Method, device, electronic device and medium for channel transmission
By generating and superimposing perturbation signals at the DECS transmitter, the security risks of DECS when facing attacks on specific semantic content are resolved, thereby improving the robustness and security of the system.
Patent Information
- Application Number
- CN202210700121.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-20
- Publication Date
- 2025-10-24
- Estimated Expiration
- 2042-06-20
AI Technical Summary
Existing end-to-end communication systems (DECS) constructed with deep neural networks are vulnerable to adversarial attacks and cannot effectively defend against attacks targeting specific semantic content, posing security risks.
A disturbance generator is deployed at the transmitting end of DECS to generate a disturbance signal that covers the target semantic information of the original signal, and superimposes it with the original signal to form a disturbance transmission signal, which is transmitted to the receiving end through the channel.
This allows for covert attacks on the transmission signals of specific semantic information without affecting the transmission of other semantic information, improving the robustness of DECS, enriching the training sample space, and enhancing the model's security and defense capabilities.
Smart Images
Figure CN116032518B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of signal processing, and in particular to a channel transmission method and device, an electronic device, and a medium. BACKGROUND
[0002] In the related art, a DNN-based End-to-End Communication System (DECS) constructed using a deep neural network can effectively learn and utilize semantic information in the transmitted content, so that its communication performance under the same channel environment far exceeds that of a traditional communication system, meeting the requirements of semantic communication.
[0003] Although DECS exhibits strong semantic communication capabilities, it also has the weakness common to deep neural network systems: it is very vulnerable to adversarial attacks. In the related art, some common deep neural network adversarial attack methods can be migrated to perform a significant attack on DECS. Among them, the attack on the communication system is mainly completed at the physical channel level: through an adversarial attack algorithm, a tiny perturbation signal is constructed, and when the signal is superimposed on the original signal, the semantic information in the signal is destroyed, and the decoding neural network cannot correctly restore the transmitted content from the superimposed signal. Since the content transmitted in the channel is unknown, a universal adversarial perturbation signal needs to be constructed, so that it has an attack effect on any transmitted content.
[0004] However, existing attack methods all focus on attacking the channel itself, making the channel unable to effectively transmit data. However, since the encoding and decoding method learned by DECS is related to the semantic content of the transmitted content, further attacks on specific semantic content can be implemented, which is a more covert and potentially more dangerous way. SUMMARY
[0005] Embodiments of the present application provide a channel transmission method, device, electronic device, and medium. To solve the problem that the communication system has security risks caused by attacks on the semantic content of the transmitted signal in the related art.
[0006] According to an aspect of an embodiment of the present application, a channel transmission method is provided, applied to a DECS based on deep learning, including:
[0007] When it is detected that the sending end of the DECS sends an original signal, a perturbation signal is generated by the perturbation generator, and the perturbation signal is used to cover the target semantic information of the original signal;
[0008] superimposing the original signal and the perturbation signal to generate a perturbation transmission signal;
[0009] transmitting the perturbation transmission signal to a receiving end of the DECS.
[0010] Optionally, in another embodiment based on the above method of the present application, before the sending end of the DECS sends the original signal, the method further comprises:
[0011] obtaining a to-be-trained perturbation generator and a training data set, the training data set being used to instruct the to-be-trained perturbation generator to replace the target semantic information in the signal with other semantic information;
[0012] training the to-be-trained perturbation generator based on the training data set until the perturbation generator is obtained.
[0013] Optionally, in another embodiment based on the above method of the present application, the training of the to-be-trained perturbation generator based on the training data set until the perturbation generator is obtained comprises:
[0014] controlling the encoder parameters and the decoder parameters in the DECS to be fixed parameters;
[0015] training the to-be-trained perturbation generator based on the training data set until the perturbation generator is obtained in the DECS.
[0016] Optionally, in another embodiment based on the above method of the present application, the training of the to-be-trained perturbation generator based on the training data set until the perturbation generator is obtained in the DECS comprises:
[0017] sending a to-be-trained perturbation signal generated by the to-be-trained perturbation generator to the receiving end of the DECS through the sending end of the DECS;
[0018] decoding the received to-be-trained perturbation signal by the receiving end of the DECS and calculating a target function of the to-be-trained perturbation signal;
[0019] updating the parameter gradient of the to-be-trained perturbation generator based on the target function and a back propagation technique to reduce the target function of the to-be-trained perturbation signal until the trained perturbation generator is obtained.
[0020] Optionally, in another embodiment based on the above method of the present application, the training of the to-be-trained perturbation generator based on the training data set until the perturbation generator is obtained in the DECS comprises:
[0021] a signal strength constraint function is set for the to-be-trained disturbance generator, where the signal strength constraint function is used to control the signal strength of the disturbance signal to be lower than a preset threshold.
[0022] Optionally, in another embodiment based on the above method of the present application, the training of the to-be-trained disturbance generator in the DECS based on the training data set until the disturbance generator is obtained comprises:
[0023] a dispersion degree function is set for the to-be-trained disturbance generator, where the dispersion degree function is used to control the dispersion degree of the disturbance signal superimposed on the original signal.
[0024] Optionally, in another embodiment based on the above method of the present application, the detection of the sending of the original signal by the sending end of the DECS, the generation of the disturbance signal by the disturbance generator comprises:
[0025] When the sending of the original signal by the sending end is detected, the disturbance generator randomly generates the disturbance signal based on a random number.
[0026] According to yet another aspect of embodiments of the present application, a device for channel transmission is provided, which is applied to a deep learning-based end-to-end communication system (DECS) and comprises:
[0027] a detection module configured to detect the sending of an original signal by a sending end of the DECS, and generate a disturbance signal by the disturbance generator, the disturbance signal being used to cover target semantic information of the original signal;
[0028] a generation module configured to superimpose the original signal and the disturbance signal to generate a disturbance transmission signal;
[0029] a transmission module configured to transmit the disturbance transmission signal to a receiving end of the DECS.
[0030] According to yet another aspect of embodiments of the present application, an electronic device is provided, which comprises:
[0031] a memory configured to store executable instructions; and
[0032] a processor configured to execute the executable instructions with the memory to complete the operations of any of the above-mentioned methods for channel transmission.
[0033] According to yet another aspect of embodiments of the present application, a computer readable storage medium is provided, which is configured to store computer readable instructions, the instructions being executed to perform the operations of any of the above-mentioned methods for channel transmission.
[0034] In the present application, when the sending end of the DECS detects the sending of the original signal, a disturbance signal can be generated by a disturbance generator, the disturbance signal being used to cover the target semantic information of the original signal; the original signal and the disturbance signal are superimposed to generate a disturbance transmission signal; and the disturbance transmission signal is transmitted to the receiving end of the DECS. By applying the technical solution of the present application, the disturbance signal generated by deploying a disturbance generator at the sending end of the DECS communication system can destroy the original transmission signal with specific semantic information in the channel. Thus, the purpose of attacking the specific transmission signal in the hidden channel without affecting the transmission signal of other semantic information is achieved. This method can be used for model security evaluation in the production environment, and can be used for training the original model by fusing the adversarial samples, enriching the training sample space, and helping to improve the robustness of the end-to-end communication system based on the deep neural network. On the other hand, the problem of security risks in the communication system caused by the inability to attack the semantic content of the transmission signal in the related art can also be avoided.
[0035] The technical solution of the present application will be described in further detail below by means of the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS
[0036] The accompanying drawings, which form a part of the specification, illustrate embodiments of the present application and, together with the description, serve to explain the principles of the present application.
[0037] The present application can be more clearly understood by referring to the following detailed description in conjunction with the accompanying drawings, in which:
[0038] Figure 1 A schematic diagram of a channel transmission method is proposed for the present application;
[0039] Figure 2 A schematic diagram of the flow of a channel transmission method is proposed for the present application;
[0040] Figure 3 A schematic diagram of the system architecture of an end-to-end communication system DECS based on deep learning is proposed for the present application;
[0041] Figure 4 A schematic diagram of the system flow of a channel transmission method is proposed for the present application;
[0042] Figure 5 A schematic diagram of the structure of a channel transmission device is proposed for the present application;
[0043] Figure 6 A schematic diagram of the structure of an electronic device is proposed for the present application. DETAILED DESCRIPTION
[0044] Various exemplary embodiments of the present application will now be described in detail with reference to the accompanying drawings. Note that the relative arrangement, numerical expressions, and numerical values of components and steps set forth in these embodiments are not limiting to the scope of the present application unless specifically stated otherwise.
[0045] It should be understood that the sizes of the respective portions depicted in the drawings are shown for illustrative purposes only and are not according to the actual scale.
[0046] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the scope of the application or its application or uses.
[0047] Techniques, methods, and apparatus known to those of ordinary skill in the relevant art can not be discussed in detail herein, but should be considered as part of the specification, where appropriate.
[0048] It should be noted that like reference numerals and letters refer to like items in the following drawings and thus, once an item is defined in one drawing, it is not necessary to discuss it further in subsequent drawings.
[0049] In addition, the technical solutions among the various embodiments of the present application can be combined with each other, but it must be based on the fact that a person of ordinary skill in the art can realize it, and when the combination of technical solutions contradicts each other or cannot be realized, it should be considered that the combination of technical solutions does not exist and is not within the scope of protection claimed by the present application.
[0050] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative positional relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications will also change accordingly.
[0051] The method for performing channel transmission according to the exemplary embodiments of the present application will be described below with reference to the accompanying drawings. Figures 1-4 It should be noted that the following application scenarios are only shown for the purpose of facilitating the understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect. On the contrary, the embodiments of the present application can be applied to any applicable scenario.
[0052] The present application also proposes a method and device for channel transmission, an electronic device and a medium.
[0053] Figure 1 A flowchart of a method for channel transmission according to an embodiment of the present application is schematically shown. As shown in Figure 1 the method is applied to a deep learning-based end-to-end communication system DECS, comprising:
[0054] S101, when detecting that the sending end of the DECS sends an original signal, a disturbance signal is generated through a disturbance generator, and the disturbance signal is used to cover the target semantic information of the original signal.
[0055] S102, the original signal and the disturbance signal are superimposed to generate a disturbance transmission signal.
[0056] S103, the disturbance transmission signal is transmitted to the receiving end of the DECS.
[0057] In the related art, the communication system has developed rapidly in the past few decades, but it has gradually approached the Shannon limit, especially in the wireless channel, the point-to-point channel capacity has approached the capacity limit. In order to break through the limitation of the Shannon limit and further improve the communication transmission capacity, more and more research attention is transferred from the traditional communication system to the next generation semantic communication system. The traditional communication model focuses on how to transmit symbol information in the channel at high speed and accurately, and the continuous symbol stream carries semantic information. However, this multi-level and complex modular system makes the artificially defined symbol coding rule unable to fully utilize the semantic information of the transmission content, and inevitably produces information redundancy. The semantic communication system refers to the communication system that can directly transmit semantic information, which requires the communication system to be able to fully extract and utilize the semantic information in the transmission content. Such a communication system has greater capacity potential and better communication performance.
[0058] The latest research shows that the end-to-end communication system constructed using deep neural network (DNN-based End-to-End Communication System, DECS) can effectively learn and utilize the semantic information in the transmission content, so that its communication performance in the same channel environment far exceeds that of the traditional communication system, and meets the requirements of semantic communication. In this system, the encoding module of the sending end and the decoding module of the receiving end are replaced by a neural network respectively. Through deep learning, the neural networks in the receiving and transmitting modules are jointly trained under the influence of simulating the real channel, so as to learn the best information coding and decoding mode.
[0059] While DECS demonstrates powerful semantic communication capabilities, it also shares a common weakness with deep neural network systems: vulnerability to adversarial attacks. Extensive experiments have demonstrated that common deep neural network adversarial attack methods can be transferred to effectively attack DECS. Attacks on communication systems primarily occur at the physical channel level: adversarial attack algorithms construct tiny perturbation signals. When superimposed on the original signal, these signals destroy the semantic information in the signal, making it impossible for the decoding neural network to correctly recover the transmitted content. Because the content transmitted in the channel is unknowable, it is necessary to construct universal adversarial perturbation signals that are effective against any transmitted content. One of the most advanced methods for physical adversarial attacks in wireless channels is the perturbation generation model. This method trains a neural network to generate a large number of random universal adversarial perturbation signals. Constrained during training, the attack signals resemble natural noise, making them difficult to detect and remove.
[0060] Existing attack methods focus on attacking the channel itself, rendering it ineffective for data transmission. However, because the encoding and decoding methods learned within DECS are semantically related to the transmitted content, attacks targeting specific semantic content can be further implemented, which is a more covert and potentially more harmful method.
[0061] Furthermore, in order to address the above-mentioned problems, Figure 2 As shown, the embodiment of the present application proposes a channel transmission method of an end-to-end communication system DECS based on deep learning, which includes:
[0062] S1: Building a training structure: Obtaining a complete model of DECS Training data set Ω; modify the original model and add an initial perturbation generator to be trained at the sending end The disturbance signal to be trained sent by the disturbance generator to be trained is superimposed on the DECS original signal.
[0063] S2: Setting up training content: Freezing the DECS model The parameter set θ is such that the training process only modifies the perturbation generator to be trained The parameter set η in .
[0064] S3: Replace dataset labels: traverse the training dataset Ω and replace all target semantic labels with random other semantic labels.
[0065] S4: Limit the attack signal strength: According to the concealment requirement of the attack signal, its signal strength should be lower than the set value Pmax, so in An output signal strength constraint function is added. As an example, the signal strength constraint function can be:
[0066]
[0067] S5: Set the objective function
[0068]
[0069] wherein, represents the loss function of the data service category (such as image recognition, semantic segmentation, speech recognition, etc.) corresponding to the original signal of this transmission; is a function of measuring the model The function of the discrete degree of the disturbance signal is generated. By increasing its value, the attack signal generated by the perturber is as dispersed as possible. Thus, the purpose of being difficult to predict and defend by the receiver is achieved.
[0070] S6: Train the to-be-trained perturbation generator model, specifically: read the data set after replacing the label, input to the communication model in batches , and add the to-be-trained perturbation signal generated by the to-be-trained perturbation generator to the original signal , and obtain the to-be-trained perturbation signal.
[0071] Further, the receiving end of the DECS receives the to-be-trained perturbation signal and decodes it, calculates the objective function corresponding to the to-be-trained perturbation signal through the back propagation technology, calculates the gradient of each parameter of the neural network, updates the parameters of the perturbation generator model , so that the objective function value decreases. Until the target perturbation generator is obtained.
[0072] S7: Verify the effect of the perturbation generator: deploy the trained perturbation generator to the sending end of the DECS. To achieve the following steps:
[0073] When the DECS detects that the sending end sends the original signal, the perturbation signal is generated by the perturbation generator, and the perturbation signal is used to cover the target semantic information of the original signal; the original signal and the perturbation signal are superimposed to generate a perturbation transmission signal; and the perturbation transmission signal is transmitted to the receiving end of the DECS.
[0074] It should be noted that the channel transmission method proposed in the embodiments of the present application can on the one hand attack the content containing specific semantic information without knowing the original signal transmission content in the channel. On the other hand, it can also be used to evaluate the security performance of the model of DECS in the actual application environment and find vulnerabilities. Still on the other hand, it can also be used for the adversarial training process of the original model, expand the original training set space, and improve the robustness of DECS.
[0075] In the present application, when it is detected that the sending end of DECS sends an original signal, a perturbation signal can be generated by a perturbation generator, the perturbation signal being used to cover the target semantic information of the original signal; the original signal and the perturbation signal are superimposed to generate a perturbation transmission signal; and the perturbation transmission signal is transmitted to the receiving end of DECS. By applying the technical solution of the present application, the perturbation signal generated by deploying a perturbation generator at the sending end of the DECS communication system can destroy the original transmission signal in the channel with specific semantic information. Thus, the purpose of attacking the specific transmission signal in the channel in a hidden manner can be achieved without affecting the transmission of other semantic information. This method can on the one hand be used for the evaluation of model security in the production environment, fuse adversarial samples to train the original model, enrich the training sample space, and help improve the robustness of the end-to-end communication system based on deep neural network. On the other hand, it can also avoid the problem of security risks in the communication system caused by the inability to attack the semantic content of the transmission signal in the related art.
[0076] Optionally, in another embodiment based on the above method of the present application, before detecting that the sending end of DECS sends an original signal, the method further comprises:
[0077] obtaining a perturbation generator to be trained and a training data set, the training data set being used to instruct the perturbation generator to be trained to replace the target semantic information in the signal with other semantic information;
[0078] training the perturbation generator to be trained based on the training data set until the perturbation generator is obtained.
[0079] Optionally, in another embodiment based on the above method of the present application, training the perturbation generator to be trained based on the training data set until the perturbation generator is obtained comprises:
[0080] controlling the encoder parameters and the decoder parameters in DECS to be fixed parameters;
[0081] In DECS, the perturbation generator to be trained is trained based on the training data set until the perturbation generator is obtained.
[0082] Optionally, in another embodiment based on the above method of the application, in the DECS, the training data set is used to train the to-be-trained perturbation generator until the perturbation generator is obtained, including:
[0083] The to-be-trained perturbation signal is sent to the receiving end of the DECS through the sending end of the DECS by using the perturbation signal generated by the to-be-trained perturbation generator.
[0084] The receiving end of the DECS decodes the received to-be-trained perturbation signal and calculates the objective function of the to-be-trained perturbation signal.
[0085] Based on the objective function and the back propagation technology, the parameter gradient of the to-be-trained perturbation generator is updated to reduce the objective function of the to-be-trained perturbation signal until the trained perturbation generator is obtained.
[0086] Optionally, in another embodiment based on the above method of the application, in the DECS, the training data set is used to train the to-be-trained perturbation generator until the perturbation generator is obtained, including:
[0087] A signal strength constraint function is set for the to-be-trained perturbation generator, wherein the signal strength constraint function is used to control the signal strength of the perturbation signal to be lower than a preset threshold.
[0088] Optionally, in another embodiment based on the above method of the application, in the DECS, the training data set is used to train the to-be-trained perturbation generator until the perturbation generator is obtained, including:
[0089] A dispersion degree function is set for the to-be-trained perturbation generator, wherein the dispersion degree function is used to control the dispersion degree of the perturbation signal superimposed on the original signal.
[0090] Optionally, in another embodiment based on the above method of the application, when the sending end of the DECS sends the original signal, the perturbation signal is generated by the perturbation generator, including:
[0091] When the sending end sends the original signal, the perturbation generator randomly generates the perturbation signal based on a random number.
[0092] Further, as shown in Figure 3 , the system architecture diagram of the DECS based on deep learning proposed in the application is shown. It includes:
[0093] Step a, obtaining the complete model of the target DECS As shown in Figure 3 , It is generally composed of two modules, namely the encoder E and the decoder D, and the function C is used to simulate the transformation of the signal by the channel.
[0094] As an example, input a picture x in the system, the encoder encodes it into a signal is denoted as After superimposing the natural noise N, the channel transformation is performed, and the output terminal receives the signal Y, which is denoted as The output terminal decodes Y into a picture x', which is denoted as x' = D(Y). The encoder E and the decoder D are trained neural networks, and the channel transformation function C depends on the specific channel type selected.
[0095] Step b, add a trained perturbation generator where z represents a random number seed, and inputting different random numbers will make the perturbation generator output different perturbation signals.
[0096] where the structure of the perturbation generator is a multi-layer perceptron, and the activation function uses LeakyReLU. The perturbation signal generated by the perturbation generator will be superimposed on the original signal, so that the perturbation transmission signal after adding the perturbation signal
[0097] where the entire communication process after adding the perturbation signal is is denoted as
[0098] Further, the generation process of the perturbation generator is described as follows:
[0099] First, freeze the parameter set θ of the DECS model, i.e., the parameters of the encoder E and the decoder D. The variable parameter during training is the parameter set η in the perturbation generator to be trained.
[0100] In addition, according to the object to be attacked, set the target label value. For example Figure 4 , the attack target is a car class picture, and the target label variable target_label is set to the corresponding value of the car class.
[0101] where the training data set Ω is traversed, and all target class labels (i.e., labels corresponding to target semantic information) are replaced with random other class labels (i.e., labels corresponding to other semantic information).
[0102] As an example, the pseudo code is denoted as:
[0103] For x j inΩ:
[0104] If x j .label==target_lable then
[0105] x j label <- random_other_lable
[0106] Further, in order to ensure the concealment of the disturbance signal, the strength of the disturbance signal needs to be reduced as much as possible in the embodiments of the present application. In one way, the signal strength should be lower than the set value Pmax, so that the signal strength constraint function is added at the output.
[0107] As an example, the signal strength constraint function can be:
[0108]
[0109] In one way, when the disturbance generator output signal exceeds the set maximum value, the Z-Score method is used to standardize the data, and a signal with a Gaussian noise-like signal strength of is obtained.
[0110] In an optional way, during the training of the to-be-trained disturbance generator, a signal dispersion degree function is also needed to be set, which is used to represent the dispersion degree of the disturbance signal generated by the model.
[0111] As an example, the signal dispersion degree function is:
[0112]
[0113] In the signal dispersion degree function, the average distance of the disturbance signal generated by the model is estimated by sampling. Wherein k is an artificially selected hyperparameter, representing the number of samples each time.
[0114] Further, during the training of the to-be-trained disturbance generator, an objective function is also needed to be set.
[0115]
[0116] Wherein, represents the loss function of image recognition, using the mean square error (MSE) function; is a function for measuring the dispersion degree of the disturbance signal generated by the model, and by increasing its value, the attack signal generated by the disturbance generator is as dispersed as possible, and it is difficult for the receiver to predict and defend. In one way, the embodiments of the present application read a batch of pictures x j of a fixed size from the training data set Ω after replacing the label, and obtain a random number z, input the to-be-trained disturbance generator, and obtain the to-be-trained disturbance signal
[0117]
[0118] Further, the input picture x j is input into the encoder of the DECS, and after interference by the perturbation signal, channel transformation, decoding by the decoder, and downstream image recognition, a recognition label l rec is obtained. The DECS decoder uses the mean square error function MSE to calculate the image recognition loss, and uses the function to calculate the model to be trained to obtain the perturbation signal dispersion.
[0119] Finally, according to the image recognition loss and the perturbation signal dispersion, the value of the target function of the perturbation transmission signal to be trained is calculated. The gradient value of is obtained by back propagation, and the parameters of the perturbation generator model are updated to complete gradient descent. Until all the training data set Ω is read, the perturbation generator is trained.
[0120] Optionally, in one way, the above steps can also be repeatedly executed according to the preset number of iterations until the preset number of iterations is determined to be completed, thereby obtaining the trained perturbation generator.
[0121] Step c: superimpose the perturbation signal generated by the trained perturbation generator with the original signal to generate a perturbation transmission signal.
[0122] Step d: the sending end of the DECS transmits the perturbation transmission signal to the receiving end of the DECS.
[0123] By applying the technical solution of the present application, the perturbation signal generated by deploying a perturbation generator at the sending end of the DECS communication system can destroy the original transmission signal with specific semantic information in the channel. Thus, the purpose of attacking the specific transmission signal in the channel in a hidden manner is achieved without affecting the transmission of other semantic information. This method can be used for model security evaluation in production environment, and can be used for training original models by fusing adversarial samples, enriching training sample space, and helping to improve the robustness of end-to-end communication systems based on deep neural networks. On the other hand, it can also avoid the problem of security risks in the communication system caused by the inability to attack the semantic content of the transmission signal in related technologies.
[0124] Optionally, in another embodiment of the present application, as shown in Figure 5 the present application also provides a channel transmission device. Applied to a deep learning-based end-to-end communication system DECS, comprising:
[0125] The detection module 201 is configured to, when detecting that the sending end of the DECS sends an original signal, generate a perturbation signal by the perturbation generator, the perturbation signal being used to cover the target semantic information of the original signal.
[0126] The generation module 202 is configured to superimpose the original signal and the perturbation signal to generate a perturbation transmission signal.
[0127] The transmission module 203 is configured to transmit the perturbation transmission signal to the receiving end of the DECS.
[0128] In another embodiment of the present application, the generation module 202 is configured to perform the following steps:
[0129] Obtain a to-be-trained perturbation generator and a training data set, the training data set being used to instruct the to-be-trained perturbation generator to replace the target semantic information in a signal with other semantic information.
[0130] Train the to-be-trained perturbation generator based on the training data set until the perturbation generator is obtained.
[0131] In another embodiment of the present application, the generation module 202 is configured to perform the following steps:
[0132] Control the encoder parameters and the decoder parameters in the DECS to be fixed parameters.
[0133] In the DECS, train the to-be-trained perturbation generator based on the training data set until the perturbation generator is obtained.
[0134] In another embodiment of the present application, the generation module 202 is configured to perform the following steps:
[0135] Use the perturbation signal generated by the to-be-trained perturbation generator to send a to-be-trained perturbation signal to the receiving end of the DECS through the sending end of the DECS.
[0136] The receiving end of the DECS decodes the received to-be-trained perturbation signal and calculates the target function of the to-be-trained perturbation signal.
[0137] Based on the target function and the back propagation technology, update the parameter gradient of the to-be-trained perturbation generator to reduce the target function of the to-be-trained perturbation signal until the trained perturbation generator is obtained.
[0138] In another embodiment of the present application, the generation module 202 is configured to perform the following steps:
[0139] a signal strength constraint function is set for the to-be-trained disturbance generator, wherein the signal strength constraint function is used to control the signal strength of the disturbance signal to be lower than a preset threshold.
[0140] In another embodiment of the present application, the generating module 202 is configured to perform steps including:
[0141] a dispersion degree function is set for the to-be-trained disturbance generator, wherein the dispersion degree function is used to control the dispersion degree of the disturbance signal superimposed on the original signal.
[0142] In another embodiment of the present application, the generating module 202 is configured to perform steps including:
[0143] When the sending end sends the original signal, the disturbance generator randomly generates the disturbance signal based on a random number.
[0144] Figure 6 is a logical structure block diagram of an electronic device according to an exemplary embodiment. For example, the electronic device 300 can be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a vehicle device, a personal digital assistant, etc.
[0145] In an exemplary embodiment, a non-transitory computer readable storage medium including instructions, such as a memory including instructions, is also provided, which can be executed by an electronic device processor to complete the above-mentioned channel transmission method, which includes: when the sending end of the DECS sends the original signal, generating a disturbance signal through the disturbance generator, the disturbance signal is used to cover the target semantic information of the original signal; superimposing the original signal and the disturbance signal to generate a disturbance transmission signal; transmitting the disturbance transmission signal to the receiving end of the DECS. Optionally, the above-mentioned instructions can also be executed by the processor of the electronic device to complete other steps involved in the above-mentioned exemplary embodiments. For example, the non-transitory computer readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.
[0146] In the example embodiment, an application program / computer program product is also provided, comprising one or more instructions executable by a processor of an electronic device to complete the method of channel transmission described above, which comprises: when detecting that a sending end of the DECS sends an original signal, generating a perturbation signal through the perturbation generator, the perturbation signal being used to cover target semantic information of the original signal; superimposing the original signal and the perturbation signal to generate a perturbation transmission signal; and transmitting the perturbation transmission signal to a receiving end of the DECS. Optionally, the above instructions can also be executed by the processor of the electronic device to complete other steps involved in the example embodiment described above.
[0147] Figure 6 An example diagram of the electronic device 300 is shown. Those skilled in the art can understand that the schematic diagram Figure 6 The electronic device 300 is only an example and does not constitute a limitation on the electronic device 300, which can include more or fewer components than shown, or combine certain components, or different components, for example, the electronic device 300 can also include an input / output device, a network access device, a bus, etc.
[0148] The processor 302 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor 302 can also be any conventional processor. The processor 302 is the control center of the electronic device 300, and is connected to various parts of the electronic device 300 through various interfaces and lines.
[0149] The memory 301 can be used to store computer-readable instructions 303, and the processor 302 implements various functions of the electronic device 300 by running or executing the computer-readable instructions or modules stored in the memory 301, and calling data stored in the memory 301. The memory 301 can mainly include a storage program area and a storage data area, wherein the storage program area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), and the like; and the storage data area can store data created according to the use of the electronic device 300, etc. In addition, the memory 301 can include a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash storage device, a read-only memory (ROM), a random access memory (RAM), or other non-volatile / volatile storage devices.
[0150] The modules integrated in the electronic device 300, if implemented in the form of software function modules and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, all or part of the flow of the method in the above-mentioned embodiments can also be implemented by computer readable instructions instructing related hardware to complete, and the computer readable instructions can be stored in a computer readable storage medium. When the computer readable instructions are executed by a processor, the steps of the above-mentioned various method embodiments can be implemented.
[0151] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the application be limited only by the scope of the claims, a properly construed, and any variations, uses, or adaptations of the application following in general the principles of the application and including such as come within the scope of the following claims are intended to be within the scope of the application. The specification and examples given are intended as illustrative only and not in a limiting sense. For example, the steps recited in the specification can be performed in any order possible.
[0152] It is to be understood that the application is not limited to the precise construction described in the specification and shown in the drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application is to be defined by the claims appended hereto.
Claims
1. A method of channel transmission applied to a deep learning based end-to-end communication system (DECS), characterized in that, The DECS comprises a disturbance generator arranged at a sending end of the DECS, and the disturbance generator comprises: When it is detected that the sending end of the DECS sends an original signal, a disturbance signal is generated by the disturbance generator, and the disturbance signal is used to cover target semantic information of the original signal; The original signal and the disturbance signal are superimposed to generate a disturbance transmission signal; The disturbance transmission signal is transmitted to a receiving end of the DECS; Before the detection that the sending end of the DECS sends the original signal, the method further comprises: obtaining a to-be-trained disturbance generator and a training data set, the training data set being used to instruct the to-be-trained disturbance generator to replace the target semantic information in the signal with other semantic information; and training the to-be-trained disturbance generator based on the training data set until the disturbance generator is obtained; training the to-be-trained disturbance generator based on the training data set until the disturbance generator is obtained, comprising: training the to-be-trained disturbance generator based on the training data set until the disturbance generator is obtained in the DECS; the training the to-be-trained disturbance generator based on the training data set until the disturbance generator is obtained in the DECS, comprising: using a disturbance signal generated by the to-be-trained disturbance generator to send a to-be-trained disturbance signal from the sending end of the DECS to the receiving end of the DECS; the receiving end of the DECS decodes the received to-be-trained disturbance signal, calculates an objective function of the to-be-trained disturbance signal, updates a parameter gradient of the to-be-trained disturbance generator based on the objective function and a back propagation technique, so that the objective function of the to-be-trained disturbance signal is reduced until the trained disturbance generator is obtained.
2. The method of claim 1, wherein, the training the to-be-trained disturbance generator based on the training data set until the disturbance generator is obtained, further comprising: controlling an encoder parameter and a decoder parameter in the DECS to be fixed parameters.
3. The method of claim 1 or 2, wherein, the training the to-be-trained disturbance generator based on the training data set until the disturbance generator is obtained in the DECS, comprising: a signal strength constraint function is arranged for the to-be-trained disturbance generator, and the signal strength constraint function is used to control a signal strength of the disturbance signal to be lower than a preset threshold.
4. The method of claim 3, wherein, the training the to-be-trained disturbance generator based on the training data set until the disturbance generator is obtained in the DECS, comprising: a dispersion degree function is arranged for the to-be-trained disturbance generator, and the dispersion degree function is used to control a dispersion degree of the disturbance signal superimposed on the original signal.
5. The method of claim 1, wherein, the generating the disturbance signal by the disturbance generator when it is detected that the sending end of the DECS sends the original signal, comprising: when it is detected that the sending end sends the original signal, the disturbance generator randomly generates the disturbance signal based on a random number.
6. An apparatus for channel transmission, characterized by application to a deep learning-based end-to-end communication system DECS, comprising: The detection module is configured to, when detecting that the sending end of the DECS sends an original signal, generate a disturbance signal through a disturbance generator, the disturbance signal being used to cover target semantic information of the original signal. The generation module is configured to superimpose the original signal and the disturbance signal to generate a disturbance transmission signal. The transmission module is configured to transmit the disturbance transmission signal to the receiving end of the DECS. The device is further configured to: Before the detection that the sending end of the DECS sends the original signal, acquire a to-be-trained disturbance generator and a training data set, the training data set being used to instruct the to-be-trained disturbance generator to replace the target semantic information in a signal with other semantic information; and train the to-be-trained disturbance generator based on the training data set until the disturbance generator is obtained. The training of the to-be-trained disturbance generator based on the training data set until the disturbance generator is obtained includes: In the DECS, the to-be-trained disturbance generator is trained based on the training data set until the disturbance generator is obtained. The training of the to-be-trained disturbance generator based on the training data set until the disturbance generator is obtained includes: A to-be-trained disturbance signal is sent from the sending end of the DECS to the receiving end of the DECS by using a disturbance signal generated by the to-be-trained disturbance generator; the receiving end of the DECS decodes the received to-be-trained disturbance signal, calculates an objective function of the to-be-trained disturbance signal, updates a parameter gradient of the to-be-trained disturbance generator based on the objective function and a back propagation technique, so that the objective function of the to-be-trained disturbance signal is reduced, and the disturbance generator trained is obtained until the disturbance generator is trained.
7. An electronic device, comprising: including: a memory for storing executable instructions; and a processor for executing the executable instructions with the memory to complete the operations of the method of any one of claims 1-5. The instructions are executed to perform the operations of the method of any one of claims 1-5.
8. A computer readable storage medium storing computer readable instructions, characterized in that, The instructions are executed to perform the operations of the method of any one of claims 1-5.