Communication method and device

Abstract

The present application relates to a communication method and a device. The method comprises: receiving first information; and inputting the first information into a first target model to obtain a first compressed codeword output by the first target model after performing de-perturbation on the first information.

Description

Communication methods and devices Technical Field

[0001] This application relates to the field of communications, and more specifically, to a communication method and apparatus. Background Technology

[0002] In large-scale multiple-input multiple-output (MIMO) communication systems, channel information feedback is crucial to system performance. To reduce feedback overhead and improve accuracy, channel feedback neural networks (CSINet) are introduced into the feedback process. In this approach, an encoder is deployed at the transmitting end to encode compressed codewords, and a decoder is deployed at the receiving end to recover or reconstruct the channel information. However, the compressed codewords transmitted between the encoder and decoder are vulnerable to adversarial attacks. Therefore, ensuring accurate channel information reconstruction at the receiving end even when the compressed codewords are subjected to such attacks during transmission becomes a critical problem.

[0003] Summary of the Invention

[0004] This application provides a communication method and device.

[0005] This application provides a communication method executed by a first device, comprising:

[0006] Receive the first message;

[0007] The first information is input into the first target model to obtain the first compressed codeword output by the first target model after the first information has been de-perturbed.

[0008] This application provides a communication method executed by a second device, comprising:

[0009] The channel information between the first device and the second device is dimensionally adjusted to obtain the adjusted channel information.

[0010] The adjusted channel information is encoded to obtain the second compressed codeword;

[0011] Send the second compressed codeword.

[0012] This application provides a communication method executed by a second device, comprising:

[0013] Receive air interface signals;

[0014] The air interface signal is input into the second target model to obtain the denoised channel information between the second device and the first device, which is output by the second target model.

[0015] This application provides a first device, including:

[0016] The first communication unit is used to receive the first information;

[0017] The first processing unit is configured to input the first information into the first target model and obtain the first compressed codeword output by the first target model after de-perturbation of the first information.

[0018] This application provides a second device, including:

[0019] The second processing unit is used to perform dimensional adjustment on the channel information between the first device and the second device to obtain adjusted channel information; and to encode the adjusted channel information to obtain a second compressed codeword.

[0020] The second communication unit is used to send the second compressed codeword.

[0021] This application provides a second device, including:

[0022] The second communication unit is used to receive air interface signals;

[0023] The second processing unit is used to input the air interface signal into the second target model to obtain the denoised channel information between the second device and the first device output by the second target model.

[0024] By adopting the above scheme, after receiving the first information, the first device obtains the denoised compressed codewords of the first information through the first target model. Thus, even if the compressed codewords are subjected to adversarial perturbation attacks during transmission, the first device can obtain the denoised compressed codewords through the first target model. This ensures the accuracy of the compressed codewords obtained by the first device as the receiving end, thereby guaranteeing the accuracy of the channel information reconstructed by the first device as the receiving end based on the compressed codewords, and ultimately ensuring the quality of wireless communication. Attached Figure Description

[0025] Figure 1 is a schematic diagram of an application scenario according to an embodiment of this application.

[0026] Figure 2 is a schematic flowchart of a communication method according to an embodiment of this application.

[0027] Figure 3 is a schematic flowchart of a communication method according to another embodiment of this application.

[0028] Figure 4 is a schematic flowchart of a communication method according to another embodiment of this application.

[0029] Figure 5 is a schematic flowchart of a communication method according to another embodiment of this application.

[0030] Figure 6 is a schematic diagram of a processing scenario in which dimension adjustment is performed on the encoder side according to an embodiment of this application.

[0031] Figure 7 is a schematic diagram of the training process of the first generator and the first discriminator according to an embodiment of this application.

[0032] Figure 8 is a schematic flowchart of a communication method according to an embodiment of this application.

[0033] Figure 9 is another schematic flowchart of a communication method according to an embodiment of this application.

[0034] Figure 10 is a schematic block diagram of a first device according to an embodiment of the present application.

[0035] Figure 11 is a schematic block diagram of a second device according to an embodiment of the present application. Detailed Implementation

[0036] The technical solutions of this application embodiment can be applied to various communication systems, such as: LTE (Long Term Evolution), LTE-A (Long Term Evolution-Advanced), NR (New Radio), evolution of NR, WLAN (Wireless Local Area Network), WiFi (Wireless Fidelity), or other communication systems.

[0037] This application describes various embodiments in conjunction with network devices and terminals. The terminal can be mobile or fixed, and may also be referred to as a mobile station, user unit, etc. The terminal can be a station in a WLAN, or a smart terminal, wireless modem, laptop, tablet, etc. In this application embodiment, the terminal can be a VR (Virtual Reality) terminal / AR (Augmented Reality) terminal, industrial control terminal, autonomous driving terminal, telemedicine terminal, smart grid terminal, transportation safety terminal, smart city terminal, or smart home wireless terminal, etc. By way of example and not limitation, in this application embodiment, the terminal can also be a wearable device.

[0038] In this embodiment, the network device can be a device for communicating with a terminal. The network device can be an access point in a WLAN, an evolved base station in LTE, a relay station, a network device in a vehicle-mounted device, wearable device, or NR network (gNB, the next generation Node B), or a network device in a future evolved PLMN (Public Land Mobile Network), or a network device in a non-terrestrial network, etc. By way of example and not limitation, in this embodiment, the network device can have mobility characteristics; for example, the network device can be a mobile device.

[0039] To facilitate understanding of the technical solutions of the embodiments of this application, the relevant technologies of the embodiments of this application are described below. The following relevant technologies are optional solutions and can be combined with the technical solutions of the embodiments of this application in any way, and they all fall within the protection scope of the embodiments of this application.

[0040] Figure 1 exemplarily illustrates a communication system 100. The communication system includes network devices 110 and terminals 120. In one possible implementation, the communication system 100 may include multiple network devices 110, and the coverage area of ​​each network device 110 may include one or more terminals 120; this embodiment does not limit this. In another possible implementation, the communication system 100 may also include other network entities such as mobility management entities and access and mobility management functions; this embodiment does not limit this. The network devices may further include access network devices and core network devices. That is, the communication system may also include multiple core networks for communicating with the access network devices. The access network devices may be base stations of LTE, LTE-A, or NR systems. Taking the communication system shown in Figure 1 as an example, the communication devices may include network devices and terminals with communication functions. The communication devices may also include other devices in the communication system, such as network controllers, mobility management entities, and other network entities; this embodiment does not limit this.

[0041] Figure 2 is a schematic flowchart of a communication method performed by a first device according to an embodiment of this application. The method includes at least a portion of the following.

[0042] S210, Receive the first information;

[0043] S220. Input the first information into the first target model to obtain the first compressed codeword after the first information has been de-perturbed, output by the first target model.

[0044] Figure 3 is a schematic flowchart of a communication method performed by a second device according to an embodiment of this application. The method includes at least a portion of the following.

[0045] S310. Adjust the dimension of the channel information between the first device and the second device to obtain the adjusted channel information;

[0046] S320. Encode the adjusted channel information to obtain the second compressed codeword;

[0047] S330, Send the second compressed codeword.

[0048] The first device is the receiver of the compressed codewords, and the second device is the transmitter of the compressed codewords. In some possible examples, the first device may also be alternatively referred to as the receiver or receiver device, and the second device may also be alternatively referred to as the transmitter or transmitter device.

[0049] The first device is an access network device, the second device is a terminal, and the channel information is downlink channel information; or, the first device is a terminal, the second device is an access network device, and the channel information is uplink channel information.

[0050] Specifically, in different scenarios, the first device and the second device refer to different devices. This embodiment defines uplink or downlink scenarios based on reference signals or channel information. In the following text, the downlink scenario refers to a scenario where the reference signal is a downlink reference signal and the channel information is downlink channel information. In the downlink scenario, the first device is an access network device (i.e., the access network device is the receiver), and the second device is a terminal (i.e., the terminal is the transmitter). Similarly, the uplink scenario refers to a scenario where the reference signal is an uplink reference signal and the channel information is uplink channel information. In the uplink scenario, the first device is a terminal (i.e., the terminal is the receiver), and the second device is an access network device (i.e., the access network device is the transmitter). It should be noted that the above division of uplink and downlink scenarios is based on the uplink or downlink direction of the reference signal or channel information. In some possible examples, the uplink or downlink scenario may also be defined by the transmission direction of the compressed codeword. This embodiment primarily uses the uplink or downlink transmission direction of the reference signal to define the uplink or downlink scenario for ease of explanation later, but this embodiment is not intended to limit the definition of uplink or downlink scenarios.

[0051] In some possible implementations, the processing of the second device may further include: measuring a reference signal from the first device to obtain channel information between the first device and the second device; and obtaining a second compressed codeword based on the channel information. Correspondingly, the processing of the first device may include: transmitting the reference signal.

[0052] In this implementation, after the second device obtains the second compressed codeword, it can send the second compressed codeword to the first device.

[0053] After the first device sends a reference signal, it can receive the first information, then input the first information into the first target model to obtain the first compressed codeword output by the first target model after the first information has been denoted; and then reconstruct the channel information between the first device and the second device based on the first compressed codeword.

[0054] In this way, after receiving the first information, the first device obtains the denoised compressed codewords by using the first target model. Thus, even if the transmitted compressed codewords received by the first device as the receiver are subjected to an adversarial perturbation attack, the first device can obtain the denoised compressed codewords through the first target model. This ensures the accuracy of the compressed codewords received by the first device as the receiver, thereby guaranteeing the accuracy of the channel information reconstructed by the first device based on these compressed codewords, and ultimately ensuring the quality of wireless communication.

[0055] In some embodiments, in the downlink scenario, the first device is an access network device and the second device is a terminal. The reference signal can be a downlink reference signal, and the channel information can be downlink channel information.

[0056] The terminal's processing specifically involves: measuring the downlink reference signal from the access network device to obtain downlink channel information between the terminal and the access network device; and obtaining a second compressed codeword based on the downlink channel information.

[0057] The timing of the access network device transmitting the downlink reference signal is within the protection scope of this embodiment as long as it occurs before receiving the first information. The downlink reference signal can be any one of the following: Channel State Information Reference Signal (CSI-RS), Demodulation Reference Signal (DMRS), etc. The specific types of downlink reference signals are not limited or exhaustively listed here.

[0058] The terminal measures the downlink reference signal from the access network device to obtain downlink channel information between the terminal and the access network device. This can be achieved by the terminal measuring the downlink reference signal from the access network device, performing channel estimation based on the measurement results, and obtaining downlink channel information between the terminal and the access network device.

[0059] The downlink channel information may include CSI (Channel State Information), etc.

[0060] Here, the algorithm used for channel estimation can be configured according to the actual situation, such as the least squares channel estimation algorithm, etc.

[0061] For example, the terminal measures the downlink reference signal from the access network device, performs channel estimation based on the measurement results, and obtains the downlink channel information between the terminal and the access network device. This can be achieved by the terminal measuring the downlink reference signal from the access network device using a receiver to obtain the pilot sequence of the downlink reference signal. Then, based on the pre-stored pilot sequence and the received pilot sequence (i.e., the pilot sequence of the downlink reference signal), the terminal estimates the channel information of the channel transmitting the downlink reference signal using a channel estimation algorithm (e.g., least squares channel estimation), and uses the channel information of the channel transmitting the downlink reference signal as the downlink channel information. It should be noted that this is only an exemplary description of how the terminal obtains downlink channel information by measuring the downlink reference signal. In actual processing, other methods may also be used to obtain downlink channel information, and this embodiment does not limit or exhaustively describe them.

[0062] The terminal obtains a second compressed codeword based on the downlink channel information, which may include: the terminal compressing the downlink channel information using an encoder to obtain the second compressed codeword. That is, the terminal can input the downlink channel information into the encoder to obtain the second compressed codeword output by the encoder. The encoder can be pre-configured in the terminal; the generation or training method of the encoder is not limited in this embodiment.

[0063] The processing after the terminal obtains the second compressed codeword may include: sending the second compressed codeword.

[0064] Specifically, the terminal sending the second compressed codeword may include: the terminal transmitting the second compressed codeword to the access network device via the air interface. The terminal may use any type of uplink AS (Access Stratum) message or information to carry or transmit the second compressed codeword; this embodiment does not impose any limitation.

[0065] Correspondingly, the access network device can receive the first information. In this embodiment, the terminal transmits the second compressed codeword through the air interface, while the access network device receives the first information through the air interface. This is because the second compressed codeword may be subject to adversarial attacks during the wireless channel transmission between the terminal and the access network device. That is, the information received by the access network device may be obtained by adding perturbation or interference to the second compressed codeword. Therefore, the information obtained by adding perturbation or interference to the second compressed codeword received by the access network device through the air interface is called the first information.

[0066] Furthermore, after the access network device receives the first information, it inputs the first information into the first target model to obtain the first compressed codeword output by the first target model after de-disturbance of the first information. Theoretically, the de-disturbance compressed codeword obtained by the access network device should be the same as, infinitely close to, or similar to the compressed codeword generated by the terminal. In order to distinguish between the de-disturbance compressed codeword obtained by the access network device and the compressed codeword generated by the terminal, this embodiment refers to the de-disturbance compressed codeword obtained by the access network device as the first compressed codeword and the compressed codeword generated by the terminal as the second compressed codeword, which will not be explained again below.

[0067] In this embodiment, the function of the first target model is to deperturbify the received signal or information to generate codewords with a distribution similar to or close to that of the real compressed codewords (i.e., the second compressed codewords). The first target model is a trained or successfully trained neural network with generative capabilities.

[0068] The structure of the first target model may include at least one of the following: fully connected neural network, convolutional neural network (CNN), deconvolutional neural network (DNN), recurrent neural network (RNN) (or variants of RNN, such as long short-term memory network (LSTM) and gated recurrent unit (GRU)), generative adversarial autoencoder (GAAE), etc. The possible structures of the first target model are not limited or exhaustively listed here.

[0069] The processing after the access network device obtains the first compressed codeword may further include: reconstructing the downlink channel information between the access network device and the terminal based on the first compressed codeword.

[0070] Specifically, the step of reconstructing the downlink channel information between the access network device and the terminal based on the first compressed codeword can be: inputting the first compressed codeword into the decoder to obtain the downlink channel information between the access network device and the terminal reconstructed (or restored) by the decoder.

[0071] It should be noted that the downlink channel information reconstructed by the decoder on the access network equipment side and the downlink channel information measured by the terminal from the downlink reference signal should theoretically be the same, similar, or infinitely close. In some possible examples, in order to distinguish between the downlink channel information reconstructed by the access network equipment side and the downlink channel information measured by the terminal, the downlink channel information measured by the terminal can be referred to as the original downlink channel information, or the original downlink CSI information, or the original CSI, or the downlink CSI, etc., while the downlink channel information reconstructed by the access network equipment can be referred to as reconstructed downlink channel information, or restored CSI, or reconstructed CSI information, or reconstructed original CSI, or reconstructed original CSI information, etc., which will not be explained again below.

[0072] In this embodiment, the decoder on the access network device side and the encoder on the terminal side can be pre-trained and matched with each other. This embodiment does not limit the composition, training method, etc. of the encoder and decoder.

[0073] Referring to Figure 4, an exemplary description of the communication method provided in this embodiment is provided, including:

[0074] Step 401: The UE (User Equipment, terminal) measures the downlink reference signal and obtains the raw downlink channel information (also called raw downlink channel information, downlink channel information, or raw CSI, etc.). Here, before step 401, it may also include: the base station (access network equipment) transmitting the downlink reference signal. This downlink reference signal may include CSI-RS, etc.

[0075] Step 402: The UE compresses the original downlink channel information through the encoder to obtain the second compressed codeword.

[0076] Step 403: The UE sends the second compressed codeword to the base station via the air interface.

[0077] Step 404: The base station receives the wireless signal and obtains the received first information. This wireless signal is used to transmit the second compressed codeword, but there is an adversarial perturbation added to the air interface signal by an attacker (e.g., represented as z). Therefore, the information received by the base station is represented as the first information (i.e., the information that adds perturbation to the second compressed codeword).

[0078] Step 405: The base station inputs the first information into the first target model to eliminate the influence of adversarial perturbations, obtaining the perturbated first compressed codeword. Here, the first target model can be a trained first generative model or a first generator, which can generate Y. generate That is, the first compressed codeword with adversarial perturbations removed (or the first compressed codeword with perturbations removed).

[0079] Step 406: The base station decodes the first compressed codeword Y. generate Decoding and reasoning yield the reconstructed downlink channel information (for example, it can be represented as...). ).

[0080] By adopting the above scheme, after receiving the first information, the access network device obtains the denoised compressed codewords from the first target model. Thus, even if the compressed codewords are subjected to adversarial perturbation attacks during transmission between the terminal and the access network device, the access network device can still obtain the denoised compressed codewords through the first target model. This ensures the accuracy of the compressed codewords obtained by the access network device, thereby guaranteeing the accuracy of the downlink channel information reconstructed by the access network device based on these compressed codewords, and ultimately ensuring the quality of wireless communication.

[0081] In some embodiments, in the uplink scenario, the first device is a terminal and the second device is an access network device. The reference signal can be an uplink reference signal, and the channel information can be uplink channel information.

[0082] The specific processing steps of the access network device are as follows: measuring the uplink reference signal from the terminal to obtain the uplink channel information between the terminal and the access network device; and obtaining the second compressed codeword based on the uplink channel information.

[0083] The timing of the terminal sending the uplink reference signal is within the protection scope of this embodiment as long as it occurs before receiving the first information. The uplink reference signal can be any one of the following: a sounding reference signal (SRS), a DMRS, etc. The specific types of uplink reference signals are not limited or exhaustively listed here.

[0084] The access network device measures the uplink reference signal from the terminal to obtain the uplink channel information between the terminal and the access network device. This can be achieved by the access network device measuring the uplink reference signal from the terminal, performing channel estimation based on the measurement results, and obtaining the uplink channel information between the terminal and the access network device. Here, the channel estimation algorithm used by the access network device is least squares channel estimation or other channel estimation algorithms; this embodiment does not limit or exhaustively list them.

[0085] The access network device obtains a second compressed codeword based on the uplink channel information. This can include: the access network device inputting the uplink channel information into an encoder to obtain the second compressed codeword output by the encoder. The encoder can be pre-configured in the access network device, and this embodiment does not limit the method of generating or training the encoder.

[0086] The processing of the access network device after obtaining the second compressed codeword may include: sending the second compressed codeword. The second compressed codeword may be any type of downlink AS message or information carrier or transmission, and this embodiment is not limited thereto.

[0087] Accordingly, the terminal can receive the first information. In this embodiment, the access network device transmits the second compressed codeword through the air interface, while the terminal receives the first information through the air interface. This is because the second compressed codeword may be subject to adversarial attacks during the wireless channel transmission between the terminal and the access network device. That is, the terminal may receive information obtained by adding perturbation or interference to the second compressed codeword, hence the distinction.

[0088] Furthermore, after the terminal receives the first information, it inputs the first information into the first target model to obtain the first compressed codeword output by the first target model after de-disturbance of the first information. Theoretically, the de-disturbed compressed codeword obtained by the terminal should be the same as, infinitely close to, or similar to the compressed codeword generated by the access network device. In this embodiment, the de-disturbed compressed codeword obtained by the terminal is referred to as the first compressed codeword, and the compressed codeword generated by the access network device is referred to as the second compressed codeword, mainly for the purpose of distinguishing the executing entities. This will not be explained again below.

[0089] In this embodiment, the functions and composition of the first target model are described in the same way as in the previous embodiments, and will not be repeated here.

[0090] The processing after the terminal obtains the first compressed codeword may further include: inputting the first compressed codeword into the decoder to obtain the uplink channel information between the access network device and the terminal reconstructed (or restored) by the decoder.

[0091] In this embodiment, the decoder on the terminal side and the encoder on the access network device side can be pre-trained and matched with each other. This embodiment does not limit the composition, training method, etc. of the encoder and decoder.

[0092] Referring to Figure 5, an exemplary description of the communication method provided in this embodiment is provided, including:

[0093] Step 501: The base station measures the uplink reference signal to obtain the raw uplink channel information.

[0094] Step 502: The base station compresses the original uplink channel information using an encoder to obtain the second compressed codeword.

[0095] Step 503: The base station sends the second compressed codeword to the UE via the air interface.

[0096] Step 504: The UE receives the wireless signal and obtains the received first information. This wireless signal is used to transmit the second compressed codeword, but there is an adversarial perturbation added to the air interface signal by an attacker (e.g., represented as z). Therefore, the information received by the UE is represented as the first information (i.e., the information that adds perturbation to the second compressed codeword).

[0097] Step 505: The UE inputs the first information into the first target model to eliminate the influence of adversarial perturbation and obtains the first compressed codeword after perturbation.

[0098] Step 506: The UE decodes the first compressed codeword through the decoder and infers the reconstructed uplink channel information.

[0099] By adopting the above scheme, even when the compressed codeword is transmitted between the terminal and the access network equipment and is subjected to adversarial perturbation attacks, the terminal can still obtain the de-perturbed compressed codeword through the first target model. This ensures the accuracy of the compressed codeword obtained by the terminal, thereby ensuring the accuracy of the uplink channel information reconstructed by the terminal based on the compressed codeword, and thus ensuring the quality of wireless communication.

[0100] In some possible implementations, the processing on the second device side includes: adjusting the dimension of the channel information between the first device and the second device to obtain adjusted channel information; encoding the adjusted channel information to obtain a second compressed codeword; and sending the second compressed codeword.

[0101] In this implementation, the processing before the second device performs dimensional adjustment on the channel information between the first device and the second device to obtain the adjusted channel information may further include: measuring a reference signal from the first device to obtain the channel information.

[0102] After the first device sends a reference signal, it can receive the first information, then input the first information into the first target model to obtain the first compressed codeword output by the first target model after the first information has been denoted; and then reconstruct the channel information between the first device and the second device based on the first compressed codeword.

[0103] In this way, the second device adjusts the dimensions of the channel information and generates a second compressed codeword based on the adjusted channel information. This increases the randomness of the second compressed codeword and resists disturbances to the compressed codeword emitted by the second device. Furthermore, after receiving the first information, the first device obtains the denoised compressed codeword based on the first target model. This ensures the accuracy of the compressed codeword obtained by the first device as the receiver, thereby guaranteeing the accuracy of the channel information reconstructed by the first device based on this compressed codeword, and ultimately ensuring the quality of wireless communication.

[0104] In some embodiments, in the downlink scenario, the first device is an access network device and the second device is a terminal. The reference signal can be a downlink reference signal, and the channel information can be downlink channel information.

[0105] The terminal-side processing includes: adjusting the dimensions of the downlink channel information between the terminal and the access network equipment to obtain adjusted downlink channel information; encoding the adjusted downlink channel information to obtain a second compressed codeword; and transmitting the second compressed codeword. The relevant explanation regarding the terminal obtaining the downlink channel information between the terminal and the access network equipment by measuring the downlink reference signal is the same as in the aforementioned embodiments and will not be repeated.

[0106] Adjusting the dimensions of the downlink channel information to obtain the adjusted downlink channel information may include: preprocessing the downlink channel information to obtain preprocessed downlink channel information; and expanding the dimensions of the preprocessed downlink channel information to obtain adjusted downlink channel information under the first target dimension, wherein the dimension of the preprocessed downlink channel information is the first initial dimension, and the first target dimension is greater than the first initial dimension.

[0107] The dimension expansion can also use pre-configured expansion methods, which are not limited here. The first target dimension can be pre-configured, and this first target dimension can be configured according to the actual situation. As long as the first target dimension is greater than the first initial dimension, it is within the protection scope of this embodiment.

[0108] The function of this data preprocessing is to ensure that the processed downlink channel information is within a suitable range. The processing methods included in the data preprocessing can be configured according to the actual situation. For example, data preprocessing can include at least one of the following: normalization, standardization, data transformation operations, etc.

[0109] It should be understood that the processing of the downlink channel information by the terminal to obtain the adjusted downlink channel information can be implemented in the terminal's encoder. For example, a dimension adjustment (i.e., dimension expansion) process (i.e., a resize operation) can be added to the input layer of the encoder. That is, the encoder can be structurally adjusted, and an input layer for expanding the dimensions can be added to the encoder. This embodiment does not limit the training method of the encoder.

[0110] For example, the way to expand the dimension of the downlink channel information under the first initial dimension can also be zero-padding. For example, the downlink channel information is a channel matrix H, whose first initial dimension is 32×32. A resize operation is performed at the input layer of the encoder, and the dimension of the input channel matrix H is changed to the first target dimension of 35×35 (i.e., the adjusted downlink channel information) by adding 0 elements.

[0111] After receiving the second compressed codeword, the terminal can send it. It should be noted that the dimensions of the second compressed codeword can also be increased on the terminal side. For example, if the downlink channel information is expanded from 32×32 to 36×36, and the compression rate of the compressed codeword is 1 / 4, then the second compressed codeword will also be expanded from 8×8 to 9×9.

[0112] Accordingly, the access network device can receive the first information. Further, after receiving the first information, the access network device inputs the first information into the first target model to obtain the first compressed codeword output by the first target model after de-perturbation of the first information. In this embodiment, the function of the first target model is to de-perturb the received signal or information to generate data or codewords with a distribution similar to or close to the compressed codeword (i.e., the second compressed codeword) sent by the terminal. The first target model is a trained or successfully trained neural network with generative capabilities. The composition structure of the first target model is the same as in the aforementioned embodiments and will not be repeated here.

[0113] The processing after the access network device obtains the first compressed codeword may further include: reconstructing the downlink channel information between the access network device and the terminal based on the first compressed codeword.

[0114] Specifically, the step of reconstructing the downlink channel information between the access network device and the terminal based on the first compressed codeword can be as follows: inputting the first compressed codeword into the decoder to obtain the initial downlink channel information between the access network device and the terminal reconstructed (or restored) by the decoder; adjusting the dimensions of the initial downlink channel information to reconstruct the downlink channel information between the access network device and the terminal.

[0115] The process of adjusting the dimensions of the initial downlink channel information to reconstruct the downlink channel information between the access network device and the terminal may include: performing dimensionality reduction processing on the initial downlink channel information under the second initial dimension to obtain the reconstructed downlink channel information between the access network device and the terminal under the second target dimension, wherein the second target dimension is smaller than the second initial dimension.

[0116] The second target dimension can be pre-configured, and can be configured according to the actual situation. The second target dimension can be equal to the first initial dimension.

[0117] It should be understood that the process of adjusting the dimensions of the initial downlink channel information and reconstructing the downlink channel information between the access network device and the terminal on the access network device side can also be implemented by the decoder on the access network device side. For example, a dimensionality reduction process (i.e., a resize operation) can be added to the output layer of the decoder. That is, the encoder and decoder can be structurally adjusted, with an output layer for dimensionality reduction added to the decoder and an input layer for dimensionality expansion added to the encoder. This embodiment does not limit the training method of the encoder and decoder.

[0118] The above embodiments can use randomized input processing to defend against adversarial attacks. Specifically, the terminal randomizes the downlink channel information input to the encoder, thereby causing the dimension of the compressed codewords output by the encoder to change randomly, thus increasing the difficulty of adversarial attacks. Referring to Figure 6, an exemplary illustration of the above embodiments is provided:

[0119] First, data preprocessing can be performed. For example, the downlink channel information input to the encoder can be preprocessed to obtain preprocessed downlink channel information, such as normalization, standardization, or other data transformation operations, to ensure that the data is within an appropriate range.

[0120] Then, a resize operation is performed on the input layer of the encoder to expand the dimensions of the preprocessed channel information (in this example, downlink channel information) to obtain the adjusted channel information in the first target dimension. For example, if the channel matrix H of the channel information has a dimension of 32×32, a resize operation is performed on the input layer of the neural network (i.e., the encoder) to change the input dimension to 35×35. Here, the dimension is generally set slightly larger than the previous dimension, and the specific size can be set randomly. When the compression ratio is 1 / 4, the dimension of the compressed codewords obtained by the encoder will also increase. Then, at the receiving end, the decoder recovers the channel information (i.e., the recovered channel information) by retrieving the compressed codewords. Since the input of the model (encoder) and the compressed codewords are vulnerable to adversarial attacks, randomizing their dimensions makes the compressed codewords transmitted by the transmitting end more resistant to disturbances.

[0121] Since the input dimension of the previous layer network is larger than the actual input dimension, it is necessary to modify the channel matrix of the downlink channel information of the actual input (e.g., represented as H). re Zero-padding can be performed, for example, it can be represented as:

[0122] Here, h11, h12, h21, and h22 constitute a 2×2 dimension channel matrix of downlink channel information (or downlink channel information in the first initial dimension). By adding "0", the 2×2 dimension channel matrix of downlink channel information (or downlink channel information in the first initial dimension) is filled into a 3×3 channel matrix of downlink channel information (or downlink channel information in the first target dimension). In this way, by adding two random layers to fine-tune the channel matrix, the computational cost is small, and the network model is more robust. It can adapt to different network structure models and adversarial defense methods.

[0123] By adopting the above scheme, the terminal adjusts the dimensionality of the downlink channel information and then generates a second compressed codeword based on the adjusted downlink channel information. This increases the randomness of the second compressed codeword and resists disturbances to the compressed codeword transmitted by the terminal. Furthermore, after receiving the first information, the access network device obtains the denoised compressed codeword using the first target model. Thus, even if the compressed codeword is subjected to adversarial disturbance attacks during transmission between the terminal and the access network device, the access network device can still obtain the denoised compressed codeword through the first target model. This ensures the accuracy of the compressed codeword obtained by the access network device, thereby guaranteeing the accuracy of the downlink channel information reconstructed by the access network device based on the compressed codeword, and ultimately ensuring the quality of wireless communication.

[0124] In some embodiments, in the uplink scenario, the first device is a terminal and the second device is an access network device. The reference signal can be an uplink reference signal, and the channel information can be uplink channel information.

[0125] The processing by the access network device includes: adjusting the dimensionality of the uplink channel information between the terminal and the access network device to obtain adjusted uplink channel information; encoding the adjusted uplink channel information to obtain a second compressed codeword; and transmitting the second compressed codeword. The relevant explanation regarding the access network device obtaining the uplink channel information between the terminal and the access network device by measuring the uplink reference signal is the same as in the aforementioned embodiments and will not be repeated.

[0126] Adjusting the dimensions of the uplink channel information to obtain the adjusted uplink channel information may include: preprocessing the uplink channel information to obtain preprocessed uplink channel information; and expanding the dimensions of the preprocessed uplink channel information to obtain the adjusted uplink channel information under the third target dimension, wherein the dimension of the preprocessed uplink channel information is the third initial dimension, and the third target dimension is greater than the third initial dimension.

[0127] The dimension expansion can also use pre-configured expansion methods, which are not limited here. The third target dimension can be pre-configured and can be configured according to the actual situation. As long as the third target dimension is greater than the third initial dimension, it is within the protection scope of this embodiment. The relevant descriptions of this data preprocessing are the same as those in the previous embodiments and will not be repeated.

[0128] It should be understood that the process of the access network device performing dimensional adjustment on the uplink channel information to obtain the adjusted uplink channel information can be implemented in the encoder of the access network device. For example, dimensional adjustment (i.e., dimensional expansion) processing (i.e., resize operation) can be added to the input layer of the encoder. This embodiment does not limit the training method of the encoder.

[0129] After obtaining the second compressed codeword, the access network device can send it. It should be noted that the second compressed codeword can also have its dimensions increased accordingly. For example, if the uplink channel information is expanded from 32×32 to 36×36, and the compression rate of the compressed codeword is 1 / 4, then the second compressed codeword will also be expanded from the original 8×8 to 9×9.

[0130] Accordingly, the terminal can receive the first information. Further, after receiving the first information, the terminal inputs the first information into the first target model to obtain the first compressed codeword output by the first target model after de-perturbation of the first information. In this embodiment, the function of the first target model is to de-perturb the received signal or information to generate data or codewords with a distribution similar to or close to the compressed codeword (i.e., the second compressed codeword) sent by the terminal. The first target model is a trained or successfully trained neural network with generative capabilities. The composition structure of the first target model is the same as in the aforementioned embodiments and will not be repeated here.

[0131] The processing after the terminal obtains the first compressed codeword may further include: reconstructing the uplink channel information between the access network device and the terminal based on the first compressed codeword.

[0132] Specifically, the step of reconstructing the uplink channel information between the access network device and the terminal based on the first compressed codeword can be as follows: inputting the first compressed codeword into the decoder to obtain the initial uplink channel information between the access network device and the terminal reconstructed (or restored) by the decoder; adjusting the dimensions of the initial uplink channel information to reconstruct the uplink channel information between the access network device and the terminal.

[0133] The process of adjusting the dimensions of the initial uplink channel information to reconstruct the uplink channel information between the access network device and the terminal may include: performing dimensionality reduction processing on the initial uplink channel information under the fourth initial dimension to obtain the reconstructed uplink channel information between the access network device and the terminal under the fourth target dimension, wherein the fourth target dimension is smaller than the fourth initial dimension. This fourth target dimension can be configured according to actual conditions, and it can be equal to the third initial dimension.

[0134] It should be understood that the process of adjusting the dimensions of the initial uplink channel information on the terminal side and reconstructing the uplink channel information between the access network device and the terminal can be implemented by the decoder on the terminal side. That is, the encoder and decoder can be structurally adjusted. An output layer for dimensionality reduction can be added to the decoder, and an input layer for dimensionality expansion can be added to the encoder. This embodiment does not limit the training method of the encoder and decoder.

[0135] The above embodiments use randomized input processing to defend against adversarial attacks. Specifically, the access network device randomizes the uplink channel information of the input encoder, causing the dimension of the compressed codewords output by the encoder to change randomly, thereby increasing the difficulty of adversarial attacks.

[0136] By adopting the above scheme, the access network device adjusts the dimensionality of the uplink channel information and then generates a second compressed codeword based on the adjusted uplink channel information. This increases the randomness of the second compressed codeword and resists disturbances to the compressed codeword transmitted by the access network device. Furthermore, after receiving the first information, the terminal obtains the denoised compressed codeword through the first target model. Thus, even under adversarial disturbance attacks during the transmission of compressed codewords between the terminal and the access network device, the terminal, as the receiving end, can still obtain the denoised compressed codeword through the first target model. This ensures the accuracy of the compressed codeword obtained by the terminal, thereby guaranteeing the accuracy of the uplink channel information reconstructed by the terminal based on the compressed codeword, and ultimately ensuring the quality of wireless communication.

[0137] It should be noted that the above embodiments describe the processing of access network devices and terminals in uplink and downlink scenarios respectively.

[0138] In practical applications, it's possible to use either one of the above uplink scenario solutions alone, or either one of the above downlink scenario solutions alone. Another possibility is to use both uplink and downlink scenario solutions. Specifically, the same pair of access network devices and terminals can have both uplink and downlink scenarios. That is, the same access network device will send downlink reference signals and also measure uplink reference signals, and the same terminal will measure downlink reference signals and also send uplink reference signals. Therefore, uplink and downlink scenarios can be combined. Both the access network device and the terminal can have a first target model. On the access network device side, the first target model is used to de-scramble the information received from the air interface and obtain de-scrambled compressed codewords, so that the decoder on the access network device side can reconstruct the downlink channel information based on the de-scrambled compressed codewords. Similarly, the terminal can also have a first target model. On the terminal side, the first target model is used to de-scramble the information received from the air interface and obtain de-scrambled compressed codewords, so that the decoder on the terminal side can reconstruct the uplink channel information based on the de-scrambled compressed codewords.

[0139] In some possible implementations, the first target model can be trained on the first device side. The timing of the first device obtaining the first target model before receiving the first information is within the scope of protection of this embodiment. In this embodiment, the first device can be an access network device or a terminal.

[0140] The method by which the first device trains to obtain the first target model may include: adding perturbation to the compressed codeword sample to obtain perturbed codeword sample; jointly training the first generation model and the first discriminant model based on the compressed codeword sample and the perturbed codeword sample to obtain the trained first generation model and the trained first discriminant model; and using the trained first generation model as the first target model.

[0141] Here, the first generative model can refer to the first untrained original generative model, the first initial generative model, or the first preset generative model, while the trained first generative model can refer to the final target generative model obtained after training (or joint training). In the following text, if only "first generative model" is mentioned without specifically referring to "trained" first generative model, it refers to the untrained first original generative model, the first initial generative model, or the first preset generative model; if "trained first generative model" is mentioned, it refers to the final target generative model obtained after training, or, for distinction, the trained first generative model can be alternatively described as the third generative model. In the following text, the first original generative model, the first initial generative model, the first preset generative model, and the first generative model (i.e., the non-trained first generative model) have the same meaning, as do the first target generative model, the third generative model, and the trained first generative model. Additionally, in some possible examples, the generative model can also be called a generator, a generative network, etc.

[0142] The first discriminant model can refer to the untrained first original discriminant model, the first initial discriminant model, or the first preset discriminant model. The trained first discriminant model can refer to the final target discriminant model obtained after training (or joint training). In the following text, if only "first discriminant model" is mentioned without specifically referring to "trained" first discriminant model, it refers to the untrained first original discriminant model, the first initial discriminant model, or the first preset discriminant model. If "trained first discriminant model" is mentioned, it refers to the final target discriminant model obtained after training, or, for distinction, the trained first discriminant model can be alternatively described as the third discriminant model. That is, in the following text, the first original discriminant model, the first initial discriminant model, and the first discriminant model (i.e., not specifically referring to the trained first discriminant model) have the same meaning, as do the first target discriminant model, the third discriminant model, and the trained first model. Additionally, in some possible examples, the discriminant model can also be alternatively called a discriminator, a classifier, or a classification model, etc., without limitation or exhaustive listing.

[0143] That is, the first device trains a first initial generation model, which is jointly trained with a first initial discrimination model. After obtaining the trained first generation model (the third generation model or the first target generation model), the first device uses the trained first generation model (the third generation model or the first target generation model) as the first target model.

[0144] The compressed codeword samples can be pre-acquired, and the method of acquisition or generation is not limited in this embodiment. The number of compressed codeword samples can be one or more, and is not limited in this embodiment.

[0145] Adding perturbation to compressed codeword samples to obtain perturbed codeword samples can include: generating a first perturbation signal that satisfies a first preset condition, and adding the first perturbation signal to the compressed codeword samples to obtain perturbed codeword samples. This first perturbation signal can also be called an adversarial perturbation, or an adversarial perturbation added to the channel, etc.

[0146] The first preset condition may include at least one of the following: the power of the first disturbance signal is less than a first specified power value; the content obtained after decoding the disturbance codeword sample with the added first disturbance signal by the decoder is maximized to be different from the channel information sample corresponding to the compressed codeword sample. The channel information sample (or original channel information sample) corresponding to the compressed codeword sample may be pre-acquired, and its acquisition or generation method is not limited in this embodiment. Here, when the first device is an access network device, the channel information sample may be a downlink channel information sample; when the first device is a terminal, the channel information sample may be an uplink channel information sample.

[0147] For example, this preset condition can be expressed as:

[0148] Among them, f de (·) represents the decoder, x represents the compressed codeword sample, z represents the first perturbation signal, "x+z" represents the perturbation codeword sample after adding the first perturbation signal, and H represents the channel information sample. Represents the square of the 2-norm. This means maximizing the difference between the content obtained after decoding the perturbation codeword sample with the first perturbation signal added by the decoder and the channel information sample corresponding to the compressed codeword sample; "st" indicates that the power of the adversarial perturbation is constrained by δ (or the first specified power value), ||·||2 indicates the calculation of the 2-norm, and ‖z‖2≤δ indicates that the 2-norm of the power of the perturbation signal is less than the specified power value δ.

[0149] The step of jointly training a first generation model and a first discriminant model based on the compressed codeword samples and the perturbed codeword samples to obtain the trained first generation model and the trained first discriminant model includes: inputting the perturbed codeword samples into the first generation model to obtain the deperturbed codeword samples output by the first generation model; obtaining a first similarity parameter between the deperturbed codeword samples and the compressed codeword samples based on the first discriminant model; calculating a first target loss based on the compressed codeword samples, the deperturbed codeword samples, and the first similarity parameter; and jointly training the first generation model and the first discriminant model based on the first target loss to obtain the trained first generation model and the trained first discriminant model.

[0150] Inputting the perturbed codeword sample into the first generation model to obtain the deperturbed codeword sample output by the first generation model means: inputting the perturbed codeword sample into the first initial generation model to obtain the deperturbed codeword sample output by the first initial generation model. The deperturbed codeword sample can also be called an adversarial sample or a pseudo codeword sample, etc., and we will not limit or exhaustively list all possible names for it here.

[0151] Specifically, obtaining the first similarity parameter between the de-perturbed codeword sample and the compressed codeword sample based on the first discrimination model can be achieved by: inputting the de-perturbed codeword sample and the compressed codeword sample into the first discrimination model to obtain the first similarity parameter between the de-perturbed codeword sample and the compressed codeword sample output by the first discrimination model. More specifically, this can be achieved by: inputting the de-perturbed codeword sample and the compressed codeword sample into a first initial discrimination model to obtain the first similarity parameter between the de-perturbed codeword sample and the compressed codeword sample output by the first initial discrimination model.

[0152] For example, the first similarity parameter can be an indication value used to indicate whether the dedisturbed codeword sample and the compressed codeword sample are similar. The first similarity parameter can be equal to a first indication value or a second indication value, where the first indication value is different from the second indication value. The first indication value can be used to indicate that the dedisturbed codeword sample and the compressed codeword sample are different, and the second indication value can be used to indicate that the dedisturbed codeword sample and the compressed codeword sample are the same. For example, the first indication value can be 0 and the second indication value can be 1, or the first indication value can be -1 and the second indication value can be 1.

[0153] For example, the first similarity parameter can be the similarity probability value between the de-perturbed codeword sample and the compressed codeword sample. The similarity probability value can be any value greater than or equal to a first value and less than or equal to a second value, where the first value is less than the second value. The first value can be used to indicate that the de-perturbed codeword sample and the compressed codeword sample are completely different, and the second value can be used to indicate that the de-perturbed codeword sample and the compressed codeword sample are completely identical. For example, the first value can be 0 and the second value can be 1, or the first value can be -1 and the second value can be 1. For instance, if the first value is 0, the second value is 1, and the similarity probability value is 0.8, then it can be said that the probability of similarity between the de-perturbed codeword sample and the compressed codeword sample is close to 1, or that the de-perturbed codeword sample and the compressed codeword sample are substantially similar.

[0154] The step of calculating the first target loss based on the compressed codeword samples, the de-perturbated codeword samples, and the first similarity parameter includes: calculating the first target generation loss based on the compressed codeword samples, the de-perturbated codeword samples, and the first similarity parameter; calculating the first target discrimination loss based on the compressed codeword samples and the first similarity parameter; and obtaining the first target loss based on the first target generation loss and the first target discrimination loss.

[0155] Specifically, calculating the first target generation loss based on the compressed codeword sample, the de-perturbated codeword sample, and the first similarity parameter includes: calculating a first generation sub-loss based on the compressed codeword sample and the de-perturbated codeword sample; calculating a second generation sub-loss based on the first similarity parameter; and calculating the first target generation loss based on the first generation sub-loss and the second generation sub-loss.

[0156] Based on the compressed codeword samples and the denoised codeword samples, the loss function or calculation method used to calculate the first generator loss can be configured according to the actual situation. For example, the MSE (Mean Squared Error) loss function can be used to calculate the first generator loss. mse It can be represented as: Where N represents the number of compressed codeword samples (e.g., one or more), n is an integer ranging from 1 to N, and x clean Represents a compressed codeword sample, x adv Indicates the perturbation codeword sample, x adv =x clean +z represents the perturbation codeword sample after adding the perturbation signal z to the compressed codeword sample, and G1 represents the first generation model (first initial generation model). G1(x adv ) represents the deperturbed codeword sample output by the first generative model (first initial generative model).

[0157] Based on the first similarity parameter, the loss function used to calculate the second generator loss can be any binary classification loss function; this embodiment does not impose any limitation. For example, the second generator loss is calculated as l adv It can be represented as Wherein, D1() is used to simply represent the first discriminant model, log represents the logarithmic calculation (the base of the logarithmic calculation is not limited in this embodiment), and the meanings of the other parameters are the same as in the previous example, and will not be repeated.

[0158] The first generation target loss is calculated based on the first generation sub-loss and the second generation sub-loss. This can be achieved by calculating the first generation target loss based on the first generation sub-loss, the first weight corresponding to the first generation sub-loss, the second generation sub-loss, and the second weight corresponding to the second generation loss. The first and second weights can be configured according to actual conditions, and this embodiment does not impose any limitations. For example, the first weight can be 100, and the second weight can be 0.1.

[0159] For example, calculate the loss l of the first generated target. g It can be represented as: l g =α*l mse +β*l adv Where α represents the first weight, β represents the second weight, and the meanings of the other parameters are the same as in the previous example, and will not be repeated here.

[0160] The calculation of the first discrimination target loss based on the compressed codeword sample and the first similarity parameter can be as follows: calculate the first discriminant sub-loss based on the compressed codeword sample; calculate the second discriminant sub-loss based on the first similarity parameter; and calculate the first discrimination target loss based on the first discriminant sub-loss and the second discriminant sub-loss.

[0161] In this embodiment, the first discriminant loss can be calculated using any binary classification loss function based on the compressed codeword sample. For example, the compressed codeword sample can be input into the first discriminator to obtain the similarity between the compressed codeword sample and itself. Then, the loss value corresponding to the similarity can be calculated based on the binary classification loss function as the first discriminant loss. In other words, the first discriminant loss can be used to represent the loss of classifying the compressed sample codeword as a compressed sample codeword by the first discriminant model.

[0162] The loss of the second discriminant can be calculated using any binary classification loss function based on the first similarity parameter, and this embodiment does not limit it.

[0163] The first discrimination target loss is calculated based on the first discriminant sub-loss and the second discriminant sub-loss. This can be achieved by adding the first discriminant sub-loss and the second discriminant sub-loss to obtain the first discrimination target loss. For example, the first discrimination target loss is l. d It can be represented as: The meanings of each parameter are the same as in the previous example, and will not be repeated here.

[0164] Based on the first target loss, the first generative model and the first discriminative model are jointly trained to obtain the trained first generative model and the trained first discriminative model. This can be achieved by: training the first generative model using the first generative target loss, and obtaining the trained first generative model if the training convergence condition of the first generative model is met; and training the first discriminative model using the first discriminative target loss, and obtaining the trained first discriminative model if the training convergence condition of the first discriminative model is met. In other words, the first initial generative model is trained using the first generative target loss, and the trained first generative model (i.e., the third generative model or the first target generative model) is obtained if the training convergence condition of the first generative model is met; the first initial discriminative model is trained using the first discriminative target loss, and the trained first discriminative model (i.e., the third discriminative model or the first target discriminative model) is obtained if the training convergence condition of the first discriminative model is met.

[0165] Training the first generative model (i.e., the first initial generative model) using the first generative target loss can refer to updating the parameters of the first generative model (i.e., the first initial generative model) based on the first generative target loss. Training the first discriminative model (i.e., the first initial discriminative model) using the first discriminative target loss can refer to updating the parameters of the first discriminative model (i.e., the first initial discriminative model) based on the first discriminative target loss.

[0166] The training convergence condition (or training termination condition) of the first generative model can be that the distribution of the denoised codeword samples generated by the first generative model is approximately the same as that of the compressed codeword samples. For example, the training convergence condition of the first generative model can be: the first generative model has been trained a specified number of times, and / or the difference between the first generation target loss and the first target value is less than a first specified difference. The specified number of training times can be configured according to the actual situation; the first target value can be configured according to the actual situation, for example, it can be 0; the first specified difference can be configured according to the actual situation, for example, it can be close to 0 (e.g., 0.1, or 0.01, etc.).

[0167] The convergence condition for training the first discriminant model can be that the first discriminant model cannot distinguish between perturbated codeword samples and compressed codeword samples. For example, the convergence condition for training the first discriminant model can be: the first discriminant model has been trained a specified number of times, and / or the difference between the first discriminant target loss and the second target value is less than a second specified difference. The specified number of training times for the first discriminant model can be the same as the specified number of training times for the first generative model, and the specified number of training times can be configured according to the actual situation; the second target value can be configured according to the actual situation; the second specified difference can be configured according to the actual situation, for example, it can be close to 0 (e.g., 0.01, or 0.05, etc.).

[0168] It should be noted that the joint training of the first generative model and the first discriminative model (i.e., the first initial generative model and the first initial discriminative model) can be performed in multiple loops. In each training loop, the first target loss can be calculated in the above manner. The processing of each training loop may input different compressed codeword samples or the same compressed codeword samples. This embodiment does not limit whether the compressed codeword samples used in each training loop are the same. For example, multiple compressed codeword samples in a batch can be generated in advance, and any compressed codeword sample can be input in each training loop to obtain the corresponding first generative target loss and first discriminative target loss as the first target loss.

[0169] It should also be noted that during the repeated training iterations, the first generative model can be trained using the first target loss and the first discriminative model using the first discriminative loss in each training iteration. Alternatively, the first generative model can be trained using the first target loss first, and after fixing the parameters of the first generative model, the first discriminative model can be trained using the first discriminative loss. This process can be repeated until the trained first generative model (i.e., the third generative model or the first target generative model) and the trained first discriminative model (i.e., the third discriminative model or the first target discriminative model) are obtained. Alternatively, the first discriminant model can be trained using the first discriminant target loss, with the parameters of the first discriminant model fixed. Then, the first generative model can be trained using the first generative target loss, with the parameters of the first generative model fixed again. This process is repeated until the trained first generative model (i.e., the third generative model or the first target generative model) and the trained first discriminant model (i.e., the third discriminant model or the first target discriminant model) are obtained. This embodiment does not limit the training order of the first generative model and the first discriminant model during multiple iterations of training.

[0170] In one example, the training process of the first generative model and the first discriminative model can be as follows: First, a batch of clean samples (i.e., a batch of compressed codeword samples) can be randomly selected from the sample dataset. Perturbed codeword samples corresponding to the compressed codeword samples are generated. The perturbed codeword samples are then deperturbed by the first generative model (first initial generative model) to obtain deperturbed codeword samples (which can be called adversarial samples). Then, based on the first discriminative model (first initial discriminative model), the first similarity parameter between the deperturbed codeword samples and the compressed codeword samples is obtained. Then, based on the above processing, the first discriminative target loss is obtained. The parameters of the first discriminative model (first initial discriminative model) are updated based on the first discriminative target loss. By minimizing the first discriminative target loss, the first discriminative model can better distinguish between clean samples and adversarial samples. Then, the parameters of the first generative model (first initial generative model) are updated. By maximizing the first discriminative target loss, the deperturbed codeword samples can be made closer to clean samples. The parameters of the first discriminator and the first generator can be updated multiple times to achieve better training results.

[0171] Additionally, it should be noted that in some possible scenarios, both the access network device and the terminal can use the first target model. In such cases, if the first device is an access network device, after training the first target model, the access network device can also send the first target model to the terminal, and the terminal can receive the first target model from the access network device. If the first device is a terminal, after training the first target model, the terminal can also send the first target model to the access network device, and the access network device can receive the first target model from the terminal.

[0172] In some embodiments, the first target model is pre-configured on the first device side. Before receiving the first information, the first device may further include: receiving the first target model from another electronic device.

[0173] In this embodiment, the first target model is the trained first generative model (i.e., the third generative model or the first target generative model). The training of the first generative model can be performed by other electronic devices. The method of training the first generative model on the other electronic device side (that is, jointly training the first generative model and the first discriminant model to obtain the trained first generative model (i.e., the third generative model or the first target generative model) and the trained first discriminant model (i.e., the third discriminant model or the first target discriminant model)) is the same as in the previous embodiment and will not be described again.

[0174] The other electronic device can be any electronic device with computing capabilities other than the first device. For example, the other electronic device can be a server (such as any server or service node in a service cluster), or a core network device (such as any core network device on the core network side), or a user-end electronic device (such as a personal computer), etc. Here, we do not limit or exhaust all possible types of other electronic devices.

[0175] Additionally, it should be noted that in some possible scenarios, both the access network device and the terminal can use the first target model. In this case, if the first device is an access network device, after receiving the first target model from other electronic devices, the access network device can also send the first target model to the terminal, and the terminal can receive the first target model from the access network device; or, other electronic devices can also send the trained first target model to the terminal, and the terminal can receive the first target model from the other electronic devices. If the first device is a terminal, after receiving the first target model from other electronic devices, the terminal can also send the first target model to the access network device, and the access network device can receive the first target model from the terminal; or, other electronic devices can also send the trained first target model to the access network device, and the access network device can receive the first target model from the other electronic devices.

[0176] In one embodiment, the encoder on the second device side and the decoder on the first device side can be trained separately from the first target model described above.

[0177] The encoder and decoder can form a channel feedback neural network (CSINet). The process of training this channel feedback neural network can be as follows: clean samples and corresponding adversarial samples are put into the training dataset, and the loss function for training the channel feedback neural network is defined as:

[0178] Here, NMSE quantizes the difference in the channel matrix, and H represents the channel information sample (or input sample) input to the encoder. The decoder output (or output sample) is used to train the channel feedback neural network using the loss function described above, so that the difference between the input sample and the output sample of the trained channel feedback neural network is minimized.

[0179] Referring to Figure 7, the training of the first generative model and the first discriminative model target model is illustrated by example:

[0180] S701: Preset conditions. Generate the first disturbance signal. Multiple counter-disturbance generation methods can be set. The generated counter-disturbance must meet the following first preset conditions:

[0181] The descriptions of the various parameters are the same as in the previous embodiments and will not be repeated. The goal of this step in generating the perturbation is to generate a first perturbation signal and add it to the compressed codeword sample x to obtain the perturbation codeword sample, thus making it impossible to reconstruct a perfect CSI (Channel Information) based on the perturbation codeword sample.

[0182] S702: Training the first generator G.

[0183] Specifically, after generating the disturbance in step S701, through x adv =x+z generates adversarial perturbation codeword samples x adv The first generator G compares x... adv and the corresponding compressed codeword sample x clean To reconstruct the output compressed codeword x generate The training objective of the first generator is to find the optimal weight parameters: Among them, l g The composition of the first generated target loss in the aforementioned embodiments will not be repeated here; This means that the first generator is trained using the first generation target loss, so that the deperturbed codeword samples reconstructed by the first generator based on the perturbed codeword samples are closest to the compressed codeword samples.

[0184] S703: Training the first discriminator D.

[0185] Here, the input to the first discriminator D is the reconstructed, denoised codeword sample x. generate =G(x) adv ) and its corresponding compressed codeword samples x The first discriminator is trained to distinguish the reconstructed, dedisturbed codeword samples x. generate and compressed codeword samples x Therefore, a reconstructed, denoised codeword sample that is highly similar to the original compressed codeword sample can be obtained by training the first discriminator D. The calculation method of the first discriminant target loss used in training the first discriminator D is the same as in the previous embodiment, and will not be repeated here.

[0186] The first generator continuously optimizes the data it generates to make it indistinguishable from the first discriminator, while the first discriminator also optimizes itself to make more accurate judgments. Specific training details are as follows: A batch of real samples (compressed codeword samples) is randomly selected from the dataset. The first generator generates corresponding adversarial examples, and the parameters of the first discriminator are updated. By minimizing the discriminator's loss function, the first discriminator can better distinguish between real samples (compressed codeword samples) and adversarial examples (de-perturbated codeword samples). The parameters of the first generator model are then updated, and by maximizing the discriminator's loss function, the generated adversarial examples (de-perturbated codeword samples) are made closer to the real samples (compressed codeword samples). Multiple updates of the discriminator and generator parameters can be performed to achieve better training results.

[0187] After the GAN network (first generator and first discriminator) is trained, the trained first discriminator D (i.e., the first target discriminator) will no longer be needed during testing or inference. Only the first generator G (i.e., the trained first generator or the first target generator) is needed. Finally, the trained first generator or the first target generator will be used as the first target model.

[0188] Furthermore, previous research combining deep learning with CSI feedback did not consider security issues. The solution provided in this embodiment addresses the vulnerability of CSI feedback to both white-box and black-box adversarial attacks due to the openness of wireless channels and the uninterpretable nature of deep learning. Therefore, this embodiment employs an external generator network (i.e., the first target model) attached to the first device (receiver) containing the decoder to simultaneously defend against both white-box and black-box adversarial attacks. Additionally, this solution does not require changes to the deep learning model related to CSI feedback or modifications to the CSI feedback network itself, thus saving training resources.

[0189] Figure 8 is a schematic flowchart of a communication method performed by a second device according to an embodiment of this application. The method includes at least a portion of the following.

[0190] S810, receives air interface signals;

[0191] S820. Input the air interface signal into the second target model to obtain the denoised channel information between the second device and the first device output by the second target model.

[0192] Furthermore, after the second device obtains the denoised channel information between the second device and the first device output by the second target model, the method further includes: encoding the channel information to obtain a second compressed codeword; and sending the second compressed codeword.

[0193] The first device is an access network device, the second device is a terminal, and the channel information is downlink channel information; or, the first device is a terminal, the second device is an access network device, and the channel information is uplink channel information.

[0194] In some possible implementations, in the downlink scenario, the first device is an access network device and the second device is a terminal. The reference signal can be a downlink reference signal, and the channel information can be downlink channel information.

[0195] The processing by the access network equipment may include transmitting a downlink reference signal. The downlink reference signal can be any one of CSI-RS, DMRS, etc., and the specific types of downlink reference signals are not limited or exhaustively listed here.

[0196] Specifically, the access network device sends the downlink reference signal, which can be done by transmitting the downlink reference signal over the air interface.

[0197] In this embodiment, the terminal receiving the air interface signal can be: the terminal receiving the downlink signal through the air interface. That is, in the downlink scenario, the air interface signal can specifically be a downlink signal, which may also be alternatively referred to as a downlink air interface signal, or an air interface downlink signal, or a downlink signal transmitted through the air interface, or a downlink signal transmitted through a wireless channel, etc.

[0198] In this embodiment, the downlink reference signal is transmitted by the access network device through the air interface, while the downlink signal is received by the terminal through the air interface. This is because the downlink reference signal may be subject to adversarial attacks during the transmission of the wireless channel. That is, the signal received by the terminal may be a signal obtained by adding perturbation or interference to the downlink reference signal. Therefore, the signal obtained by adding perturbation or interference to the downlink reference signal received by the terminal through the air interface is called the downlink signal.

[0199] The function of the second target model can be: to dedisturb the downlink signal to obtain a dedisturbed downlink reference signal, and to perform channel estimation based on the dedisturbed downlink reference signal to obtain the dedisturbed downlink channel information between the terminal and the access network device. The second target model is a neural network with generative capabilities. The composition of this second target model can include at least one of the following: fully connected neural network, CNN, DNN, RNN, GAAE, etc. The possible composition of this second target model is not limited or exhaustively listed here.

[0200] This downlink channel information may include CSI, etc.

[0201] The processing after the terminal obtains the de-disturbed downlink channel information between the terminal and the access network device further includes: encoding the downlink channel information to obtain a second compressed codeword; and sending the second compressed codeword.

[0202] The encoding of the downlink channel information to obtain the second compressed codeword can be achieved by the terminal inputting the downlink channel information into the encoder to obtain the second compressed codeword output by the encoder. The encoder can be pre-configured in the terminal, and this embodiment does not limit the method of generating or training the encoder.

[0203] Sending the second compressed codeword by the terminal may include: the terminal transmitting the second compressed codeword to the access network device via the air interface. The terminal may use any type of uplink AS message or information to carry or transmit the second compressed codeword; this embodiment does not impose any limitation.

[0204] In one embodiment, the access network device may receive a second compressed codeword; input the second compressed codeword into a decoder to obtain reconstructed downlink channel information.

[0205] In this embodiment, the access network device does not need to perform de-disturbance processing on the information received over the air interface, and directly obtains the second compressed codeword sent by the terminal. The decoder on the access network device side and the encoder on the terminal side can be pre-trained and matched with each other. This embodiment does not limit the composition, training method, etc. of the encoder and decoder.

[0206] In another embodiment, the access network device can receive the first information. In this embodiment, the terminal transmits the second compressed codeword through the air interface, while the access network device receives the first information through the air interface. This is because the second compressed codeword may be subject to adversarial attacks during transmission between the terminal and the access network device. That is, the information received by the access network device may be information that adds perturbation or interference to the second compressed codeword. Therefore, the information obtained by adding perturbation or interference to the second compressed codeword received by the access network device through the air interface is called the first information.

[0207] Furthermore, after the access network device receives the first information, it inputs the first information into the first target model to obtain the first compressed codeword output by the first target model after de-perturbation of the first information. In this embodiment, the relevant descriptions of the first target model on the access network device side and its training or acquisition methods are the same as in the previous embodiments, and will not be repeated.

[0208] The processing after the access network device obtains the first compressed codeword may further include: inputting the first compressed codeword into the decoder to obtain the downlink channel information between the access network device and the terminal reconstructed by the decoder. In this embodiment, the descriptions of the decoder on the access network device side and the encoder on the terminal side are the same as in the previous embodiments and will not be repeated.

[0209] Referring to Figure 9, another exemplary description of the communication method provided in this embodiment is given, including:

[0210] S901: The UE receives the downlink signal and inputs it into the second target model to obtain the denoised downlink channel information between the terminal and the access network device. Here, before step 801, it may also include: the base station (access network device) sending a downlink reference signal. This downlink reference signal may include CSI-RS, etc. Since attackers may add adversarial perturbations to the air interface signal, the downlink reference signal received by the UE is described as a downlink signal (i.e., a signal that adds perturbations to the downlink reference signal).

[0211] S902: The UE compresses the downlink channel information through the encoder to obtain the second compressed codeword.

[0212] S903: The UE sends the second compressed codeword to the base station via the air interface.

[0213] S904: The base station decodes the second compressed codeword using a decoder and infers the reconstructed downlink channel information.

[0214] In some possible implementations, in the uplink scenario, the first device is a terminal and the second device is an access network device. The reference signal can be an uplink reference signal, and the channel information can be uplink channel information.

[0215] The terminal's processing may include sending an uplink reference signal. Specifically, the terminal transmits the uplink reference signal over the air interface.

[0216] When an access network device receives an air interface signal, it can mean that the access network device receives an uplink signal through the air interface. That is, in an uplink scenario, the air interface signal can specifically be an uplink signal, which may also be referred to as an uplink air interface signal, an air interface uplink signal, an uplink signal transmitted through the air interface, or an uplink signal transmitted through a wireless channel, etc.

[0217] In this embodiment, the terminal transmits an uplink reference signal through the air interface, while the access network device receives an uplink signal through the air interface. This is because the uplink reference signal may be subject to adversarial attacks during wireless channel transmission. That is, the signal received by the access network device may be a signal obtained by adding perturbation or interference to the uplink reference signal. Therefore, the signal obtained by adding perturbation or interference to the uplink reference signal received by the access network device through the air interface is called the uplink signal.

[0218] The function of the second target model is to: denoise the uplink signal to obtain a denoised uplink reference signal, perform channel estimation based on the denoised uplink reference signal, and obtain the denoised uplink channel information between the terminal and the access network device. The relevant descriptions of the second target model are similar to those of the aforementioned implementation methods and will not be repeated here.

[0219] The processing of the uplink channel information by the access network device further includes: encoding the uplink channel information to obtain a second compressed codeword; and sending the second compressed codeword.

[0220] The encoding of the uplink channel information to obtain the second compressed codeword can be achieved by the access network device inputting the uplink channel information into the encoder to obtain the second compressed codeword output by the encoder. The encoder can be pre-configured, and this embodiment does not limit the method of generating or training the encoder.

[0221] Sending the second compressed codeword by the access network device may include: the access network device transmitting the second compressed codeword to the terminal via the air interface. The message carrying or transmitting the second compressed codeword is the same as in the aforementioned related embodiments and will not be described again.

[0222] In one embodiment, the terminal may receive a second compressed codeword; input the second compressed codeword into a decoder to obtain the reconstructed uplink channel information.

[0223] In another embodiment, the terminal may receive first information. This is because the second compressed codeword may be subject to adversarial attacks during transmission between the terminal and the access network device, meaning that the terminal may receive information that adds perturbation or interference to the second compressed codeword. Therefore, the information obtained by the terminal through the air interface that adds perturbation or interference to the second compressed codeword is called the first information.

[0224] Furthermore, after receiving the first information, the terminal inputs the first information into the first target model to obtain the first compressed codeword output by the first target model after de-perturbation of the first information; the first compressed codeword is then input into the decoder to obtain the uplink channel information between the access network device and the terminal reconstructed by the decoder. In this embodiment, the relevant descriptions of the decoder and encoder are the same as those in the aforementioned uplink scenario embodiments, and will not be repeated.

[0225] In some possible implementations, the second target model is locally trained on the second device side. In this implementation, the second device can be an access network device or a terminal.

[0226] Specifically, the method by which the second device trains to obtain the second target model may include: adding perturbation to the signal sample to obtain a perturbed signal sample; jointly training the second generation model and the second discriminant model based on the real channel information sample corresponding to the signal sample and the perturbed signal sample to obtain the trained second generation model and the trained second discriminant model; and using the trained second generation model as the second target model.

[0227] Here, the second generative model can refer to an untrained second original generative model, a second initial generative model, or a second preset generative model, while the trained second generative model can refer to the final second target generative model obtained after training (or joint training). In the following text, if only "second generative model" is mentioned without specifically referring to "trained" second generative model, it refers to the untrained second original generative model, second initial generative model, or second preset generative model; if "trained second generative model" is mentioned, it refers to the final second target generative model obtained after training, or, for distinction, the trained second generative model can be alternatively described as a fourth generative model; in the following text, the second original generative model, second initial generative model, second preset generative model, and second generative model (i.e., when not specifically referring to the trained second generative model) have the same meaning, as do the second target generative model, fourth generative model, and trained second generative model.

[0228] The second discriminant model can refer to an untrained second original discriminant model, a second initial discriminant model, or a second preset discriminant model, while the trained second discriminant model can refer to the final second target discriminant model obtained after training (or joint training). In the following text, if only "second discriminant model" is mentioned without specifically referring to "trained" second discriminant model, it refers to the untrained second original discriminant model, the second initial discriminant model, or the second preset discriminant model. If "trained second discriminant model" is mentioned, it refers to the final second target discriminant model obtained after training, or, for distinction, the trained second discriminant model can be alternatively described as the fourth discriminant model. That is, in the following text, the second original discriminant model, the second initial discriminant model, the second preset discriminant model, and the second discriminant model (i.e., not specifically referring to the trained second discriminant model) have the same meaning, as do the second target discriminant model, the fourth discriminant model, and the trained second discriminant model.

[0229] That is, the second device will train a second initial generation model, which is jointly trained with the second initial discrimination model. After obtaining the trained second generation model (the fourth generation model or the second target generation model), the second device will use the trained second generation model (the fourth generation model or the second target generation model) as the second target model.

[0230] The signal samples can be pre-acquired, and the method of acquisition or generation is not limited in this embodiment. The number of signal samples can be one or more, and is not limited in this embodiment.

[0231] Adding a perturbation to a signal sample to obtain a perturbation signal sample can include: generating a second perturbation signal that meets a second preset condition, and adding the second perturbation signal to the signal sample to obtain the perturbation signal sample. This second perturbation signal can also be called an adversarial perturbation, or an adversarial perturbation added to the channel, etc.

[0232] The second preset condition may include at least one of the following: the power of the second disturbance signal is less than a second specified power value; the codeword obtained by encoding the channel information obtained by the encoder after adding the second disturbance signal to the disturbance signal sample is maximized to be different from the compressed codeword sample corresponding to the signal sample; the channel information obtained by adding the second disturbance signal to the disturbance signal sample is maximized to be different from the real channel information sample corresponding to the signal sample.

[0233] In this embodiment, the compressed codeword sample corresponding to the signal sample can be pre-acquired, and the real channel information sample corresponding to the signal sample can also be pre-acquired. This embodiment does not limit the acquisition method of the compressed codeword sample corresponding to the signal sample or the real channel information sample corresponding to the signal sample. Furthermore, when the second device is a terminal, the signal sample can be a downlink signal sample, the compressed codeword sample corresponding to the signal sample can be the compressed codeword sample corresponding to the downlink signal sample, and the real channel information sample corresponding to the signal sample can be the real downlink channel information sample corresponding to the downlink signal sample. When the second device is an access network device, the signal sample can be an uplink signal sample, the compressed codeword sample corresponding to the signal sample can be the compressed codeword sample corresponding to the uplink signal sample, and the real channel information sample corresponding to the signal sample can be the real uplink channel information sample corresponding to the uplink signal sample.

[0234] The step of jointly training a second generation model and a second discriminant model based on the real channel information sample corresponding to the signal sample and the perturbation signal sample to obtain a trained second generation model and a trained second discriminant model includes: inputting the perturbation signal sample into the second generation model to obtain the channel information sample output by the second generation model; obtaining a second similarity parameter between the channel information sample and the real channel information sample corresponding to the signal sample based on the second discriminant model; calculating a second target loss based on the real channel information sample corresponding to the signal sample, the channel information sample, and the second similarity parameter; and jointly training the second generation model and the second discriminant model based on the second target loss to obtain the trained second generation model and the trained second discriminant model.

[0235] The phrase "inputting the disturbance signal sample into the second generation model to obtain the channel information sample output by the second generation model" refers to: inputting the disturbance signal sample into the second initial generation model to obtain the channel information sample output by the second initial generation model.

[0236] Based on the second discrimination model, obtaining the second similarity parameter between the channel information sample and the corresponding real channel information sample of the signal sample can be achieved by: inputting the channel information sample and the corresponding real channel information sample of the signal sample into the first discrimination model to obtain the second similarity parameter between the channel information sample output by the first discrimination model and the corresponding real channel information sample of the signal sample. Specifically, this can be achieved by inputting the channel information sample and the corresponding real channel information sample of the signal sample into the second initial discrimination model to obtain the second similarity parameter between the channel information sample output by the second initial discrimination model and the corresponding real channel information sample of the signal sample. The description of this second similarity parameter is similar to that of the first similarity parameter in the aforementioned embodiments and will not be repeated here.

[0237] The calculation of the second target loss based on the real channel information sample corresponding to the signal sample, the channel information sample, and the second similarity parameter includes: calculating the second target generation loss based on the real channel information sample corresponding to the signal sample, the channel information sample, and the second similarity parameter; calculating the second target discrimination loss based on the real channel information sample corresponding to the signal sample and the second similarity parameter; and obtaining the second target loss based on the second target generation loss and the second target discrimination loss.

[0238] Specifically, calculating the second target generation loss based on the real channel information sample corresponding to the signal sample, the channel information sample, and the second similarity parameter includes: calculating a third generation sub-loss based on the real channel information sample corresponding to the signal sample and the channel information sample; calculating a fourth generation sub-loss based on the second similarity parameter; and calculating the second target generation loss based on the third generation sub-loss and the fourth generation sub-loss.

[0239] Based on the actual channel information sample corresponding to the signal sample and the channel information sample, the loss function or calculation method used to calculate the third generator loss can be configured according to the actual situation. For example, the Mean Squared Error (MSE) loss function can be used to calculate the third generator loss, but this is not limited or exhaustive.

[0240] Based on the second similarity parameter, the loss function used to calculate the loss of the fourth generator can be any binary classification loss function, and this embodiment does not limit it.

[0241] The second target generation loss is calculated based on the third and fourth generation sub-losses. This can be achieved by calculating the second target generation loss based on the third generation sub-loss, its corresponding third weight, the fourth generation sub-loss, and its corresponding fourth weight. The third and fourth weights can be configured according to actual conditions, and this embodiment does not impose limitations. For example, the third generation sub-loss can be multiplied by the third weight to obtain a third value, the fourth generation loss can be multiplied by the fourth weight to obtain a fourth value, and the third and fourth values ​​can be added together to obtain the second target generation loss.

[0242] The second target discrimination loss is calculated based on the real channel information sample corresponding to the signal sample and the second similarity parameter. This can be achieved by: calculating the third discrimination sub-loss based on the real channel information sample corresponding to the signal sample; calculating the fourth discrimination sub-loss based on the second similarity parameter; and calculating the second target discrimination loss based on the third discrimination sub-loss and the fourth discrimination sub-loss.

[0243] In this embodiment, the third discriminant loss can be calculated using any binary classification loss function based on the real channel information sample corresponding to the signal sample. For example, the real channel information sample can be input into the first discriminator to obtain the similarity between the real channel information sample and itself, and then the loss value corresponding to the similarity can be calculated as the third discriminant loss based on the binary classification loss function.

[0244] Based on the second similarity parameter, the loss of the fourth discriminant can be calculated using any binary classification loss function, and this embodiment does not limit it.

[0245] The second discrimination target loss is calculated based on the third discriminant loss and the fourth discriminant loss. This can be achieved by adding the third discriminant loss and the fourth discriminant loss together to obtain the second discrimination target loss.

[0246] Based on the second target loss, the second generative model and the second discriminative model are jointly trained to obtain the trained second generative model and the trained second discriminative model. This can be achieved by: training the second generative model using the second generative target loss, and obtaining the trained second generative model if the training convergence condition of the second generative model is met; and training the second discriminative model using the second discriminative target loss, and obtaining the trained second discriminative model if the training convergence condition of the second discriminative model is met. In other words, the second initial generative model is trained using the second generative target loss, and the trained second generative model (i.e., the fourth generative model or the second target generative model) is obtained if the training convergence condition of the second generative model is met; the second initial discriminative model is trained using the second discriminative target loss, and the trained second discriminative model (i.e., the fourth discriminative model or the second target discriminative model) is obtained if the training convergence condition of the second discriminative model is met.

[0247] Training the second generative model (i.e., the second initial generative model) using the second generative target loss can refer to updating the parameters of the second generative model (i.e., the second initial generative model) based on the second generative target loss. Similarly, training the second discriminative model (i.e., the second initial discriminative model) using the second discriminative target loss can refer to updating the parameters of the second discriminative model (i.e., the second initial discriminative model) based on the second discriminative target loss.

[0248] The training convergence conditions for the second generative model and the second discriminative model are similar to those in the aforementioned embodiments, and will not be repeated here.

[0249] It should be noted that the joint training of the second generative model and the second discriminative model (i.e., the second initial generative model and the second initial discriminative model) can be performed in multiple loops. In each training loop, the second target loss can be calculated in the above manner.

[0250] It should also be noted that during the repeated training iterations, the second generative model can be trained using the second generative target loss and the second discriminative model using the second discriminative target loss in each training iteration. Alternatively, the second generative model can be trained first using the second generative target loss, with the parameters of the second generative model fixed, and then trained using the second discriminative target loss. This process can be repeated until the trained second generative model (i.e., the fourth generative model or the second target generative model) and the trained second discriminative model (i.e., the fourth discriminative model or the second target discriminative model) are obtained. Alternatively, the second discriminant model can be trained using the second discriminant target loss. With the parameters of the second discriminant model fixed, the second generative model can be trained using the second generative target loss. After fixing the parameters of the second generative model, the second discriminant model can be trained again using the second discriminant target loss. This process is repeated until a trained second generative model (i.e., the fourth generative model or the second target generative model) and a trained second discriminant model (i.e., the fourth discriminant model or the second target discriminant model) are obtained. This embodiment does not limit the training order of the second generative model and the second discriminant model during multiple iterations of training.

[0251] In some embodiments, the second target model is pre-configured on the second device side. The processing of the second device (terminal or access network device) may further include receiving a second target model from other electronic devices.

[0252] In this embodiment, the second target model is the trained second generative model (i.e., the second target generative model). The training of the second generative model can be performed by other electronic devices. The method of training the second generative model on the other electronic device side (that is, jointly training the second generative model and the second discriminant model (i.e., the second initial generative model and the second initial discriminant model) to obtain the trained second generative model (i.e., the second target generative model) and the trained second discriminant model (i.e., the second target discriminant model) is the same as in the previous embodiment and will not be repeated.

[0253] The other electronic device can be any electronic device with computing capabilities other than the terminal. For example, the other electronic device can be a server (such as any server or service node in a service cluster), or a core network device (such as any core network device on the core network side), or a user-end electronic device (such as a personal computer), etc. Here, we do not limit or exhaust all possible types of other electronic devices.

[0254] By adopting the above scheme, the second device can input the received air interface signal into the second target model, and obtain the denoised channel information through the second target model. In this way, even if the air interface signal is subjected to adversarial perturbation attacks, the second device can still process the adversarially perturbed air interface signal through the second target model to obtain denoised channel information. This ensures the accuracy of the channel information obtained by the second device, thereby ensuring the accuracy of the final obtained second compressed codeword and guaranteeing the quality of wireless communication.

[0255] Figure 10 is a schematic diagram of the composition structure of a first device according to an embodiment of the present application, including:

[0256] The first communication unit 1001 is used to receive first information;

[0257] The first processing unit 1002 is used to input the first information into the first target model and obtain the first compressed codeword after the first information has been de-perturbed by the first target model.

[0258] The first processing unit is configured to reconstruct the channel information between the first device and the second device based on the first compressed codeword.

[0259] The first device is an access network device, the second device is a terminal, and the channel information is downlink channel information; or, the first device is a terminal, the second device is an access network device, and the channel information is uplink channel information.

[0260] The first processing unit is configured to add perturbation to the compressed codeword sample to obtain perturbed codeword sample; based on the compressed codeword sample and the perturbed codeword sample, jointly train the first generation model and the first discriminant model to obtain the trained first generation model and the trained first discriminant model; and use the trained first generation model as the first target model.

[0261] The first processing unit is configured to input the perturbed codeword sample into the first generation model to obtain the deperturbed codeword sample output by the first generation model; based on the first discriminant model, obtain a first similarity parameter between the deperturbed codeword sample and the compressed codeword sample; calculate a first target loss based on the compressed codeword sample, the deperturbed codeword sample and the first similarity parameter; and jointly train the first generation model and the first discriminant model based on the first target loss to obtain the trained first generation model and the trained first discriminant model.

[0262] The first processing unit is configured to calculate a first target generation loss based on the compressed codeword sample, the de-perturbed codeword sample, and the first similarity parameter; calculate a first target discrimination loss based on the compressed codeword sample and the first similarity parameter; and obtain the first target loss based on the first target generation loss and the first target discrimination loss.

[0263] The first processing unit is configured to calculate a first generator loss based on the compressed codeword sample and the de-perturbed codeword sample; calculate a second generator loss based on the first similarity parameter; and calculate a first target generation loss based on the first generator loss and the second generator loss.

[0264] The first processing unit is configured to calculate a first discriminant loss based on the compressed codeword sample; calculate a second discriminant loss based on the first similarity parameter; and calculate the first target discriminant loss based on the first discriminant loss and the second discriminant loss.

[0265] The first target model is pre-configured.

[0266] Figure 11 is a schematic diagram of the composition structure of a second device according to an embodiment of this application, including:

[0267] The second processing unit 1102 is used to perform dimensional adjustment on the channel information between the first device and the second device to obtain adjusted channel information; and to encode the adjusted channel information to obtain a second compressed codeword.

[0268] The second communication unit 1101 is used to send the second compressed codeword.

[0269] The second processing unit is used to measure the reference signal from the first device to obtain the channel information.

[0270] The first device is an access network device, the second device is a terminal, and the channel information is downlink channel information; or, the first device is a terminal, the second device is an access network device, and the channel information is uplink channel information.

[0271] A second device according to an embodiment of this application includes:

[0272] The second communication unit is used to receive air interface signals;

[0273] The second processing unit is used to input the air interface signal into the second target model to obtain the denoised channel information between the second device and the first device output by the second target model.

[0274] The second processing unit is used to encode the channel information to obtain a second compressed codeword;

[0275] The second communication processing unit is used to send the second compressed codeword.

[0276] The second processing unit is used to add perturbation to the signal sample to obtain a perturbed signal sample; based on the real channel information sample corresponding to the signal sample and the perturbed signal sample, jointly train the second generation model and the second discrimination model to obtain the trained second generation model and the trained second discrimination model; and use the trained second generation model as the second target model.

[0277] The second processing unit is configured to input the perturbation signal sample into the second generation model to obtain the channel information sample output by the second generation model; based on the second discrimination model, obtain a second similarity parameter between the channel information sample and the real channel information sample corresponding to the signal sample; calculate a second target loss based on the real channel information sample corresponding to the signal sample, the channel information sample, and the second similarity parameter; and jointly train the second generation model and the second discrimination model based on the second target loss to obtain the trained second generation model and the trained second discrimination model.

[0278] The second target model is pre-configured.

[0279] The first device is an access network device, the second device is a terminal, and the channel information is downlink channel information; or, the first device is a terminal, the second device is an access network device, and the channel information is uplink channel information.

[0280] The device in this application embodiment can realize the corresponding functions of the various devices in the foregoing communication method embodiments. The processes, functions, implementation methods, and beneficial effects of each module (sub-module, unit, or component, etc.) in this device can be found in the corresponding descriptions in the above method embodiments, and will not be repeated here. It should be noted that the functions described for each module (sub-module, unit, or component, etc.) in the device of this application embodiment can be implemented by different modules (sub-modules, units, or components, etc.) or by the same module (sub-module, unit, or component, etc.).

[0281] It should be understood that the sequence number of each process in the various embodiments of this application does not imply the order of execution; the execution order of each process should be determined by its function and internal logic. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. The above descriptions are merely specific embodiments of this application, and the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method of communication performed by a first device, comprising: receiving first information; inputting the first information into a first target model to obtain a first compressed codeword output by the first target model, the first compressed codeword being de-disturbed from the first information.

2. The method of claim 1, wherein, The method further comprises: reconstructing channel information between the first device and a second device based on the first compressed codeword. 3.The method of claim 2, wherein: the first device is an access network device, the second device is a terminal, and the channel information is downlink channel information; or the first device is a terminal, the second device is an access network device, and the channel information is uplink channel information. 4.The method of any one of claims 1-3, further comprising: adding a disturbance to a compressed codeword sample to obtain a disturbed codeword sample; jointly training a first generation model and a first discriminative model based on the compressed codeword sample and the disturbed codeword sample to obtain a trained first generation model and a trained first discriminative model; using the trained first generation model as the first target model. The jointly training the first generation model and the first discriminative model based on the compressed codeword sample and the disturbed codeword sample to obtain the trained first generation model and the trained first discriminative model comprises:

5. The method of claim 4, wherein, inputting the disturbed codeword sample into the first generation model to obtain a de-disturbed codeword sample output by the first generation model; obtaining a first similarity parameter between the de-disturbed codeword sample and the compressed codeword sample based on the first discriminative model; calculating a first target loss based on the compressed codeword sample, the de-disturbed codeword sample, and the first similarity parameter; jointly training the first generation model and the first discriminative model based on the first target loss to obtain the trained first generation model and the trained first discriminative model. The calculating the first target loss based on the compressed codeword sample, the de-disturbed codeword sample, and the first similarity parameter comprises:

6. The method of claim 5, wherein, calculating a first target generation loss based on the compressed codeword sample, the de-disturbed codeword sample, and the first similarity parameter; calculating a first target discriminative loss based on the compressed codeword sample and the first similarity parameter; obtaining the first target loss based on the first target generation loss and the first target discriminative loss. The calculating the first target generation loss based on the compressed codeword sample, the de-disturbed codeword sample, and the first similarity parameter comprises:

7. The method of claim 6, wherein, calculating a first generation sub-loss based on the compressed codeword sample and the de-disturbed codeword sample; calculating a second generation sub-loss based on the first similarity parameter; calculating the first target generation loss based on the first generation sub-loss and the second generation sub-loss. The calculating the first target discriminative loss based on the compressed codeword sample and the first similarity parameter comprises:

8. The method of claim 6, wherein, calculating a first discriminative sub-loss based on the compressed codeword sample; calculating a second discriminative sub-loss based on the first similarity parameter; calculating the first target discriminative loss based on the first discriminative sub-loss and the second discriminative sub-loss. ​ 9. The method of any one of claims 1-3, wherein, The first target model is preconfigured.

10. A communication method performed by a second device, comprising: dimensionally adjusting channel information between a first device and the second device to obtain adjusted channel information; encoding the adjusted channel information to obtain a second compressed codeword; transmitting the second compressed codeword.

11. The method of claim 10, wherein, The method further comprises: measuring a reference signal from the first device to obtain the channel information.

12. The method of claim 10 or 11, wherein: the first device is an access network device, the second device is a terminal, and the channel information is downlink channel information; or the first device is a terminal, the second device is an access network device, and the channel information is uplink channel information.

13. A communication method performed by a second device, comprising: receiving an air interface signal; inputting the air interface signal into a second target model to obtain channel information between the second device and a first device output by the second target model after disturbance removal.

14. The method of claim 13, further comprising: encoding the channel information to obtain a second compressed codeword; transmitting the second compressed codeword.

15. The method of claim 13 or 14, further comprising: adding disturbance to signal samples to obtain disturbed signal samples; jointly training a second generation model and a second discriminant model based on real channel information samples corresponding to the signal samples and the disturbed signal samples to obtain a trained second generation model and a trained second discriminant model; using the trained second generation model as the second target model. The jointly training the second generation model and the second discriminant model based on the real channel information samples corresponding to the signal samples and the disturbed signal samples to obtain the trained second generation model and the trained second discriminant model comprises:

16. The method of claim 15, wherein, inputting the disturbed signal samples into the second generation model to obtain channel information samples output by the second generation model; obtaining a second similarity parameter between the channel information samples and real channel information samples corresponding to the signal samples based on the second discriminant model; calculating a second target loss based on the real channel information samples corresponding to the signal samples, the channel information samples, and the second similarity parameter; jointly training the second generation model and the second discriminant model based on the second target loss to obtain the trained second generation model and the trained second discriminant model. The second target model is preconfigured.

17. The method of claim 13 or 14, wherein, 18. The method of any one of claims 13-17, wherein: the first device is an access network device, the second device is a terminal, and the channel information is downlink channel information; or the first device is a terminal, the second device is an access network device, and the channel information is uplink channel information.

19. A first device, comprising: a first communication unit configured to receive first information; a first processing unit configured to input the first information into a first target model to obtain a first compressed codeword output by the first target model after disturbance removal on the first information.

20. A second device, comprising: ​ ​ a second processing unit, configured to perform dimension adjustment on channel information between the first device and the second device to obtain adjusted channel information; encode the adjusted channel information to obtain a second compressed code word; a second communication unit, configured to send the second compressed code word.

21. A second device, comprising: a second communication unit, configured to receive an air interface signal; a second processing unit, configured to input the air interface signal into a second target model to obtain disturbance-removed channel information between the second device and a first device output by the second target model.

Images