 This is a training block diagram for the WDM-RF over Optical Fiber secure transmission model based on end-to-end learning.
[0043] Figure 6 This is a heat map of point-to-point communication security performance.
[0044] Figure 7 The bit error rate results of different TransNN-ReceivNN deployments under different wavelength transmission links. DETAILED DESCRIPTION
[0045] The present invention will be explained in more detail below with reference to the accompanying drawings and specific embodiments.
[0046] The process of the wavelength division multiplexing-optical radio frequency secure transmission method based on end-to-end learning of the present invention is as follows: Figure 1 As shown, the specific steps include:
[0047] Step 1: Build a standard WDM-RoF transmission system and conduct transmission experiments. The link structure is as follows Figure 2 As shown in (B1), the system consists of a transmitter, a fiber channel, and a receiver. The transmitter includes an arbitrary waveform generator, a multi-wavelength light source array, an intensity modulator, and a wavelength division multiplexer. Each light source generates an optical carrier of a different wavelength. The arbitrary waveform generator generates a radio frequency signal that is loaded onto the corresponding optical carrier through the intensity modulator. The wavelength division multiplexer synthesizes the multi-wavelength signal and inputs it into the fiber channel. The receiver is equipped with a demultiplexer, a photodetector array, an amplifier, and an oscilloscope to achieve wavelength separation, photoelectric conversion, signal amplification, and digital acquisition.
[0048] Step 2: Collect RF signals from the transmitter and receiver of each wavelength link in the system to construct a wavelength-characterized channel model training dataset. The transmitter uses single-carrier quadrature amplitude modulation (SC-QAM) to generate the RF signal. An oscilloscope is used to capture the distorted signal after fiber transmission. This signal comprehensively characterizes signal impairments such as optoelectronic device response, fiber dispersion, and nonlinear effects. Training data is constructed independently by wavelength, with the number of samples strictly corresponding to the number of wavelength division multiplexing channels. That is, when the system uses N wavelength multiplexing, N datasets containing corresponding wavelength transmit-receive signal pairs will be generated. This design effectively supports independent modeling of multi-wavelength signal characteristics.
[0049] Step 3: Establish a deep learning-based signal model, ChannelNN, and complete differentiated training of each wavelength channel model based on the data obtained in Step 2. A series of ChannelNN models (ChannelNN1-ChannelNN4) are constructed for wavelengths λ1, λ2, λ3, and λ4, respectively. By minimizing the mean square error between the transmitted signal and the distorted signal, the stochastic gradient descent algorithm is used to optimize the network parameters to achieve accurate modeling of the transmission characteristics of each wavelength channel.
[0050] Step 4: Build an end-to-end transmission model and train it. The end-to-end transmission model is built by cascading the transmitter TransNN modulation model, shaping filter, up-conversion module, ChannelNN channel model and the receiver ReceivNN demodulation model. Its innovative architecture is as follows: Figure 2(B3) As shown in the figure. The present invention uses a neural network to optimize the entire process of physical layer signal processing: at the transmitting end, the symbol mapping and upsampling model is replaced by a neural network model TransNN model, which can adaptively optimize the constellation shape after upsampling; at the receiving end, the traditional demodulation process (including Figure 2 Frequency compensation, down-conversion, and matched filtering (shown in (B2)) are used to achieve end-to-end nonlinear mapping of received signals to bits. ChannelNN models based on different wavelengths are trained in parallel for the corresponding end-to-end transmission systems.
[0051] In the present invention, symbol encoding and upsampling rules are implicitly represented through the TransNN neural network. This model has typical black box characteristics, and its input-output mapping cannot be accurately described by traditional closed-form formulas. From the perspective of information security, even if the attacker reversely obtains the neural network structure or part of the weight parameters, the retrained model still cannot demodulate the previously trained TransNN signal. This security feature stems from the deep coupling mechanism of the end-to-end system: the modulation and encryption module and the demodulation and decryption module form a dynamic interlocking relationship during the collaborative optimization process. The models generated by different training rounds produce uncontrollable offsets in the constellation diagram distribution due to random initialization and noise perturbations. This endogenous randomness significantly increases the difficulty for eavesdroppers to reconstruct the demodulation model.
[0052] like Figure 3 As shown in the structural diagram of the TransNN model, the input data is a bit sequence (D1). The bits are first one-hot encoded (D2). The input layer (D3) and output layer (D5) of the TransNN model use fully connected neural networks, while the hidden layer (D4) is composed of four layers of convolutional neural networks, which are responsible for feature extraction and nonlinear mapping at different levels. It is worth noting that the hidden layer structure of the TransNN model is highly scalable and can adopt neural networks of other types and structures. The innovation of the present invention is that the neural network model is intended to replace the symbol encoding and upsampling functions in traditional communication systems. The number of neurons in the TransNN input layer is related to the modulation order. For example, when the modulation order is 64, the number of neurons in the input layer is set to 64. The number of neurons in the output layer is determined by the upsampling factor. For example, if the upsampling factor is 20, the number of neurons in the output layer of the TransNN is 40, where two consecutive numerical values represent the real part and imaginary part of a complex number, respectively. In addition, the input and output dimensions of the ChannelNN model are equal to the upsampling coefficient. The input dimension of the ReceivNN model is the upsampling coefficient, and the output dimension is the base-2 logarithm of the modulation order. For example, if the modulation order is 16, the output dimension is 4. In addition, the output value must be limited to the interval [0, 1].
[0053] After building the end-to-end transmission model, it is necessary to train the model. Figure 4 The figure shows the flow chart for collaborative training of an end-to-end transmission model. This flow illustrates that training an end-to-end transmission system involves the joint training and optimization of multiple models. Specifically, the target channel model (C1) and the non-target channel model (C2) are first trained separately. Subsequently, a randomly generated bit sequence is input into the TransNN model (C3) and processed by the target and non-target channel models (C4 and C5). The processed signal features are then sent to the ReceivNN model for demodulation, resulting in fitted bits (C6 and C7). The mean squared error (MSE) between the fitted bits of the target channel and the input bits (C8) and the mean squared error (C9) between the fitted bits of the non-target channel and the input bits are then calculated. In step C10, the two loss functions are weighted and combined using different weight coefficients to form the final loss function. Subsequently, the TransNN and ReceivNN models are jointly optimized by the optimizer (C11). Finally, the system determines whether the training cycle is complete. If so, training ends; otherwise, the training process repeats (C12).
[0054] Figure 5 The secure training architecture shown here consists of three core modules: the TransNN neural network model at the transmitter performs symbol encoding and upsampling. The channel model implements the physical loss model and gradient propagation, and the ReceivNN at the receiver is responsible for signal modulation and decryption. In specific implementation, the input transmit bits first pass through the neural network model (TransNN) for signal processing. The processed signal passes through the shaping filter and upconversion module in sequence to generate the input signals for the target channel and non-target channels. After the signal is transmitted to the receiver, it is demodulated by the receiving neural network module (ReceivNN) and the corresponding received bits are output, including the received bits of the target channel and the received bits of the non-target channel. At the same time, the mean square error (MSE) is used to evaluate the error between the transmitted bits at the transmitter and the demodulated bits at the receiver in a normal transmission system. Here, target_mseloss represents the mean square error between the transmitted bits and the received bits of the main target channel link, and nontarget_mseloss represents the mean square error between the transmitted bits and the received bits of the non-target channel link. Therefore, the loss function of this model is:
[0055] loss=ω×target_mseloss-(1-ω)×nontarget_mseloss
[0056] Wherein, ω 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 ω is 0.
[0057] like Figure 5 As shown in the figure, during the training process, in order to reduce the error between the channel model and the actual channel, perturbation noise is introduced at both ends of the channel to compensate for the lack of channel modeling accuracy. It should be noted that in the non-target link, no perturbation noise is added between the eavesdropping channel model and the ReceivNN model. The core objectives of the training mainly cover two aspects: first, maximizing the mean square error (MSE) of the non-target channel, significantly improving the eavesdropper's bit error rate, enhancing the confidentiality of information transmission on the main channel, and effectively curbing information leakage on the non-target channel; second, minimizing the mean square error of the target channel transmission, improving the accuracy of the main channel transmission signal, and ensuring that the receiving end can accurately demodulate the information. By continuously updating the weights of TransNN and ReceivNN, the entire system can gradually adapt to the channel characteristics, achieving more efficient communication performance and more reliable information security.
[0058] Step 5: After completing single-wavelength link training, the process of step 4 is repeated until all operating wavelengths are covered. For the WDM-RoF system, this invention sets four characteristic wavelengths, λ1 to λ4, to form an optical carrier array and establishes a dynamic channel pairing mechanism: when λ1 is the target channel, λ2 automatically becomes a non-target channel; when λ2 is the target channel, it is associated with λ1 and λ3; and so on, forming four sets of parallel optimized end-to-end transmission models.
[0059] Step 6: Deploy the trained wavelength encryption demodulation modules to the WDM-RoF system, and run the corresponding ReceivNN model to implement multi-channel channel demodulation testing. The WDM-RoF transmission system proposed in the present invention adopts an innovative architectural design to achieve highly secure point-to-point communication. In the system, optical signals with wavelengths of λ1, λ2, λ3, and λ4 correspond to four transmitting terminals (Tx1-Tx4) and their corresponding four receiving terminals (Rx1-Rx4). Through precise end-to-end transmission model training, the system obtains the TransNN and ReceivNN neural network modules, and deploys them according to the wavelength allocation principle: TransNN1 to TransNN4 are installed on Tx1 to Tx4, respectively, and ReceivNN1 to ReceivNN4 are deployed on Rx1 to Rx4, forming an orthogonal mapping relationship in the wavelength dimension.
[0060] The present invention constructs a point-to-point secure transmission mechanism based on channel binding, the core feature of which lies in the strong coupling relationship established between TransNN and ReceivNN. Taking Tx1 as an example, the specific coded signal generated by TransNN1 can only be effectively demodulated in ReceivNN1 of Rx1, reflecting significant security features: (1) The unpaired ReceivNN model in the same receiving terminal (such as ReceivNN2 deployed in Rx1) cannot demodulate the signal from TransNN1. (2) When TransNN1 is deployed in Tx2 and transmitted via the λ2 wavelength, the ReceivNN models of all receiving terminals (including ReceivNN1 to ReceivNN4) cannot achieve effective signal demodulation. This innovative security mechanism achieves system security through two dimensions: (1) The modulation and demodulation models have a dynamic interlocking relationship: ensuring that only neural network models that have been jointly trained can recognize and communicate with each other. (2) Wavelength-model deep binding: the model performance is tightly bound to the specific wavelength channel.
[0061] Simulation experiment:
[0062] To verify the above conclusions, this study designed a system simulation experiment with the following specific parameter configurations:
[0063] (1) Figure 2 In (B1), the RF signal uses a 64-QAM modulation order, a symbol rate of 2 GBaud, and a center frequency of 10 GHz. (2) The laser transmitter is equipped with four wavelength channels (1552.393 nm, 1552.795 nm, 1553.197 nm, and 1553.599 nm), corresponding to the transmit-receive link transmit-receive end 1 to transmit-receive end 4, with a channel spacing of 50 GHz and a laser linewidth of 1 MHz. (3) The extinction ratio of the intensity modulator is 20 dB, and the responsivity of the photodetector is 1. This multi-channel experimental architecture effectively simulates the system operating state in a wavelength division multiplexing environment.
[0064] The corresponding TransNN models are deployed at different transmitting and receiving ends, and different ReceivNN models are used to demodulate the received signals. The demodulation performance is as follows: Figure 6 As shown in the figure, only matching transmit-receive model combinations (such as TransNN1-ReceivNN1) can achieve correct demodulation. When the signal sent by TransNN1 is processed by other demodulation models (ReceivNN2-ReceivNN4), the bit error rate is close to 0.5. This shows that different models are not compatible and cannot correctly demodulate each other's signals.
[0065] Figure 7 (a) and Figure 7(b) A comprehensive demonstration of the bit error rate (BER) performance of four neural network transceiver systems under different transmission-reception configurations, reflecting one of the core security features of the present invention. Each neural network transceiver combination exhibits significant pairing specificity and channel dependence. TransNN1-ReceivNN1 achieves optimal demodulation performance in the Tx1-Rx1 configuration, TransNN2-ReceivNN2 in the Tx2-Rx2 configuration, TransNN3-ReceivNN3 in the Tx3-Rx3 configuration, and TransNN4-ReceivNN4 in the Tx4-Rx4 configuration, with BERs reaching the order of 1e-5, far exceeding the 7% FEC limit requirement. Under non-optimal configurations, the bit error rates of these systems are generally in the order of 1e-1, which is lower than the communication system requirements.
[0066] These experimental results strongly validate the system's dual-factor security mechanism: On the one hand, unpaired neural network models cannot effectively demodulate signals, resulting in high bit error rates. On the other hand, even for paired neural network models, demodulation performance is significantly degraded if their transmission wavelengths do not match (for example, when TransNN1 is transmitted over a wavelength other than λ1). This dual protection mechanism, based on pairing constraints for neural network modulation and demodulation models and wavelength-model binding, effectively ensures the security of point-to-point transmission and provides an innovative security solution for WDM-RoF systems.
[0067] The above embodiments are intended to help those skilled in the art to better understand the present invention, but are not intended to limit the present invention in any way. It should be noted that, without departing from the spirit of the present invention, those skilled in the art may make various changes and improvements, which are all within the scope of protection of the present invention.
Claims
1. A WDM-RoF secure transmission method based on end-to-end learning, characterized in that: The following steps are involved: Step 1: Build a WDM-RoF system and conduct transmission experiments; Step 2: Systematically collect the RF signals of the transmit and receive ends of each wavelength link to construct a wavelength-characterized channel modeling training dataset; Step 3: Establish a deep learning channel model ChannelNN and complete differentiated training of each wavelength channel model based on the data obtained in step 2; Step 4: Build an end-to-end secure transmission system consisting of the transmit modulation model TransNN, the receive demodulation model ReceivNN, and the pre-trained channel model ChannelNN. Perform target wavelength training according to the following process: (1) Generate a random bit stream as a training data set; (2) Importing training data into the system model; (3) Establish a dual evaluation index by calculating the mean square error of the target system output and the mean square error of the non-target channel output; (4) Construct a dynamic weighted joint loss function and use the optimizer to implement collaborative optimization of TransNN and ReceivNN, where the weight factor is adaptively adjusted according to the training process; (5) Iterate the training until the preset termination condition is reached; Step 5: After completing the single wavelength link training, repeat the process in step 4 until the working wavelength is covered; Step 6: Deploy the trained encryption and demodulation modules for each wavelength to the WDM-RoF system and run the corresponding ReceivNN model to perform multi-channel demodulation testing.
2. A WDM-RoF secure transmission method based on end-to-end learning according to claim 1, characterized in that: The network structure characteristics of the transmission modulation model TransNN are as follows: the dimension of the input layer is equal to the modulation order, and the dimension of the output layer is twice the upsampling coefficient.
3. The WDM-RoF secure transmission method based on end-to-end learning according to claim 1, characterized in that: The structural characteristics of the receiving demodulation model ReceivNN are as follows: a) The input dimension matches the upsampling coefficient; b) The output dimension is the base-2 logarithm of the modulation order; c) The fitted bit value output by the demodulation model is in the interval [0,1].
4. The WDM-RoF secure transmission method based on end-to-end learning according to claim 1, characterized in that: The training process is divided into two stages: (1) target channel performance optimization stage; (2) non-target channel security enhancement stage; During the training process, disturbance noise is added to both the front-end and back-end of the target channel model, and disturbance noise is added to the front-end of the non-target channel model.
5. The WDM-RoF secure transmission method based on end-to-end learning according to claim 1, characterized in that: The loss function is constructed by: the mean square error target_mesloss between the fitting bits output by the target channel and the transmitted bits and the mean square error nontarget_mseloss between the fitting bits output by the non-target channel and the transmitted bits, and the two are weighted and summed according to different weight coefficients. The expression is: loss=ω×target_mseloss-(1-ω)×nontarget_mseloss Among them, ω is the weight coefficient, and its value range is (0, 1); the weight coefficient ω will be dynamically modulated according to the size of the loss value during the training process.
6. The WDM-RoF secure transmission method based on end-to-end learning according to claim 1, characterized in that: It is necessary to complete the construction of the end-to-end transmission system model for all working wavelength channels in parallel, and to train in parallel to obtain the transmission model group with bound wavelength characteristics.