Communication method, and devices

Abstract

The present application relates to a communication method, and devices. The method comprises: receiving first information; and on the basis of the first information and a target model, obtaining a first compressed codeword after perturbation removal is performed on the first information.

Description

Communication method and device TECHNICAL FIELD

[0001] The present application relates to the field of communication, and more particularly, to a communication method and device. BACKGROUND

[0002] In a large-scale multiple-input multiple-output communication system, feedback of channel information is crucial to system performance. In order to reduce the feedback overhead of channel information and improve the feedback accuracy of channel information, AI / ML (Artificial Intelligence / Machine Learning) is introduced into the feedback of channel information in the related technology. In this scheme, an encoder can be deployed at the sending end of channel information to encode to obtain compressed code words, and a decoder can be deployed at the receiving end of channel information to recover or reconstruct the channel information.

[0003] However, due to the open nature of the wireless channel and the fragile nature of the neural network, the compressed code words transmitted between the encoder and the decoder are vulnerable to adversarial perturbation attacks by attackers. Therefore, how to ensure that the receiving end can accurately reconstruct the channel information under the condition that the compressed code words are subjected to adversarial perturbation attacks during transmission has become a problem to be solved.

[0004] SUMMARY

[0005] Embodiments of the present application provide a communication method and device.

[0006] Embodiments of the present application provide a communication method performed by a first device, comprising:

[0007] receiving first information;

[0008] obtaining, based on the first information and a target model, a first compressed code word after the first information is de-perturbed.

[0009] Embodiments of the present application provide a communication method performed by a second device, comprising:

[0010] dimensionally processing the second compressed code word to obtain a third compressed code word;

[0011] sending the third compressed code word.

[0012] Embodiments of the present application provide a first device, comprising:

[0013] a first communication unit configured to receive first information;

[0014] a first processing unit configured to obtain, based on the first information and a target model, a first compressed code word after the first information is de-perturbed.

[0015] The embodiment of the present application provides a second device, comprising:

[0016] a second processing unit, configured to perform dimension reduction processing on the second compressed code word to obtain a third compressed code word;

[0017] a second communication unit, configured to send the third compressed code word.

[0018] By adopting the above scheme, after receiving the first information, the first device side obtains the first compressed code word after the first information is de-disturbed through the target model. Thus, in the case that the compressed code word is transmitted between the second device and the first device and is subjected to the adversarial disturbance attack, the first device can obtain the de-disturbed first compressed code word through the target model processing, so that the accuracy of the first compressed code word obtained by the first device side is ensured, thereby ensuring the accuracy of the channel information reconstructed by the first device based on the first compressed code word, and further ensuring the quality of wireless communication. BRIEF DESCRIPTION OF DRAWINGS

[0019] FIG. 1 is a schematic diagram of an application scenario according to an embodiment of the present application.

[0020] FIG. 2 is a schematic flowchart of a communication method according to an embodiment of the present application.

[0021] FIG. 3 is a schematic flowchart of a communication method according to another embodiment of the present application.

[0022] FIG. 4 is a schematic flowchart of a communication method according to still another embodiment of the present application.

[0023] FIG. 5 is a schematic diagram of a communication method according to an embodiment of the present application.

[0024] FIG. 6 is a schematic flowchart of a communication method according to still another embodiment of the present application.

[0025] FIG. 7 is a schematic diagram of a communication method according to another embodiment of the present application.

[0026] FIG. 8 is a schematic diagram of a training scenario of a first generator and a first discriminator according to an embodiment of the present application.

[0027] FIG. 9 is a schematic flowchart of training of a first generator and a first discriminator according to an embodiment of the present application.

[0028] FIG. 10 is another schematic flowchart of a communication method according to an embodiment of the present application.

[0029] FIG. 11 is another schematic diagram of a communication method according to an embodiment of the present application.

[0030] FIG. 12 is a schematic block diagram of a first device according to an embodiment of the present application.

[0031] FIG. 13 is a schematic block diagram of a second device according to an embodiment of the present application. DETAILED DESCRIPTION

[0032] The technical solutions of the embodiments of the present application can be applied to various communication systems, for example, 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, etc.

[0033] The embodiments of the present application describe various embodiments in combination with network devices and terminals. The terminals can be mobile or fixed, and can also be referred to as mobile stations, user units, etc. The terminals can be stations in WLAN, and can be smart terminals, wireless modems, notebook computers, tablet computers, etc. In the embodiments of the present application, the terminals can be VR (Virtual Reality) terminals / AR (Augmented Reality) terminals, industrial control terminals, unmanned driving terminals, remote medical terminals, smart grid terminals, transportation safety terminals, smart city terminals, or wireless terminals of smart homes, etc. As an example but not limitation, in the embodiments of the present application, the terminals can also be wearable devices.

[0034] In the embodiments of the present application, the network devices can be devices for communicating with the terminals. The network devices can be access points in WLAN, can be evolved base stations in LTE, or relay stations, or network devices in vehicle-mounted devices, wearable devices, and NR networks (gNB, the next Generation Node B), or network devices in future evolved PLMN (Public Land Mobile Network), or network devices in non-ground networks, etc. As an example but not limitation, in the embodiments of the present application, the network devices can have mobile characteristics, for example, the network devices can be mobile devices.

[0035] In order to facilitate understanding of the technical solutions of the embodiments of the present application, the related technologies of the embodiments of the present application are described below. The following related technologies can be combined with the technical solutions of the embodiments of the present application in any way, and all belong to the protection scope of the embodiments of the present application.

[0036] FIG. 1 illustrates a communication system 100. The communication system includes a network device 110 and a terminal 120. In a possible implementation, the communication system 100 can include a plurality of network devices 110, and each network device 110 can include one or more terminals 120 within its coverage, which is not limited in the embodiments of the present application. In a possible implementation, the communication system 100 can further include a mobility management entity, an access and mobility management function, and other network entities, which are not limited in the embodiments of the present application. The network device can include an access network device and a core network device. That is, the communication system can include a plurality of core networks for communicating with the access network device. The access network device can be a base station of an LTE, LTE-A, or NR system. For example, the communication system shown in FIG. 1 can include network devices and terminals with communication functions, and can further include other devices in the communication system, such as a network controller, a mobility management entity, and other network entities, which are not limited in the embodiments of the present application.

[0037] FIG. 2 is a schematic flowchart of a communication method performed by a first device according to an embodiment of the present application. The method includes at least part of the following.

[0038] S210, receiving first information;

[0039] S220, obtaining a first compressed codeword after the first information is de-disturbed based on the first information and a target model.

[0040] FIG. 3 is a schematic flowchart of a communication method performed by a second device according to an embodiment of the present application. The method includes at least part of the following.

[0041] S310, performing dimension reduction processing on the second compressed codeword to obtain a third compressed codeword;

[0042] S320, transmitting the third compressed codeword.

[0043] The first device is a receiving end of the compressed codeword, and the second device is a transmitting end of the compressed codeword. In some possible examples, the first device can also be alternatively referred to as a receiving end or a receiving end device, and the second device can also be alternatively referred to as a transmitting end or a transmitting end device.

[0044] 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.

[0045] Specifically, in different scenarios, the first device and the second device refer to different devices. The embodiments define the uplink scenario or the downlink scenario with reference to the reference signal or the channel information. In the downlink scenario referred to below, the reference signal is the downlink reference signal, and the channel information is the downlink channel information. In the downlink scenario, the first device is the access network device, that is, the access network device is the receiving end, and the second device is the terminal, that is, the terminal is the sending end. In the uplink scenario referred to below, the reference signal is the uplink reference signal, and the channel information is the uplink channel information. In the uplink scenario, the first device is the terminal, that is, the terminal is the receiving end, and the second device is the access network device, that is, the access network device is the sending end. It should be pointed out that the above is to divide the uplink and downlink scenarios according to the uplink and downlink of the reference signal or the channel information. In some possible examples, the uplink or downlink scenario can also be defined according to the transmission direction of the compressed code word. The embodiments mainly define the uplink and downlink scenarios according to the uplink and downlink transmission directions of the reference signal for subsequent convenience of description, but the embodiments do not limit the definition of the uplink and downlink scenarios.

[0046] In some possible implementation manners, the processing of the second device can include: measuring the reference signal from the first device to obtain the channel information between the first device and the second device; and encoding the channel information to obtain the second compressed code word. Correspondingly, the processing of the first device can include: transmitting the reference signal.

[0047] In this implementation manner, after the second device obtains the second compressed code word, the second device can transmit the second compressed code word to the first device.

[0048] After the first device transmits the reference signal, the first device can receive the first information, and then obtain the first compressed code word after the first information is disturbed based on the first information and the target model. After the first device obtains the first compressed code word, the processing of the first device can include: decoding the first compressed code word to obtain the reconstructed channel information between the first device and the second device.

[0049] In the first device, the first compressed code word after the first information is disturbed based on the first information and the target model includes: generating a plurality of first random noise vectors; obtaining a first target noise vector based on the first information and the plurality of first random noise vectors; and inputting the first target noise vector into the target model to obtain the first compressed code word after the first information is disturbed output by the target model.

[0050] In this way, after receiving the first information, the first device obtains the first compressed codeword after the first information is de-disturbed through the target model. In this way, in the case that the transmission of the compressed codeword between the second device and the first device is subjected to the adversarial disturbance attack, the first device can obtain the first compressed codeword after the de-disturbance through the processing of the target model. In this way, the accuracy of the first compressed codeword obtained by the first device is ensured, so that the accuracy of the channel information reconstructed by the first device based on the first compressed codeword is ensured, and then the quality of the wireless communication is ensured.

[0051] In some embodiments, in a 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.

[0052] The processing of the terminal includes: measuring a downlink reference signal from the access network device to obtain downlink channel information between the terminal and the access network device; and encoding the downlink channel information to obtain a second compressed codeword.

[0053] The timing at which the access network device transmits the downlink reference signal is within the protection scope of the present embodiment as long as the timing is before the first information is received. The downlink reference signal can be any one of a CSI-RS (Channel State Information Reference Signal), a DMRS (Demodulation Reference Signal), and the like, and the specific type of the downlink reference signal is not limited or enumerated herein.

[0054] The terminal measures the downlink reference signal from the access network device to obtain the downlink channel information between the terminal and the access network device. This can include: the terminal measures the downlink reference signal from the access network device, performs channel estimation based on the measurement result, and obtains the downlink channel information between the terminal and the access network device.

[0055] The downlink channel information can include CSI (Channel State Information) and the like.

[0056] Here, the algorithm used for channel estimation can be configured according to actual conditions, such as a least square channel estimation algorithm and the like.

[0057] For example, the terminal measures a downlink reference signal from the access network device, performs channel estimation based on the measurement result, and obtains downlink channel information between the terminal and the access network device. For example, the terminal can measure the downlink reference signal from the access network device through a receiver, obtain a pilot sequence of the downlink reference signal, and then perform channel estimation on the pilot sequence of the downlink reference signal through a channel estimation algorithm (for example, least square channel estimation) based on a pre-stored pilot sequence and the received pilot sequence (i.e., the pilot sequence of the downlink reference signal), to obtain channel information of a channel for transmitting the downlink reference signal, and take the channel information of the channel for transmitting the downlink reference signal as the downlink channel information. It should be noted that the above is only an example of obtaining the downlink channel information by measuring the downlink reference signal by the terminal, and other methods for obtaining the downlink channel information can also be used in actual processing, which is not limited or exhausted by the present embodiment.

[0058] The terminal encodes the downlink channel information to obtain the second compressed code word, which can include that the terminal compresses the downlink channel information through an encoder to obtain the second compressed code word. That is, the terminal can input the downlink channel information into the encoder to obtain the second compressed code word output by the encoder. The encoder can be pre-configured in the terminal, and the generation or training method of the encoder is not limited by the present embodiment.

[0059] The processing of the terminal after obtaining the second compressed code word can include that the terminal transmits the second compressed code word.

[0060] Specifically, the terminal transmitting the second compressed code word can include that the terminal transmits the second compressed code word to the access network device through an air interface. The terminal can use any type of uplink access layer (Access Stratum, AS) message or information to carry or transmit the second compressed code word, which is not limited by the present embodiment.

[0061] Correspondingly, the access network device can receive the first information. In the present embodiment, the terminal transmits the second compressed code word through the air interface, and the access network device receives the first information through the air interface. This is because the second compressed code word can be subjected to an adversarial attack during transmission between the terminal and the access network device, that is, the access network device side can receive information obtained by adding disturbance or interference to the second compressed code word. Therefore, the information obtained by adding disturbance or interference to the second compressed code word and received by the access network device through the air interface is referred to as the first information.

[0062] Further, after the access network device receives the first information, a de-disturbed compressed code word is obtained based on the target model and the first information. In theory, the de-disturbed compressed code word obtained by the access network device should be the same as or infinitely close to or similar to the compressed code word generated by the terminal, but in order to distinguish between the de-disturbed compressed code word obtained by the access network device and the compressed code word generated by the terminal, the embodiment refers to the de-disturbed compressed code word obtained by the access network device as a first compressed code word, and refers to the compressed code word generated by the terminal as a second compressed code word, which will not be repeated below.

[0063] At the access network device side, based on the first information and the target model, a first compressed code word after de-disturbance of the first information is obtained, including: generating a plurality of first random noise vectors; based on the first information and the plurality of first random noise vectors, a first target noise vector is obtained; the first target noise vector is input into the target model, and the first compressed code word after de-disturbance of the first information output by the target model is obtained.

[0064] Wherein, the plurality of first random noise vectors can be generated by a random noise generator in the access network device, and the random noise generator generates the plurality of first random noise vectors in a manner which is not limited in the embodiment. The number of the plurality of first random noise vectors is not limited in the embodiment, as long as it is greater than or equal to 2, it is within the protection scope of the embodiment. The noise vector can also be referred to as a noise variable, and the embodiment does not limit or exhaust all possible names thereof.

[0065] The target model is a neural network with generation capability, and in the embodiment, the function of the target model is to generate data or code words similar or close to the distribution of the real compressed code word (i.e. the second compressed code word).

[0066] The composition structure of the target model can include at least one of the following: a fully connected neural network, a CNN (Convolutional Neural Networks), a DNN (Deconvolutional Neural Network), an RNN (recurrent neural network) (or a variant of RNN, such as LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit)), a GAAE (Generative Adversarial Autoencoder), and the like, and the possible composition structure of the target model is not limited or exhausted here.

[0067] The first target noise vector can be a feature vector closest to a real feature. Here, the real feature is a feature of real data, which in this embodiment can be the first information received by the access network device; the feature vector closest to the real feature can also be referred to as a feature vector close to the real feature, or a real vector closest to the real feature, or an optimal noise vector.

[0068] The first target noise vector can be a feature vector closest to a real feature. Here, the real feature is a feature of real data, which in this embodiment can be the first information received by the access network device; the feature vector closest to the real feature can also be referred to as a feature vector close to the real feature, or a real vector closest to the real feature, or an optimal noise vector.

[0069] The mapping network can also be alternatively referred to as a mapping model, or a mapping neural network, etc., and the present embodiment does not limit or exhaustively enumerate all possible names of the mapping network.

[0070] The function of the mapping network is to generate a feature vector close to a real feature according to a set of random noise vectors (i.e., a plurality of first random noise vectors (or a plurality of first random noise variables)). In this embodiment, the feature vector closest to the real feature is the first target noise vector, which can be the same as any one of the first random noise vectors, or different from each of the first random noise vectors. That is, in this embodiment, the input information or data of the mapping network includes a plurality of first random noise vectors and the first information, and the output information or data is the feature vector closest to the real feature (the first target noise vector).

[0071] In some possible examples, the mapping network can use a GD (Gradient Descent) algorithm in its composition or construction or processing. For example, the mapping network can use the gradient descent algorithm and a set of random noise vectors (i.e., a plurality of first random noise vectors (or a plurality of first random noise variables)) to generate a feature vector (i.e., the first target noise vector) closer to the real feature (the feature of the first information).

[0072] In other possible examples, the mapping network can also use other algorithms or other processing methods. For example, the mapping network can determine a feature vector (i.e., the first target noise vector) closer to the real feature (the feature of the first information) from a plurality of first random noise vectors based on a distribution similarity between each first random noise vector and the first information; and the present embodiment does not limit the calculation method of the distribution similarity between each first random noise vector and the first information.

[0073] It should be noted that the above is only an exemplary description of the construction (or composition) or processing manner of the mapping network, and the present embodiment does not limit or exhaust the training, composition or construction, processing manner, etc. of the mapping network, as long as the mapping network can generate a feature vector closest to the real feature according to a plurality of first random noise vectors.

[0074] The processing after the access network device obtains the first compressed code word can further include: decoding the first compressed code word to obtain reconstructed downlink channel information between the access network device and the terminal.

[0075] Specifically, the decoding of the first compressed code word by the access network device to obtain the reconstructed downlink channel information between the access network device and the terminal can be: inputting the first compressed code word into a decoder to obtain the reconstructed downlink channel information between the access network device and the terminal.

[0076] It should be noted that the downlink channel information reconstructed by the decoder on the access network device side and the downlink channel information measured by the terminal on the terminal side should be the same or similar or infinitely close in theory. In some possible examples, in order to distinguish the downlink channel information reconstructed by the access network device and the downlink channel information measured by the terminal, the downlink channel information measured by the terminal can be referred to as original downlink channel information, or original downlink CSI information, or original CSI, or downlink CSI, etc., and the reconstructed downlink channel information obtained by the access network device can be referred to as reconstructed downlink channel information, or recovered CSI, or reconstructed CSI information, or reconstructed original CSI, or reconstructed original CSI information, etc., which will not be repeated hereinafter.

[0077] In the present 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, and the composition, training manner, etc. of the encoder and the decoder are not limited in the present embodiment.

[0078] In combination with FIG. 4 and FIG. 5, the communication method provided by the present embodiment is exemplarily described, including:

[0079] S400: A BS (Base Station, access network device) sends a downlink reference signal. The downlink reference signal can include CSI-RS, etc.

[0080] S401: A UE (User Equipment, terminal) measures the downlink reference signal to obtain original CSI.

[0081] Here, after receiving the reference signal, the UE can perform channel estimation. In one possible implementation, the UE receiver estimates the channel information (original CSI) of the channel through which the downlink reference signal is transmitted, based on the pre-stored pilot sequence and the received pilot sequence, through a channel estimation algorithm (for example, least square channel estimation).

[0082] S402: The UE compresses the original CSI through an encoder to obtain a second compressed code word.

[0083] For example, in combination with FIG. 5, the UE inputs the original CSI into the encoder to obtain the second compressed code word compressed by the encoder.

[0084] S403: The UE sends the second compressed code word compressed by the encoder to the BS through a wireless channel (air interface).

[0085] For example, in combination with FIG. 5, the second compressed code word is sent by the transmitter on the UE side to the BS through the wireless channel (air interface).

[0086] S404: The BS receives the wireless signal to obtain the received first information; and the BS finds the optimal random noise vector (i.e., the first target noise vector in the foregoing embodiment) using the first information.

[0087] For example, the first information is received by the receiver on the BS side from the wireless channel (or air interface). Specifically, there can be an attacker (for example, an attack device, etc.) in the wireless channel that adds disturbance to the second compressed code word transmitted in the wireless channel, and thus the received information on the BS side can be the first information obtained by adding disturbance to the second compressed code word.

[0088] Further, in combination with FIG. 5, S404 specifically includes: generating, by the BS, R first random noise vectors {z1, z2, …, zR} using a random noise generator, where R is an integer greater than or equal to 2; and generating, using a mapping network, the optimal random noise vector z* based on the first information and the R first random noise vectors. R * (i.e., the feature vector closest to the true feature (the feature of the first information)).

[0089] S405: The BS performs de-disturbance processing on the optimal random noise vector through a target model deployed on the BS side to obtain a clean intermediate code word (the clean intermediate code word is the first compressed code word in the foregoing embodiment).

[0090] For example, in combination with FIG. 5, the optimal random noise vector is input into the target model on the BS side to obtain the clean intermediate code word de-disturbed by the target model.

[0091] ​S406: The BS inputs the clean intermediate codeword to the decoder to recover the CSI.

[0092] For example, in combination with FIG. 5, the BS inputs the clean intermediate codeword after the disturbance removed by the target model to the decoder to obtain the recovered CSI output by the decoder.

[0093] By using the above scheme, after receiving the first information, the access network device side obtains the first compressed codeword after the disturbance of the first information by using the target model. In this way, in the case that the compressed codeword is transmitted between the terminal and the access network device and is subjected to the adversarial disturbance attack, the access network device can also obtain the first compressed codeword after the disturbance by processing the target model, so as to ensure the accuracy of the first compressed codeword obtained by the access network device side, thereby ensuring the accuracy of the downlink channel information reconstructed by the access network device based on the first compressed codeword, and further ensuring the quality of wireless communication.

[0094] 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 is uplink channel information.

[0095] The processing of the access network device includes: measuring the uplink reference signal from the terminal to obtain the uplink channel information between the terminal and the access network device; and encoding the uplink channel information to obtain the second compressed codeword.

[0096] The terminal transmits the uplink reference signal at a time before receiving the first information, which is within the protection scope of the present embodiment. The uplink reference signal can be any one of SRS (Sounding Reference Signal), DMRS, etc., and the specific type of the uplink reference signal is not limited or enumerated herein.

[0097] 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, which can be that the access network device measures the uplink reference signal from the terminal, performs channel estimation based on the measurement result, and obtains the uplink channel information between the terminal and the access network device. Here, the channel estimation algorithm used by the access network device side can be least square channel estimation or other channel estimation algorithms, which are not limited or enumerated herein.

[0098] The access network device encodes the uplink channel information to obtain the second compressed codeword, which can include that the access network device can input the uplink channel information to the encoder to obtain the second compressed codeword output by the encoder. The encoder can be pre-configured in the access network device, and the generation or training method of the encoder is not limited herein.

[0099] The processing of the access network device after obtaining the second compressed code word can comprise: sending the second compressed code word. Wherein, the access network device can use any type of downlink AS (Access Stratum, access layer) message or information to carry or transmit the second compressed code word, and the embodiments are not limited.

[0100] Correspondingly, the terminal can receive the first information. In the present embodiment, the access network device sends the second compressed code word through the air interface, and the terminal receives the first information through the air interface. This is because the second compressed code word may be subjected to an adversarial attack in the process of wireless channel transmission between the terminal and the access network device, that is, the terminal can receive information obtained by adding disturbance or interference to the second compressed code word, so a distinction is made in the description.

[0101] Further, after the terminal receives the first information, the first compressed code word after disturbance removal is obtained based on the target model and the first information. In theory, the first compressed code word obtained by the terminal should be the same as or infinitely close to or similar to the compressed code word generated by the access network device, but in order to distinguish the first compressed code word obtained by the terminal and the second compressed code word generated by the access network device, the first compressed code word obtained by the terminal is referred to as the first compressed code word, and the compressed code word generated by the access network device is referred to as the second compressed code word, which will not be repeated in the following description.

[0102] On the terminal side, the first compressed code word after disturbance removal of the first information is obtained based on the first information and the target model, comprising: generating a plurality of first random noise vectors; obtaining a first target noise vector based on the first information and the plurality of first random noise vectors; inputting the first target noise vector into the target model to obtain the first compressed code word after disturbance removal of the first information output by the target model. Wherein, the related description of the plurality of first random noise vectors, the target model and the first target noise vector is similar to the foregoing embodiments, and will not be repeated.

[0103] On the terminal side, the first target noise vector is obtained based on the first information and the plurality of first random noise vectors, which can be: the terminal obtains the first target noise vector based on the first information, the plurality of first random noise vectors and the mapping network. Here, the related description of the mapping network and its processing is also similar to the foregoing embodiments, and the only difference is that the mapping network is arranged on the terminal side in the present embodiment, while the mapping network is arranged on the access network device side in the foregoing embodiments, so no repeated description is made.

[0104] The processing after the terminal obtains the first compressed code word can further comprise: the terminal can input the first compressed code word into a decoder for decoding to obtain reconstructed uplink channel information between the access network device and the terminal.

[0105] It should be noted that the uplink channel information reconstructed by the decoder at the terminal side and the uplink channel information measured by the access network device on the uplink reference signal should be the same or similar or infinitely close in theory. In some possible examples, in order to distinguish the uplink channel information reconstructed by the terminal and the uplink channel information measured by the access network device, the uplink channel information measured by the access network device can be referred to as original uplink channel information, and the reconstructed uplink channel information obtained by the terminal can be referred to as reconstructed uplink channel information or recovered uplink channel information, and the like, which will not be repeatedly explained below.

[0106] In this embodiment, the decoder at the terminal side and the encoder of the access network device can be pre-trained and matched with each other, and the composition of the encoder and the decoder, the training manner and the like are not limited in this embodiment.

[0107] In combination with FIGS. 6 and 7, a communication method provided by the embodiment is exemplarily explained, including the following steps.

[0108] S600: The UE sends the uplink reference signal.

[0109] S601: The BS measures the uplink reference signal to obtain original uplink channel information.

[0110] S602: The BS compresses the original uplink channel information by using the encoder to obtain a second compressed code word.

[0111] For example, in combination with FIG. 7, the BS inputs the original uplink channel information into the encoder to obtain the second compressed code word compressed by the encoder.

[0112] S603: The BS sends the second compressed code word compressed by the encoder to the UE through a wireless channel (air interface).

[0113] For example, in combination with FIG. 7, the second compressed code word is sent by the transmitter at the BS side to the UE through the wireless channel (air interface).

[0114] S604: The UE receives the wireless signal to obtain received first information, and finds an optimal random noise vector (i.e., a first target noise vector) by using the first information.

[0115] For example, the first information is received by the receiver of the UE from the wireless channel (or air interface). Specifically, there can be an attacker (for example, an attack device or the like) in the wireless channel to add disturbance to the second compressed code word transmitted in the wireless channel, and therefore, the first information received at the UE side can be the first information obtained by adding disturbance to the second compressed code word.

[0116] Further, in combination with FIG. 7, it is specified that: the UE generates R first random noise vectors {z1, z2, …, zR} using a random noise generator, R is an integer greater than or equal to 2; and an optimal random noise vector z* is generated based on the first information, the R first random noise vectors, and a mapping network. R} is generated based on the first information, the R first random noise vectors, and a mapping network. * (i.e., the feature vector closest to the real feature (the feature of the first information)).

[0117] S605: The UE performs a de-disturbance processing on the optimal random noise vector by a target model deployed at the UE side, to obtain a clean intermediate code word (the clean intermediate code word is the first compressed code word in the foregoing embodiments).

[0118] For example, in combination with FIG. 7, the optimal random noise vector is input into the target model at the UE side, to obtain the clean intermediate code word output by the target model after the de-disturbance.

[0119] S606: The UE inputs the clean intermediate code word into a decoder, to recover the uplink channel information.

[0120] For example, in combination with FIG. 7, the clean intermediate code word output by the target model after the de-disturbance is input into the decoder at the UE side, to obtain the recovered uplink channel information output by the decoder.

[0121] By adopting the above scheme, after receiving the first information, the terminal obtains the first compressed code word after the de-disturbance of the first information by the target model. In this way, in the case that the transmission of the compressed code word between the terminal and the access network device is subjected to the adversarial disturbance attack, the terminal can also obtain the de-disturbed first compressed code word by the processing of the target model. Thus, the accuracy of the first compressed code word obtained by the terminal is ensured, so that the accuracy of the uplink channel information reconstructed by the terminal based on the first compressed code word is ensured, and further the quality of the wireless communication is ensured.

[0122] In some possible implementation manners, the target model can be trained at the first device side. The timing at which the first device trains the target model is within the protection scope of the present embodiment, as long as the timing is before the first information is received. In the present implementation manner, the first device can be an access network device or a terminal.

[0123] The manner in which the first device trains the target model can include: generating a plurality of second random noise vectors; obtaining a first noise sample based on a sample compressed code word and the plurality of second random noise vectors; jointly training a first generation model and a first discrimination model based on the first noise sample and the sample compressed code word, to obtain a trained second generation model and a trained second discrimination model; and taking the trained second generation model as the target model.

[0124] That is, the first generation model is trained on the first device side, the first generation model is trained in conjunction with the first discriminant model, and the first device obtains the trained second generation model as the target model. In some examples below, for the sake of brevity, the process of training the first generation model and the first discriminant model to obtain the trained second generation model and the trained second discriminant model, and taking the trained second generation model as the target model, can be simply described as training the target model, and will not be repeated below.

[0125] Here, the first generation model can refer to an untrained generation model, or an untrained original generation model, or an untrained initial generation model. The trained second generation model can refer to a final generation model obtained after training is completed, or a target generation model obtained after training is completed, or a target generation model obtained after joint training is completed. It should be understood that the present embodiment refers to the untrained original (or initial) generation model as the first generation model, and refers to the final generation model (or target generation model) obtained after training is completed as the trained second generation model, which is described to distinguish the generation model in different stages before and after training is completed. The composition or structure of the generation model in different stages before and after training is completed should be the same, but the parameters can be different. The above distinction is not intended to represent two different generation models (i.e., not intended to represent two generation models with different structures), and will not be repeated below. In addition, in some possible examples, the generation model can also be referred to as a generator, or a generation network, etc.

[0126] The first discriminant model can refer to an untrained discriminant model, or an untrained original discriminant model, or an untrained initial discriminant model. The trained second discriminant model can refer to a final discriminant model obtained after training is completed, or a target discriminant model obtained after training is completed, or a target discriminant model obtained after joint training is completed. It should be understood that the present embodiment refers to the untrained original (or initial) discriminant model as the first discriminant model, and refers to the final discriminant model (or target discriminant model) obtained after training is completed as the trained second discriminant model, which is described to distinguish the discriminant model in different stages before and after training is completed. The composition or structure of the discriminant model in different stages before and after training is completed should be the same, but the parameters can be different. The above distinction is not intended to represent two different discriminant models (i.e., not intended to represent two discriminant models with different structures), and will not be repeated below. In addition, in some possible examples, the discriminant model can also be referred to as a discriminator, or a classifier, or a classification model, etc., which is not limited or exhaustive here.

[0127] The plurality of second random noise vectors can be generated by a random noise generator. The number of the plurality of second random noise vectors is greater than 1, and the number of the plurality of second random noise vectors is not limited in this embodiment.

[0128] The sample compression code word can be pre-acquired or generated. The acquisition or generation manner of the sample compression code word is not limited in this embodiment. The sample compression code word can also be referred to as a clean sample, or a real compression code word, or a real sample compression code word, and the like. The possible names of the sample compression code word are not limited or enumerated in this embodiment.

[0129] The first noise sample is obtained based on the sample compression code word and the plurality of second random noise vectors. The first noise sample is obtained based on the sample compression code word, the plurality of second random noise vectors, and a mapping network. Specifically, the sample compression code word and the plurality of second random noise vectors are input into the mapping network to obtain a second target noise vector output by the mapping network, and the second target noise vector is taken as the first noise sample.

[0130] In this embodiment, the mapping network also generates a feature vector close to a real feature based on a set of random noise vectors (i.e., the plurality of second random noise vectors (or the plurality of second random noise variables)). Different from the foregoing embodiments, in the training scenario involved in this embodiment, the real data can be a sample compression code word, i.e., a clean sample; the feature vector close to the real feature can also be referred to as a feature vector closer to the real feature, or a real vector closest to the real feature, or an optimal noise vector. However, in this embodiment, the feature vector closest to the real feature is a second target noise vector, which can be the same as any one of the second random noise vectors, or different from each of the second random noise vectors. That is, in this embodiment, the input information or data of the mapping network includes the plurality of second random noise vectors and the sample compression code word, and the output information or data is a feature vector (i.e., the second target noise vector) closest to the real feature (the feature of the clean sample).

[0131] In some examples, a GD (Gradient Descent) algorithm can be used in the construction or processing of the mapping network. For example, the mapping network can generate a feature vector that is closer to the real feature (the feature of the sample compression code word) by using the gradient descent algorithm and a set of random noise vectors (i.e., a plurality of second random noise vectors (or a plurality of first random noise variables)). In other examples, the mapping network can also use other algorithms or other processing methods. For example, the mapping network can determine a feature vector that is closer to the real feature (the feature of the sample compression code word) from a plurality of second random noise vectors based on a distribution similarity between each second random noise vector and the sample compression code word. The manner in which the distribution similarity between each second random noise vector and the sample compression code word is calculated is not limited in the present embodiment.

[0132] It should be noted that the above is only an exemplary description of the construction (or composition) or processing of the mapping network, and the present embodiment does not limit or exhaust the training, construction, or processing of the mapping network, etc. as long as the mapping network can generate a feature vector that is closest to the real feature according to a plurality of second random noise vectors.

[0133] The first generation model and the first discriminant model are jointly trained based on the first noise sample and the sample compression code word to obtain a trained second generation model and a trained second discriminant model, including: inputting the first noise sample into the first generation model to obtain a pseudo sample compression code word output by the first generation model; obtaining a first similarity parameter between the pseudo sample compression code word and the sample compression code word based on the first discriminant model; calculating a first target loss based on the pseudo sample compression code word, the sample compression code word, 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 second generation model and the trained second discriminant model.

[0134] The first similarity parameter between the pseudo sample compression code word and the sample compression code word based on the first discriminant model can be obtained by inputting the pseudo sample compression code word and the sample compression code word into the first discriminant model to obtain the first similarity parameter between the pseudo sample compression code word and the sample compression code word output by the first discriminant model.

[0135] For example, the first similarity parameter can be an indication value indicating whether the pseudo sample compressed code and the sample compressed code are similar. The first similarity parameter can be equal to a first indication value or a second indication value, the first indication value being different from the second indication value, the first indication value being used to represent that the pseudo sample compressed code and the sample compressed code are different, and the second indication value being used to represent that the pseudo sample compressed code and the sample compressed code 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.

[0136] For example, the first similarity parameter can be a similarity probability value between the pseudo sample compressed code and the sample compressed code. 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, the first value being less than the second value, the first value being used to represent that the pseudo sample compressed code and the sample compressed code are completely different, and the second value being used to represent that the pseudo sample compressed code and the sample compressed code are completely the same. 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 example, if the first value is 0, the second value is 1, and the similarity probability value is 0.8, it can be represented that the probability of the pseudo sample compressed code and the sample compressed code being similar is close to 1, or the pseudo sample compressed code and the sample compressed code are basically similar.

[0137] Optionally, calculating the first target loss based on the pseudo sample compressed code, the sample compressed code, and the first similarity parameter can include: calculating a first generation target loss based on the pseudo sample compressed code, the sample compressed code, and the first similarity parameter; calculating a first discrimination target loss based on the sample compressed code and the first similarity parameter; and obtaining the first target loss based on the first generation target loss and the first discrimination target loss.

[0138] Specifically, calculating the first generation target loss based on the pseudo sample compressed code, the sample compressed code, and the first similarity parameter can include: calculating a first generation sub-loss based on the sample compressed code and the pseudo sample compressed code; calculating a second generation sub-loss based on the first similarity parameter; and calculating the first generation target loss based on the first generation sub-loss and the second generation sub-loss.

[0139] The loss function or calculation method used for calculating the first generation sub-loss based on the sample compressed code and the pseudo sample compressed code can be configured according to actual conditions.

[0140] For example, the first generation sub-loss can be calculated by using an MSE (Mean Squared Error) loss function, which calculates the first generation sub-loss l mse may be represented as wherein N can represent the number of sample compression code words (which can be 1, for example), n is an integer, which can be an integer from 1 to N, y represents a sample compression code word, z*'represents a first noise sample, G represents a first generation model, and G(z* ') represents a pseudo sample compression code word output by the first generation model.

[0141] The loss function used to calculate the second generation sub-loss based on the first similarity parameter can be any binary classification loss function, which is not limited in the embodiment.

[0142] For example, the second generation sub-loss l adv may be represented as wherein D() is used to simply represent a discrimination model, log represents logarithmic calculation (the base number of the logarithmic calculation is not limited in the embodiment), and the meanings of the remaining parameters are the same as those in the foregoing example and will not be repeated here.

[0143] Based on the first generation sub-loss and the second generation sub-loss, the first generation target loss can be calculated, which can be calculated based on the first generation sub-loss, a first weight corresponding to the first generation sub-loss, the second generation sub-loss, and a second weight corresponding to the second generation sub-loss. The first weight and the second weight can be configured according to actual conditions, which are not limited in the embodiment. For example, the first generation target loss l g may be represented as: l g = a * l mse + β * l adv wherein a represents the first weight, β represents the second weight, and the meanings of the remaining parameters are the same as those in the foregoing example and will not be repeated here.

[0144] Based on the sample compression code word and the first similarity parameter, the first discrimination target loss can be calculated, which can be calculated based on the sample compression code word, the first discrimination sub-loss is calculated; based on the first similarity parameter, the second discrimination sub-loss is calculated; and based on the first discrimination sub-loss and the second discrimination sub-loss, the first discrimination target loss is calculated.

[0145] wherein the first discrimination sub-loss can be calculated based on the sample compression code word by using any binary classification loss function, which is not limited in the embodiment. For example, the sample compression code word can be input into a first discrimination model to obtain the similarity between the sample compression code word and itself, and then the loss value corresponding to the similarity is calculated based on a binary classification loss function as the first discrimination sub-loss.

[0146] The second discriminant sub-loss can be calculated based on the first similarity parameter using any binary classification loss function, which is not limited in the embodiment.

[0147] The first discriminant target loss can be calculated based on the first discriminant sub-loss and the second discriminant sub-loss by adding the first discriminant sub-loss and the second discriminant sub-loss to obtain the first discriminant target loss. For example, the first discriminant target loss l d may be expressed as: wherein D(y) can be the similarity between the sample compression code word and itself, and the meanings of the remaining parameters are the same as the aforementioned examples, which are not described herein.

[0148] The first target loss can be obtained based on the first generation target loss and the first discriminant target loss, which can be two parts of the first target loss.

[0149] The trained second generation model and the trained second discriminant model can be obtained by jointly training the first generation model and the first discriminant model based on the first target loss, which can be: training the first generation model based on the first generation target loss, and obtaining the trained second generation model when the training convergence condition of the first generation model is met; training the first discriminant model based on the first discriminant target loss, and obtaining the trained second discriminant model when the training convergence condition of the first discriminant model is met.

[0150] The training of the first generation model based on the first generation target loss can refer to the parameter adjustment of the first generation model based on the first generation target loss. The training of the first discriminant model based on the first discriminant target loss can refer to the parameter adjustment of the first discriminant model based on the first discriminant target loss.

[0151] The training convergence condition (or training end condition) of the first generation model can be that the pseudo sample compression code word generated by the first generation model is approximately distributed with the sample compression code word. For example, the training convergence condition of the first generation model can be that the training number of the first generation model reaches a specified training number, and / or the difference between the first generation target loss and the first target value is less than a first specified difference. The specified training number can be configured according to actual conditions; the first target value can be configured according to actual conditions, such as 0; and the first specified difference can be configured according to actual conditions, such as close to 0 (such as 0.1, or 0.01, etc.).

[0152] The training convergence condition of the first discriminant model can be that the first discriminant model cannot distinguish between the pseudo sample compression code and the sample compression code. For example, the training convergence condition of the first discriminant model can be that the number of training times of the first discriminant model reaches a specified training number, and / or the difference between the first discriminant target loss and a second target value is less than a second specified difference. The specified training number corresponding to the first discriminant model can be the same as the specified training number corresponding to the first generation model, and the specified training number can be configured according to actual conditions; the second target value can be configured according to actual conditions; and the second specified difference can be configured according to actual conditions, such as close to 0 (such as 0.01, or 0.05, etc.).

[0153] It should be pointed out that the joint training of the first generation model and the first discriminant model can be performed in multiple cycles, and in the processing of each training cycle, the first target loss can be calculated in the above manner. Different compression code samples can be input in each training cycle, or the same compression code samples can be input. The embodiment does not limit whether the compression code samples used in each training cycle are the same. For example, a batch of multiple compression code samples can be generated in advance, and any one compression code sample can be used in each training cycle to obtain the corresponding first generation target loss and first discriminant target loss as the first target loss.

[0154] It should also be pointed out that in the process of multiple cycles of training, the first generation model can be trained using the first generation target loss and the first discriminant model can be trained using the first discriminant target loss at each training. Alternatively, the first generation model can be trained using the first generation target loss, the parameters of the first generation model are fixed, the first discriminant model is trained using the first discriminant target loss, the parameters of the first discriminant model are fixed, the first generation model is trained using the first generation target loss, and the process is repeated until the trained second generation model and the trained second discriminant model are obtained. Alternatively, the first discriminant model can be trained using the first discriminant target loss, the parameters of the first discriminant model are fixed, the first generation model is trained using the first generation target loss, the parameters of the first generation model are fixed, the first discriminant model is trained using the first discriminant target loss, and the process is repeated until the trained second generation model and the trained second discriminant model are obtained. The embodiment does not limit the training order of the first generation model and the first discriminant model in the process of multiple cycles of training.

[0155] Optionally, the calculating the first target loss based on the pseudo sample compressed code, the sample compressed code and the first similarity parameter can comprise: calculating a first generation sub-loss based on the sample compressed code and the pseudo sample compressed code; calculating a second generation sub-loss based on the first similarity parameter; and calculating the first target loss based on the first generation sub-loss and the second generation sub-loss.

[0156] The difference between this example and the foregoing examples is that the first discriminant model can not be adjusted in parameters, that is, the first discriminant model can be a trained model by default, that is, the second discriminant model after training is the same as the initial first discriminant model, and only the first generation model needs to be adjusted in parameters.

[0157] The calculating the first generation sub-loss based on the sample compressed code and the pseudo sample compressed code and the calculating the second generation sub-loss based on the first similarity parameter are the same as in the foregoing embodiments, and will not be described herein.

[0158] The calculating the first target loss based on the first generation sub-loss and the second generation sub-loss can be: calculating the first target loss based on the first generation sub-loss, a first weight corresponding to the first generation sub-loss, the second generation sub-loss, and a second weight corresponding to the second generation sub-loss. The first weight and the second weight can be configured according to actual conditions, and are not limited in the present embodiment. For example, the calculating the first target loss l can be represented as: l = a * l mse + b * l adv wherein a represents the first weight, b represents the second weight, and the meanings of the remaining parameters are the same as in the foregoing examples and will not be described herein.

[0159] In this case, the training the first generation model and the first discriminant model based on the first target loss to obtain a second generation model after training and a second discriminant model after training can be: training the first generation model based on the first target loss, and obtaining the second generation model after training and the second discriminant model after training when a training convergence condition of the first generation model is met. The training convergence condition of the first generation model is the same as in the foregoing embodiments and will not be described herein.

[0160] It should be noted that in some possible cases, the target model can be used by both the access network device and the terminal. In this case, if the first device is the access network device, the access network device can send the target model to the terminal after training the target model, and correspondingly, the terminal can receive the target model from the access network device; if the first device is the terminal, the terminal can send the target model to the access network device after training the target model, and correspondingly, the access network device can receive the target model from the terminal.

[0161] In some possible implementation, on the side of the first device, the target model is preconfigured. The first device can be an access network device or a terminal.

[0162] For example, before receiving the first information, the first device can further include: receiving the target model from another electronic device.

[0163] In this embodiment, the target model is a trained second generation model. The process of training the target model can be performed by another electronic device. The another electronic device jointly trains the first generation model and the first discriminant model to obtain a trained second generation model and a trained second discriminant model, and the process is the same as the previous embodiment, and thus is not repeated here.

[0164] The another electronic device can be any electronic device with computing capability other than the first device. For example, the another electronic device can be a server (such as any one of the servers or service nodes in a service cluster), or a core network device (such as any one of the core network devices on the side of the core network), or an electronic device on the user side (such as a personal computer), and the like. Here, all possible types of the another electronic device are not limited or enumerated.

[0165] It should be noted that in some possible cases, the target model can be used by both the access network device and the terminal. In this case, if the first device is the access network device, the access network device can further send the target model to the terminal after receiving the target model from the another electronic device, and correspondingly, the terminal can receive the target model from the access network device; or the another electronic device can further send the trained target model to the terminal, and correspondingly, the terminal can receive the target model from the another electronic device. If the first device is the terminal, the terminal can further send the target model to the access network device after receiving the target model from the another electronic device, and correspondingly, the access network device can receive the target model from the terminal; or the another electronic device can further send the trained target model to the access network device, and correspondingly, the access network device can receive the target model from the another electronic device.

[0166] In the above embodiments, the first generation model and the first discrimination model can be a first generation model (or an initial or original generation model) and a first discrimination model (or an initial or original discrimination model) in a Defense-GAN (Generative Adversarial Networks). Both the Defense-GAN and the APE-GAN (Adversarial Perturbation Elimination with GAN) are based on a GAN (W-GAN) and aim to reconstruct an adversarial sample by using a first generator (or a first generation model) for the purpose of defense. However, the training processes of the APE-GAN and the Defense-GAN are different. In the APE-GAN, clean samples and adversarial samples are respectively input into a first discriminator (a first discrimination model, or simply represented as D) and a first generator (a first generation model, or simply represented as G) for training. In the Defense-GAN, random noise is used as input during training.

[0167] Next, the training process of the first generator G and the first discriminator D of the Defense-GAN will be described in an exemplary manner with reference to FIG. 8. First, a set of random noise vectors (which can include R second random noise vectors z, represented as z(1), z(2)~z(R) in FIG. 8) is generated by using a random noise generator. The set of random noise vectors and clean samples (real data or sample compressed codewords) are input into a mapping network to generate feature vectors (optimal noise vectors, i.e., first noise samples) that are closer to real features (i.e., features of clean samples). Then, the feature vectors are input into the Defense-GAN (including the first generator G and the first discriminator D) for training. During the training process, the first generator G converts the input feature vectors into pseudo sample data (i.e., pseudo sample compressed codewords) so that the pseudo sample data is as similar as possible to the clean samples (i.e., real data or sample compressed codewords). Then, the first discriminator D identifies the real data and the pseudo sample data generated by the first generator G (i.e., obtains first similarity parameters) to determine whether the pseudo sample data is consistent with the clean samples (i.e., real data or sample compressed codewords). The above training aims to make the pseudo sample data infinitely close to the clean samples.

[0168] Next, another exemplary description of the training process of the first generator and the first discriminator will be given with reference to FIG. 9.

[0169] S901: Channel information is collected, and based on the collected channel information, a training set, a test set, and an adversarial set for generating adversarial samples are obtained.

[0170] For example, based on the collected channel information, the training set, the test set and the adversarial set for generating adversarial samples can be obtained by preprocessing the collected channel information to obtain the training set, the test set and the adversarial set for generating adversarial samples.

[0171] The way of collecting the channel information is not limited in the embodiment. For example, the channel information can be downlink channel information, and the collected downlink channel information can include CSI original information, reconstructed information, compressed information, etc. It should be understood that this is only an example, and the channel information can also be uplink channel information, and the embodiment does not limit or exhaust all possible cases.

[0172] The training set and the test set are used to train the deep neural network to obtain the autoencoder model, and the embodiment is not limited. The adversarial set for generating adversarial samples can at least include compressed information (or sample compressed information, or sample compressed codeword).

[0173] S902: training the first generator G using the adversarial set to obtain adversarial sample codewords (i.e. pseudo sample compressed codewords) similar to the original codeword (sample compressed codeword) distribution.

[0174] S903: inputting the adversarial sample codeword into the first discriminator D to obtain a first similarity parameter between the adversarial sample codeword and the original codeword (sample compressed codeword). Here, the input of the first discriminator D can include the adversarial sample codeword (i.e. G'(z * )) and the original codeword, and the output is the first similarity parameter between the two. The related description of the first similarity parameter is the same as the previous embodiment, and is not repeated here. Further, the termination condition of step S903 is as follows: the first discriminator D cannot distinguish between the adversarial sample codeword and the original codeword (sample compressed codeword).

[0175] The training of the first generator G and the first discriminator D is performed in a loop of S902-S903. The first generator G continuously optimizes its network during the training process to make the generated adversarial sample codeword indistinguishable from the original codeword by the first discriminator D, and the first discriminator D also continuously optimizes its network during the training process to make its judgment more accurate.

[0176] In some possible implementation manners, the processing of the second device can further include: measuring a reference signal from the first device to obtain channel information between the first device and the second device; and encoding the channel information to obtain the second compressed codeword. Correspondingly, the processing of the first device can include: transmitting the reference signal.

[0177] In this implementation, after the second device obtains the second compressed code word, the second device can perform dimension reduction processing on the second compressed code word to obtain a third compressed code word, and send the third compressed code word. The dimension reduction processing on the second compressed code word to obtain the third compressed code word includes: generating a plurality of third random noise vectors, and performing dimension reduction processing on the second compressed code word based on the plurality of third random noise vectors to obtain the third compressed code word.

[0178] After the first device sends the reference signal, the first device can receive the first information, and then obtain, based on the first information and a target model, a first compressed code word after the first information is de-disturbed. The obtaining, based on the first information and the target model, of the first compressed code word after the first information is de-disturbed includes: inputting the first information into the target model to obtain the first compressed code word after the first information is de-disturbed, which is output by the target model.

[0179] The processing performed by the first device after obtaining the first compressed code word can include: decoding the first compressed code word to obtain reconstructed channel information between the first device and the second device.

[0180] By using the above scheme, the second device performs dimension reduction on the real second compressed code word to obtain a third compressed code word. As a result, the dimension of the compressed code word sent by the second device is reduced, so that the amount of data transmitted between the second device and the first device is reduced, and the data transmission efficiency is ensured. Moreover, the third compressed code word obtained by the second device after the dimension reduction of the real second compressed code word plays a role of encryption of the real second compressed code word, so that the real compressed code word is not transmitted over the air, and the information security is ensured. Correspondingly, after receiving the first information, the first device obtains the first compressed code word after the first information is de-disturbed through the target model. In this way, in the case where the transmission of the compressed code word between the second device and the first device is subjected to an adversarial disturbance attack, the first device can also obtain the first compressed code word after de-disturbance through the target model. As a result, the accuracy of the first compressed code word obtained by the first device is ensured, and thus the accuracy of the channel information reconstructed by the first device based on the first compressed code word is ensured, and the quality of wireless communication is ensured.

[0181] In some embodiments, in a 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.

[0182] The processing performed by the terminal after obtaining the second compressed code word can include: performing dimension reduction processing on the second compressed code word to obtain a third compressed code word, and sending the third compressed code word.

[0183] The dimension reduction processing on the second compression code word to obtain a third compression code word comprises: generating a plurality of third random noise vectors; and performing dimension reduction processing on the second compression code word based on the plurality of third random noise vectors to obtain the third compression code word. The plurality of third random noise vectors can be generated by using a random noise generator, and the number of third random noise vectors is greater than or equal to 2, which is within the protection scope of the embodiment, and the embodiment does not limit the number of third random noise vectors.

[0184] The dimension reduction processing on the second compression code word based on the plurality of third random noise vectors to obtain the third compression code word can comprise: obtaining a third target noise vector based on the second compression code word, the plurality of third random noise vectors and a mapping network, and taking the third target noise vector as the third compression code word obtained by the dimension reduction processing on the second compression code word.

[0185] The function of the mapping network is to generate a feature vector close to a real feature according to a set of random noise vectors (that is, a plurality of third random noise vectors (or a plurality of third random noise variables)). Here, the real feature is the feature of the real data, and in the embodiment, the real data can be the second compression code word generated by the terminal; the feature vector close to the real feature can also be referred to as a feature vector closer to the real feature, or a real vector closest to the real feature. In the embodiment, the feature vector closest to the real feature is the third target noise vector, which can be the same as any one of the third random noise vectors or different from each of the third random noise vectors. That is, in the embodiment, the input information or input data of the mapping network includes the plurality of third random noise vectors and the real second compression code word, and the output information or output data is the feature vector (third target noise vector) closest to the real feature.

[0186] In some possible examples, the GD (Gradient Descent) algorithm can be used in the composition or construction or processing of the mapping network. For example, the mapping network can generate a feature vector closer to the real feature (the feature of the real second compression code word) by using the gradient descent algorithm and a set of random noise vectors (that is, a plurality of third random noise vectors (or a plurality of third random noise variables)). In another possible example, the mapping network can also use other algorithms or other processing modes. For example, the mapping network can determine a feature vector (that is, a third target noise vector) closer to the real feature (the feature of the real second compression code word) from the plurality of third random noise vectors based on the distribution similarity between each third random noise vector and the real second compression code word. The embodiment does not limit the calculation method of the distribution similarity between each third random noise vector and the real second compression code word.

[0187] It should be noted that the above is only an exemplary description of the construction (or composition) or processing manner of the mapping network, and the present embodiment does not limit or exhaust the training, composition or construction, processing manner, etc. of the mapping network, as long as the mapping network can generate a feature vector closest to the real feature according to a plurality of third random noise vectors.

[0188] The terminal sending the third compressed code word can include that the terminal transmits the third compressed code word to the access network device through the air interface. Wherein, the terminal can use any type of uplink AS message or information bearer to transmit the third compressed code word, and the present embodiment does not limit.

[0189] Correspondingly, the access network device can receive the first information. In the present embodiment, the terminal transmits the third compressed code word through the air interface, and the access network device receives the first information through the air interface, because the third compressed code word may be subjected to an adversarial attack in the process of transmission between the terminal and the access network device, that is, the access network device side may receive information added with disturbance or interference to the third compressed code word, therefore, the information obtained by adding disturbance or interference to the third compressed code word received by the access network device through the air interface is called the first information.

[0190] Further, after the access network device receives the first information, the de-disturbed compressed code word is obtained based on the target model and the first information. Although, in theory, the de-disturbed compressed code word obtained by the access network device should be the same as or infinitely close to or similar to the compressed code word generated by the terminal, in order to distinguish and describe the de-disturbed compressed code word obtained by the access network device and the compressed code word generated by the terminal, the present embodiment calls the de-disturbed compressed code word obtained by the access network device as the first compressed code word, and calls the compressed code word generated by the terminal as the second compressed code word, which will not be repeated hereinafter.

[0191] On the access network device side, based on the first information and the target model, the first compressed code word after de-disturbing the first information is obtained, including inputting the first information into the target model to obtain the first compressed code word after de-disturbing the first information output by the target model.

[0192] The target model is a neural network with generation capability, and in the present embodiment, the function of the target model is to generate data or code word similar or close to the distribution of the real compressed code word (i.e. the second compressed code word). The composition structure of the target model is similar to the foregoing embodiments, which will not be repeated here.

[0193] The processing after the access network device obtains the first compressed code word can further include: decoding the first compressed code word to obtain reconstructed downlink channel information between the access network device and the terminal. Specifically, the access network device decodes the first compressed code word to obtain reconstructed downlink channel information between the access network device and the terminal can be: inputting the first compressed code word into a decoder to obtain downlink channel information between the access network device and the terminal reconstructed by the decoder. In this embodiment, the decoder on the access network device side and the encoder on the terminal side are described as before, and will not be repeated.

[0194] In combination with FIG. 10 and FIG. 11, the communication method provided by the present embodiment is further exemplarily described, including:

[0195] S1000-S1002 are the same as S400-S402 shown in FIG. 4, and will not be repeated.

[0196] S1003: The UE finds an optimal random noise vector (i.e., a third compressed code word after dimensionality reduction processing on the second compressed code word) based on the second compressed code word and the second target model.

[0197] For example, in combination with FIG. 11, S1003 specifically includes: inputting the original CSI into the encoder to obtain the second compressed code word compressed by the encoder; and then generating R third random noise vectors {z1, z2, …, zR} by using the random noise generator, and generating the optimal random noise vector z* based on the second compressed code word and the R third random noise vectors by using the mapping network (i.e., the feature vector closest to the real feature). Further, the UE can also take the optimal random noise vector as the third compressed code word after dimensionality reduction processing on the second compressed code word. R * For example, in combination with FIG. 11, S1003 specifically includes: inputting the original CSI into the encoder to obtain the second compressed code word compressed by the encoder; and then generating R third random noise vectors {z1, z2, …, zR} by using the random noise generator, and generating the optimal random noise vector z* based on the second compressed code word and the R third random noise vectors by using the mapping network (i.e., the feature vector closest to the real feature). Further, the UE can also take the optimal random noise vector as the third compressed code word after dimensionality reduction processing on the second compressed code word.

[0198] S1004: The UE sends the optimal random noise vector (i.e., the third compressed code word after dimensionality reduction processing on the second compressed code word) to the BS through the wireless channel (air interface). For example, in combination with FIG. 11, the transmitter on the UE side sends the optimal random noise vector (i.e., the third compressed code word) to the BS through the wireless channel (air interface).

[0199] S1005: The BS receives the wireless signal sent by the UE to obtain the received first information; and the BS performs the de-disturb processing on the first information by using the target model deployed on the BS side to obtain the clean intermediate code word (i.e., the first compressed code word).

[0200] S1006: The BS inputs the clean intermediate code word into the decoder to obtain the recovered CSI.

[0201] ​For example, in combination with FIG. 11, S1005 and S1006 can be that the receiver of the BS inputs the received first information (which can be the optimal random noise vector after adding disturbance, that is, the third compressed code word) into the target model to obtain the clean intermediate code word output by the target model; and inputs the clean intermediate code word into the decoder to obtain the recovered CSI.

[0202] By adopting the above scheme, the terminal reduces the dimension of the real second compressed code word to obtain the third compressed code word. As the dimension of the compressed code word sent by the terminal is low, the amount of data transmitted between the terminal and the access network device can be reduced, the transmission efficiency of the terminal is ensured, and the third compressed code word obtained by the terminal after reducing the dimension of the real second compressed code word plays a role of encrypting the real second compressed code word, so that the real compressed code word can be avoided to be transmitted in the air interface, thereby ensuring the information security. Correspondingly, the access network device side obtains the compressed code word after removing the disturbance of the first information through the target model, so that in the case that the transmission of the compressed code word between the terminal and the access network device is attacked by the adversary, the access network device can also obtain the compressed code word after removing the disturbance, the accuracy of the compressed code word obtained by the access network device side is ensured, and the accuracy of the downlink channel information reconstructed by the access network device based on the compressed code word is ensured, thereby ensuring the quality of wireless communication.

[0203] 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 is uplink channel information.

[0204] The processing of the access network device includes: measuring the uplink reference signal from the terminal to obtain the uplink channel information between the terminal and the access network device; and encoding the uplink channel information to obtain the second compressed code word.

[0205] The processing of the access network device after obtaining the second compressed code word can include: reducing the dimension of the second compressed code word to obtain a third compressed code word; and sending the third compressed code word.

[0206] The access network device reduces the dimension of the second compressed code word to obtain the third compressed code word, which includes: generating a plurality of third random noise vectors; and reducing the dimension of the second compressed code word based on the plurality of third random noise vectors to obtain the third compressed code word. The plurality of third random noise vectors can be generated by a random noise generator, and the number of third random noise vectors is greater than or equal to 2, which is within the protection scope of the present embodiment, and the present embodiment does not limit it.

[0207] The dimension reduction processing of the second compressed code word based on the plurality of third random noise vectors by the access network device to obtain the third compressed code word can comprise: obtaining a third target noise vector based on the second compressed code word, the plurality of third random noise vectors and a mapping network, and taking the third target noise vector as the third compressed code word obtained by the dimension reduction processing of the second compressed code word. Here, the related description about the mapping network and its processing is similar to the foregoing embodiments, and the only difference is that the mapping network is arranged at the access network device side in the present embodiment, while the mapping network is arranged at the terminal side in the foregoing embodiments, and thus the repeated description is not given.

[0208] The access network device transmits the third compressed code word, which can comprise that the terminal transmits the third compressed code word to the access network device through the air interface. The access network device can use any type of downlink AS message or information bearer to transmit the third compressed code word, and the present embodiment is not limited in this regard.

[0209] Correspondingly, the access network device can receive the first information. In the present embodiment, the terminal transmits the third compressed code word through the air interface, while the access network device receives the first information through the air interface, because the third compressed code word can be subjected to a counter-attack in the process of transmission between the terminal and the access network device, that is, the information received by the access network device side can be information added with disturbance or interference to the third compressed code word, and thus the information obtained by adding disturbance or interference to the third compressed code word received by the access network device through the air interface is called the first information.

[0210] Further, after receiving the first information, the terminal obtains the de-disturbed compressed code word based on the target model and the first information. Although, in theory, the de-disturbed compressed code word obtained by the terminal should be the same as or infinitely close to or similar to the compressed code word generated by the access network device, in order to distinguish and describe the de-disturbed compressed code word obtained by the terminal and the compressed code word generated by the access network device, the de-disturbed compressed code word obtained by the terminal is called the first compressed code word, and the compressed code word generated by the access network device is called the second compressed code word, and the repeated explanation is not given below.

[0211] At the terminal side, the first compressed code word obtained by de-disturbing the first information based on the first information and the target model comprises: inputting the first information into the target model to obtain the first compressed code word de-disturbed by the first information output by the target model.

[0212] The target model is a neural network with generation capability, and in the present embodiment, the function of the target model is to generate data or code words similar or close to the distribution of the real compressed code word (i.e. the second compressed code word). The composition structure of the target model is similar to the foregoing embodiments, and thus the repeated description is not given.

[0213] The processing after the terminal obtains the first compressed code word can further include: decoding the first compressed code word to obtain reconstructed uplink channel information between the access network device and the terminal. Specifically, the terminal inputs the first compressed code word into a decoder for decoding to obtain the uplink channel information between the access network device and the terminal reconstructed by the decoder. In this embodiment, the decoder of the terminal and the encoder on the side of the access network device are described as in the foregoing embodiments, and thus no further description is given.

[0214] By using the above scheme, the access network device obtains the third compressed code word by reducing the dimension of the real second compressed code word. Thus, since the compressed code word transmitted by the access network device has a low dimension, the amount of data transmitted between the terminal and the access network device can be reduced, and the data transmission efficiency is ensured. Moreover, the third compressed code word obtained by the access network device by reducing the dimension of the real second compressed code word plays a role of encrypting the real second compressed code word, and can avoid transmitting the real compressed code word over the air, thereby ensuring information security. Correspondingly, the terminal obtains the compressed code word after the first information is de-disturbed by using the target model. Thus, in the case that the terminal and the access network device are attacked by an adversary in the transmission of the compressed code word, the terminal can also obtain the de-disturbed compressed code word, ensuring the accuracy of the compressed code word obtained by the terminal, and further ensuring the accuracy of the uplink channel information reconstructed by the terminal based on the compressed code word, thereby ensuring the quality of wireless communication.

[0215] In some embodiments, the target model used by the first device can be locally trained, and the training manner can be the same as that in the foregoing embodiments, which will not be repeated.

[0216] In some embodiments, the target model used by the first device can be locally trained, and the training manner can be different from that in the foregoing embodiments. In this embodiment, the first device can also be the access network device or the terminal.

[0217] Specifically, the manner in which the first device trains the target model can include: generating a plurality of second random noise vectors; obtaining first noise samples based on a sample compressed code word and the plurality of second random noise vectors; jointly training a first generation model and a first discriminator model based on the first noise samples and the sample compressed code word to obtain a trained second generation model and a trained second discriminator model; and taking the trained second generation model as the target model.

[0218] The manner in which the first noise samples are obtained based on the sample compressed code word and the plurality of second random noise vectors is the same as that in the foregoing embodiments, and thus no further description is given.

[0219] The first noise sample is added with a disturbance to obtain a second noise sample; the second noise sample is input into the first generation model to obtain a pseudo sample compression codeword output by the first generation model; a second similarity parameter between the pseudo sample compression codeword and the sample compression codeword is obtained based on the first discrimination model; a second target loss is calculated based on the pseudo sample compression codeword, the sample compression codeword and the second similarity parameter; the first generation model and the first discrimination model are jointly trained based on the second target loss to obtain the trained second generation model and the trained second discrimination model.

[0220] The first noise sample is added with a disturbance to obtain a second noise sample; the second noise sample is input into the first generation model to obtain a pseudo sample compression codeword output by the first generation model; a second similarity parameter between the pseudo sample compression codeword and the sample compression codeword is obtained based on the first discrimination model; a second target loss is calculated based on the pseudo sample compression codeword, the sample compression codeword and the second similarity parameter; the first generation model and the first discrimination model are jointly trained based on the second target loss to obtain the trained second generation model and the trained second discrimination model.

[0221] The preset condition can include at least one of the following: the power of the disturbance signal is less than a specified power value; the content obtained by decoding the first noise sample to which the disturbance signal is added is different from channel information. Here, the channel information can be uplink channel information, that is, the uplink channel information corresponding to the sample compression codeword; or the channel information can be downlink channel information, that is, the downlink channel information corresponding to the sample compression codeword.

[0222] The specific generation method of the disturbance signal can be configured according to actual conditions, and the embodiment does not limit it.

[0223] The first noise sample is added with a disturbance to obtain a second noise sample; the second noise sample is input into the first generation model to obtain a pseudo sample compression codeword output by the first generation model; a second similarity parameter between the pseudo sample compression codeword and the sample compression codeword is obtained based on the first discrimination model; a second target loss is calculated based on the pseudo sample compression codeword, the sample compression codeword and the second similarity parameter; the first generation model and the first discrimination model are jointly trained based on the second target loss to obtain the trained second generation model and the trained second discrimination model.

[0224] Optionally, the second target loss can be calculated based on the pseudo sample compression codeword, the sample compression codeword and the second similarity parameter, which can include: a second generation target loss is calculated based on the pseudo sample compression codeword, the sample compression codeword and the second similarity parameter; a second discrimination target loss is calculated based on the sample compression codeword and the second similarity parameter; and the second target loss is obtained based on the second generation target loss and the second discrimination target loss.

[0225] Specifically, the calculating the second generation target loss based on the pseudo sample compressed code, the sample compressed code and the second similarity parameter can include: calculating a third generation sub-loss based on the sample compressed code and the pseudo sample compressed code; calculating a fourth generation sub-loss based on the second similarity parameter; and calculating the second generation target loss based on the third generation sub-loss and the fourth generation sub-loss.

[0226] The loss function or calculation method used for calculating the third generation sub-loss based on the sample compressed code and the pseudo sample compressed code can be configured according to actual conditions. For example, the second generation sub-loss can be calculated by using an MSE (Mean Squared Error) loss function, which is not limited herein.

[0227] The loss function used for calculating the fourth generation sub-loss based on the second similarity parameter can be any binary classification loss function, which is not limited herein.

[0228] The calculating the second generation target loss based on the third generation sub-loss and the fourth generation sub-loss can be: calculating the second generation target loss based on the third generation sub-loss, a third weight corresponding to the third generation sub-loss, the fourth generation sub-loss and a fourth weight corresponding to the fourth generation sub-loss. The third weight and the fourth weight can be configured according to actual conditions, which are not limited herein.

[0229] The calculating the second discrimination target loss based on the sample compressed code and the second similarity parameter can be: calculating a third discrimination sub-loss based on the sample compressed code; calculating a fourth discrimination sub-loss based on the second similarity parameter; and calculating the second discrimination target loss based on the third discrimination sub-loss and the fourth discrimination sub-loss.

[0230] The calculating the third discrimination sub-loss based on the sample compressed code can be calculated by using any binary classification loss function, which is not limited herein. The calculating the fourth discrimination sub-loss based on the second similarity parameter can be calculated by using any binary classification loss function, which is not limited herein.

[0231] The calculating the second discrimination target loss based on the third discrimination sub-loss and the fourth discrimination sub-loss can be adding the third discrimination sub-loss and the fourth discrimination sub-loss to obtain the second discrimination target loss.

[0232] Exemplarily, the obtaining the second target loss based on the second generation target loss and the second discrimination target loss can be taking the second generation target loss and the second discrimination target loss as two parts of the second target loss.

[0233] The first generation model and the first discriminant model are jointly trained based on the second target loss to obtain a trained second generation model and a trained second discriminant model. The first generation model can be trained using the second generation target loss, and the trained second generation model is obtained when a training convergence condition of the first generation model is met. The first discriminant model can be trained using the second discriminant target loss, and the trained second discriminant model is obtained when a training convergence condition of the first discriminant model is met.

[0234] The training convergence condition (or the training end condition) of the first generation model can be that the pseudo sample compressed code generated by the first generation model is approximately distributed as the sample compressed code. For example, the training convergence condition of the first generation model can be that the number of training times of the first generation model reaches a specified number of training times, and / or the difference between the second generation target loss and a fourth target value is less than a fourth specified difference. The specified number of training times can be configured according to actual conditions. The fourth target value can be configured according to actual conditions, for example, it can be 0. The fourth specified difference can be configured according to actual conditions, for example, it can be close to 0 (such as 0.1, or 0.01, etc.).

[0235] The training convergence condition of the first discriminant model can be that the first discriminant model cannot distinguish between the pseudo sample compressed code and the sample compressed code. For example, the training convergence condition of the first discriminant model can be that the number of training times of the first discriminant model reaches a specified number of training times, and / or the difference between the second discriminant target loss and a fifth target value is less than a fifth specified difference. The specified number of training times corresponding to the first discriminant model can be the same as the specified number of training times corresponding to the first generation model, and the specified number of training times can be configured according to actual conditions. The fifth target value can be configured according to actual conditions. The fifth specified difference can be configured according to actual conditions, for example, it can be close to 0 (such as 0.01, or 0.05, etc.).

[0236] For example, based on the second generation target loss and the second discriminant target loss, the second target loss can be calculated using a preset calculation method based on the second generation target loss and the second discriminant target loss. The preset calculation method can be configured according to actual conditions, for example, it can be weighted summation, or direct summation, or other calculation methods, which are not limited in the present embodiment.

[0237] The first generation model and the first discriminant model are jointly trained based on the second target loss to obtain a trained second generation model and a trained second discriminant model. The first generation model and the first discriminant model can be trained using the second target loss, and the trained second generation model and the trained second discriminant model are obtained when a model training convergence condition is met.

[0238] The training of the first generation model and the first discrimination model based on the second target loss can mean that the first generation model and the first discrimination model are respectively adjusted in parameters based on the second target loss. In this case, the model training convergence condition can be that the pseudo sample compressed code generated by the first generation model is approximately distributed with the sample compressed code, and the first discrimination model cannot distinguish the pseudo sample compressed code from the sample compressed code. For example, the model training convergence condition can include that the number of joint training reaches a specified training number, and / or the difference between the second target loss and a sixth target value is less than a sixth specified difference. The specified training number can be configured according to actual conditions; the sixth target value can be configured according to actual conditions; and the sixth specified difference can be configured according to actual conditions, such as close to 0 (such as 0.1, or 0.01, etc.).

[0239] For example, the calculation of the second target loss based on the pseudo sample compressed code, the sample compressed code and the second similarity parameter can include: calculating a third generation sub-loss based on the sample compressed code and the pseudo sample compressed code; calculating a fourth generation sub-loss based on the second similarity parameter; calculating the second generation target loss based on the third generation sub-loss and the fourth generation sub-loss; and calculating the second target loss based on the third generation sub-loss and the fourth generation sub-loss.

[0240] The difference between this example and the foregoing examples is that the first discrimination model can not be adjusted in parameters, that is, the first discrimination model can be a trained model by default, that is, the first discrimination model can be the same as the trained second discrimination model, and only the first generation model needs to be loss calculated and adjusted in parameters.

[0241] The calculation of the third generation sub-loss based on the sample compressed code and the pseudo sample compressed code, and the calculation of the fourth generation sub-loss based on the second similarity parameter are the same as in the foregoing embodiments, and will not be repeated.

[0242] The calculation of the second target loss based on the third generation sub-loss and the fourth generation sub-loss can be: calculating the second target loss based on the third generation sub-loss, a third weight corresponding to the third generation sub-loss, the fourth generation sub-loss, and a fourth weight corresponding to the fourth generation sub-loss. The meanings of the third weight and the fourth weight are the same as in the foregoing examples, and will not be repeated.

[0243] In this case, the first generation model and the first discriminant model are jointly trained based on the second target loss to obtain a trained second generation model and a trained second discriminant model. The first generation model can be trained using the second target loss, and the trained second generation model and the trained second discriminant model are obtained when the training convergence condition of the first generation model is met. The training convergence condition of the first generation model is the same as that of the foregoing embodiments, and will not be described again.

[0244] It should be noted that in some possible cases, the target model can be used by the access network device and the terminal. In this case, if the first device is the access network device, the access network device can send the target model to the terminal after training the target model, and correspondingly, the terminal can receive the target model from the access network device. If the first device is the terminal, the terminal can send the target model to the access network device after training the target model, and correspondingly, the access network device can receive the target model from the terminal.

[0245] In some embodiments, the first device side target model is preconfigured. For example, the other electronic device can send the trained second generation model to the first device as the target model, and correspondingly, the first device can receive the target model from the other electronic device. The other electronic device is similar to the foregoing embodiments, and will not be described again. The way of obtaining the target model on the other electronic device side is the same as the way of training the target model in the foregoing embodiments, and will not be described again.

[0246] It should be noted that in some possible cases, the target model can be used by the access network device and the terminal. In this case, if the first device is the access network device, the access network device can send the target model to the terminal after receiving the target model from the other electronic device, and correspondingly, the terminal can receive the target model from the access network device. Alternatively, the other electronic device can also send the trained target model to the terminal, and correspondingly, the terminal can receive the target model from the other electronic device. If the first device is the terminal, the terminal can send the target model to the access network device after receiving the target model from the other electronic device, and correspondingly, the access network device can receive the target model from the terminal. Alternatively, the other electronic device can also send the trained target model to the access network device, and correspondingly, the access network device can receive the target model from the other electronic device.

[0247] FIG. 12 is a schematic diagram of the composition structure of the first device according to an embodiment of the present application, which includes:

[0248] The first communication unit 1201 is configured to receive the first information.

[0249] The first processing unit 1202 is configured to obtain, based on the first information and a target model, a first compressed code word after de-disturbance of the first information.

[0250] The first processing unit is configured to decode the first compressed code word to obtain reconstructed channel information between the first device and the second device.

[0251] The first device is an access network device, and the second device is a terminal, and the channel information is downlink channel information; or the first device is a terminal, and the second device is an access network device, and the channel information is uplink channel information.

[0252] The first processing unit is configured to generate a plurality of first random noise vectors, obtain a first target noise vector based on the first information and the plurality of first random noise vectors, and input the first target noise vector into the target model to obtain the first compressed code word after de-disturbance of the first information output by the target model.

[0253] The first processing unit is configured to input the first information into the target model to obtain the first compressed code word after de-disturbance of the first information output by the target model.

[0254] The first processing unit is configured to generate a plurality of second random noise vectors, obtain a first noise sample based on a sample compressed code word and the plurality of second random noise vectors, and jointly train a first generation model and a first discriminant model based on the first noise sample and the sample compressed code word to obtain a trained second generation model and a trained second discriminant model, and use the trained second generation model as the target model.

[0255] The first processing unit is configured to input the first noise sample into the first generation model to obtain a pseudo sample compressed code word output by the first generation model, obtain a first similarity parameter between the pseudo sample compressed code word and the sample compressed code word based on the first discriminant model, calculate a first target loss based on the pseudo sample compressed code word, the sample compressed code word, 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 second generation model and the trained second discriminant model.

[0256] The first processing unit is configured to add a disturbance to the first noise sample to obtain a second noise sample, input the second noise sample into the first generation model to obtain a pseudo sample compressed code word output by the first generation model, obtain a second similarity parameter between the pseudo sample compressed code word and the sample compressed code word based on the first discrimination model, calculate a second target loss based on the pseudo sample compressed code word, the sample compressed code word and the second similarity parameter, and jointly train the first generation model and the first discrimination model based on the second target loss to obtain the trained second generation model and the trained second discrimination model.

[0257] The target model is preconfigured.

[0258] FIG. 13 is a schematic diagram of the composition structure of a second device according to an embodiment of the present application, including:

[0259] The second processing unit is configured to perform dimensionality reduction processing on the second compressed code word to obtain a third compressed code word.

[0260] The second communication unit is configured to send the third compressed code word.

[0261] The second processing unit is configured to measure a reference signal from a first device to obtain channel information between the second device and the first device, and encode the channel information to obtain the second compressed code word.

[0262] The first device is an access network device, and the second device is a terminal, and the channel information is downlink channel information; or the first device is a terminal, and the second device is an access network device, and the channel information is uplink channel information.

[0263] The second processing unit is configured to generate a plurality of third random noise vectors, and perform dimensionality reduction processing on the second compressed code word based on the plurality of third random noise vectors to obtain the third compressed code word.

[0264] The device of the embodiments of the present application can realize the corresponding functions of each device in the communication method embodiments described above. The processes, functions, implementation manners and advantages of each module (sub-module, unit or component, etc.) in the device correspond to the descriptions of the corresponding modules in the method embodiments described above, and will not be described here. It should be noted that the functions described with respect to each module (sub-module, unit or component, etc.) in the device of the embodiments of the present application can be realized by different modules (sub-modules, units or components, etc.), or can be realized by the same module (sub-module, unit or component, etc.).

[0265] It should be understood that the magnitude of the serial number of each process in various embodiments of the present application does not mean the order of execution, the execution order of each process should be determined by its function and inherent logic. Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here. The above is only a specific embodiment of the present application, and the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A communication method performed by a first device, comprising: Receive the first message; Based on the first information and the target model, the first compressed codeword after de-perturbation of the first information is obtained.

2. The method according to claim 1, wherein, The method further includes: Decoding the first compressed codeword yields the reconstructed channel information between the first and second devices.

3. The method according to 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; Alternatively, the first device may be a terminal, the second device may be an access network device, and the channel information may be uplink channel information.

4. The method according to any one of claims 1-3, wherein, The step of obtaining the first compressed codeword after de-perturbation of the first information based on the first information and the target model includes: Generate multiple first random noise vectors; Based on the first information and the plurality of first random noise vectors, a first target noise vector is obtained; The first target noise vector is input into the target model to obtain the first compressed codeword output by the target model after the first information has been de-perturbed.

5. The method according to any one of claims 1-3, wherein, The step of obtaining the first compressed codeword after de-perturbation of the first information based on the first information and the target model includes: The first information is input into the target model to obtain the first compressed codeword output by the target model after the first information has been de-perturbed.

6. The method according to any one of claims 1-5, further comprising: Generate multiple second random noise vectors; Based on the sample compressed codewords and the multiple second random noise vectors, the first noise sample is obtained; Based on the first noise sample and the sample compressed codeword, the first generation model and the first discriminant model are jointly trained to obtain the trained second generation model and the trained second discriminant model; The trained second generative model is used as the target model.

7. The method according to claim 6, wherein, The step of jointly training the first generation model and the first discriminant model based on the first noise sample and the sample compressed codeword to obtain the trained second generation model and the trained second discriminant model includes: The first noise sample is input into the first generation model to obtain the pseudo sample compressed codewords output by the first generation model; Based on the first discrimination model, the first similarity parameter between the pseudo-sample compressed codeword and the sample compressed codeword is obtained; The first target loss is calculated based on the pseudo-sample compressed codeword, the sample compressed codeword, and the first similarity parameter. The first generative model and the first discriminative model are jointly trained based on the first target loss to obtain the trained second generative model and the trained second discriminative model.

8. The method according to claim 6, wherein, The step of jointly training the first generation model and the first discriminant model based on the first noise sample and the sample compressed codeword to obtain the trained second generation model and the trained second discriminant model includes: A perturbation is added to the first noise sample to obtain a second noise sample; The second noise sample is input into the first generation model to obtain the pseudo sample compressed codewords output by the first generation model; Based on the first discrimination model, a second similarity parameter between the pseudo-sample compressed codeword and the sample compressed codeword is obtained; The second target loss is calculated based on the pseudo-sample compressed codeword, the sample compressed codeword, and the second similarity parameter. The first generative model and the first discriminative model are jointly trained based on the second target loss to obtain the trained second generative model and the trained second discriminative model.

9. The method according to any one of claims 1-5, wherein, The target model is pre-configured.

10. A communication method performed by a second device, comprising: The second compressed codeword is dimensionality reduced to obtain the third compressed codeword; Send the third compressed codeword.

11. The method according to claim 10, wherein, The method further includes: By measuring the reference signal from the first device, channel information between the second device and the first device is obtained; The channel information is encoded to obtain the second compressed codeword.

12. The method according to claim 11, wherein, The first device is an access network device, the second device is a terminal, and the channel information is downlink channel information; Alternatively, the first device may be a terminal, the second device may be an access network device, and the channel information may be uplink channel information.

13. The method according to any one of claims 10-12, wherein, The dimensionality reduction process of the second compressed codeword to obtain the third compressed codeword includes: Generate multiple third random noise vectors; The second compressed codeword is obtained by dimensionality reduction based on the multiple third random noise vectors.

14. A first device, comprising: The first communication unit is used to receive the first information; The first processing unit is used to obtain a first compressed codeword after de-perturbation of the first information based on the first information and the target model.

15. A second device, comprising: The second processing unit is used to perform dimensionality reduction processing on the second compressed codeword to obtain the third compressed codeword; The second communication unit is used to send the third compressed codeword.

Images