A Distributed Radio Map Sensing Method for Privacy Protection under Wireless Networks

Abstract

The present invention discloses a distributed radio map sensing method for privacy protection under a wireless network, comprising the following steps: S1. Each user is assigned an independent local sub-model, and the user uses their respective local sub-models to extract local data features, and finally generates a deep local feature and uploads it to the server; S2. A decision sub-model is configured for the server, and the server receives the local features from all users and uses them as the input of the decision sub-model. The decision sub-model visualizes the transmitter information and forms a two-channel picture with the static geographical environment image; S3. The two-channel picture is input into a pre-trained U-Net model to generate a radio map, and finally a radio map estimation result is obtained. The present invention realizes high-precision real-time radio map estimation on the premise of protecting user privacy, and has better application universality; and assigns importance to the features from different users to improve the estimation accuracy and reduce the model complexity.

Description

Technical Field

[0001] The present invention relates to the field of radio map perception, and particularly to a distributed radio map perception method for privacy protection in a wireless network. Background Art

[0002] A radio map is a tool that can effectively depict the key features of the surrounding radio environment, such as the received signal strength (RSS) at any position within the area, path gain, interference strength, etc. Therefore, accurate and efficient radio map estimation (RME) has been a research hotspot in recent years. Traditional radio map estimation techniques such as interpolation algorithms, matrix completion algorithms, and ray tracing methods are difficult to achieve real-time and accurate radio map estimation. In recent years, the emerging deep learning (DNN)-based methods can achieve real-time and high-precision radio map estimation due to their high-quality datasets and low-complexity DNN structures.

[0003] However, the current deep learning methods can only be applied to the case of centralized data processing. Since the received signal strength information of the user terminals in a wireless network is distributed among the user terminals, considering the issue of data privacy protection, such information is often difficult to be centrally uploaded to the server side. Therefore, how to handle real-time high-precision radio map perception in a distributed scenario remains one of the key challenges faced by current research. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a distributed radio map perception method for privacy protection in a wireless network, which can achieve high-precision real-time radio map estimation on the premise of protecting user privacy (i.e., without uploading local raw data), and has better application universality; and assigns importance to the features from different users to improve the estimation accuracy and reduce the model complexity.

[0005] The purpose of the present invention is achieved by the following technical solutions: A distributed radio map perception method for privacy protection in a wireless network, comprising the following steps:

[0006] S1. Assign an independent local sub-model to each user. The user uses their respective local sub-models to extract local data features, and finally generates a deep local feature and uploads it to the server;

[0007] S2. Configure a decision sub-model for the server. The server receives local features from all users and uses them as the input of the decision sub-model. The decision sub-model first assigns importance levels to data from different users using the attention mechanism in the model input layer, then infers the transmitter information through subsequent neural network layers, and finally visualizes the transmitter information to obtain a picture representing the signal strength of the transmitter position, which is combined with the static geographical environment image to form a two-channel picture;

[0008] S3. Input the two-channel picture into a pre-trained U-Net model to generate a radio map, and finally obtain the radio map estimation result.

[0009] The beneficial effects of the present invention are as follows: 1. The present invention proposes a two-stage radio map estimation framework from distributed estimation of transmitter information to radio map estimation. First, a distributed online estimation scheme for transmitter information based on the received signal strength information at the user side is designed, which can obtain high-precision transmitter position and strength estimation; after obtaining accurate transmitter information estimation, combined with the deep learning method based on UNet, the final radio map is inferred using an offline-trained deep model. This method can adapt to scenarios with dynamic changes in the number of transmitters and users, and has a wide range of applicable scenarios.

[0010] 2. In the transmitter information estimation stage, the architecture of vertical federated learning (VFL) is adopted for model training and inference. Under this architecture, sub-neural networks are deployed at both the user side and the server side. The sub-neural network at the user side is used to extract local data (i.e., the received signal strength information at the user side) features, and the sub-neural network at the server side is used for decision-making. All sub-neural networks cooperate for training and inference. When using this architecture for training and inference, users do not need to upload the original data (i.e., the received signal strength information at the user side), but only upload the local data features extracted by the sub-neural network, thus protecting user privacy.

[0011] 3. For the sub-neural network at the server side, an attention layer is introduced, and the attention mechanism is used to intelligently assign importance levels to different users, thereby improving the accuracy of transmitter information estimation. Description of the Drawings

[0012] Figure 1 is the flowchart of the method of the present invention;

[0013] Figure 2 is the schematic diagram of vertical federated learning;

[0014] Figure 3 is the schematic diagram of the collaborative training effect of the model under the vertical federated learning framework;

[0015] Figure 4Schematic diagram of the difference in estimation accuracy for different numbers of users. Detailed implementation manners

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0017] The present invention aims at distributed radio map perception that takes into account privacy protection in a wireless network. The overall inference framework consists of three parts, namely, a user feature extraction part, a transmitter information estimation part, and a radio map generation part. Among them, user feature extraction and transmitter information estimation are implemented on a vertical federated learning model, which consists of multiple local sub-models located at the user side and a decision sub-model located at the server, as Figure 2 shown; actual historical data can be collected to form a data set for training. The format of a single data sample in the data set is: the received signal strength at the user side location, the transmitter location and the transmitted signal strength; the radio map generation part is implemented on a U-Net model. This model can be pre-trained based on the existing data set RadioMapSeer (this part directly uses existing methods to generate a radio map using transmitter and geographical environment information), as Figure 1 shown. The method of the present invention includes:

[0018] User feature extraction: In this solution, each user is assigned an independent local sub-model. The user uses their respective local sub-models to extract the features of local data (i.e., the received signal strength information at the user side), and finally generates a deep local feature and uploads it to the server;

[0019] Transmitter information estimation: The server receives the local features from all users and uses them as the input of the decision sub-model. The decision sub-model first uses an attention mechanism in the model input layer to assign importance levels to the data from different users, and then infers the transmitter information through subsequent neural network layers. Finally, the transmitter information is visualized and combined with a static geographical environment image to form a two-channel picture;

[0020] Radio map generation: The two-channel picture is input into the pre-trained U-Net model to generate a radio map, and finally the radio map estimation result is obtained.

[0021] Specifically in the application scenario, assume that there are K users in each time period within the same L1*L2 area. Then the vertical federated learning model requires K + 1 sub-models, including K user local sub-models located at the user side and 1 decision sub-model located at the server side. The received signal strength at the user's location is determined by large-scale fading (including path loss and shadow fading), transmitter power, and additive noise. The specific implementation process of the solution in this scenario is:

[0022] Each user takes the user location and the received signal strength as the input of the local sub-model, and calculates the local features. The local sub-models of each user are all fully connected neural networks, but the parameters are independent and not shared;

[0023] The user uploads the local features to the server;

[0024] The server takes the local features received from all K users as the input of the decision sub-model. The decision sub-model consists of an attention layer at the input layer and a subsequent fully connected neural network.

[0025] The attention mechanism in the input layer of the decision sub-model performs adaptive importance detection and weight assignment on the user features;

[0026] The FCNN (fully connected neural network) processes the user features with the importance weights added, and obtains the transmitter information including the transmitter location and the transmitted signal strength;

[0027] The transmitter information is visualized as a grayscale image of size L1*L2, that is, the image of the signal strength representing the transmitter location, where the coordinates of the transmitter correspond to the gray value of 255, and the gray value at the remaining coordinates is 0, and it forms a two-channel image with the geographical grayscale image of the building environment image of size L1*L2 corresponding to the area;

[0028] The two-channel input is fed into the U-Net network to generate the radio map;

[0029] The transmitter information and the radio map are output.

[0030] The local sub-model and the decision sub-model are trained in the way of vertical federated learning. The training process includes:

[0031] A1. For data sample n, each user k extracts the deep data features using the local sub-model and uploads them to the server; the server receives the deep data features from all users as the input of the decision sub-model and obtains the predicted value of the transmitter information;

[0032] A2. At the server, calculate the mean square error between the predicted value of the transmitter information and the true label as the training loss;

[0033] A3. Traverse all data samples, calculate the training loss, and then calculate the gradient of the decision sub-model based on this training loss by the backpropagation method;

[0034] A4. For each user k, calculate the partial derivative of the training loss with respect to the deep data features output by user k (traverse the data samples and sum) and backpropagate it to this user;

[0035] A5. Based on the above-mentioned backpropagated partial derivatives, each user k calculates the gradient of the local sub-model using the backpropagation method;

[0036] A6. Update the parameters of all local sub-models and the decision sub-model;

[0037] A7. Loop through A1 - A6 until the set number of training rounds is reached.

[0038] In the sample set used for the vertical federated learning, there are multiple samples;

[0039] The label of each sample is saved in the server, which is the transmitter position and intensity;

[0040] The sample feature of each sample includes: the local feature at each user, and the local feature is the user position and the corresponding received signal strength;

[0041] Each sample is obtained by the server collecting the transmitter position and intensity at different historical moments, and each user terminal collecting its own position and the corresponding received signal strength.

[0042] When the U-Net model is pre-trained, it is necessary to first construct a sample set. The sample feature is: a two-channel picture, which consists of a picture representing the signal strength of the transmitter position and a geographical grayscale picture of the building environment picture; the sample label is radio map information;

[0043] After pre-training the U-Net model with the sample set, the U-Net model can predict the radio map information according to the input two-channel picture;

[0044] As Figure 3 shown, it is the model collaborative training effect under the vertical federated learning framework (the number of users is 128, and the error situations of estimating the transmitter position and intensity with and without attention are respectively shown); the error change situation during the training process is recorded. The error includes two parts - the transmitter position estimation error and the transmitter intensity estimation error. Among them, the error of the network with attention decreases faster, indicating that adding the attention mechanism can help the VFL network achieve higher estimation accuracy. Figure 4For the differences in estimation accuracy under different numbers of users, the numbers of users are respectively (a): 64, (b): 128, (c): 200. (d) shows the relationship between the number of users and the accuracy of the final radiomap output with the true label as a comparison (the final radiomap is the final output of two cascaded networks, VFL + UNet). When the number of users > 64, the estimated radiomap is already relatively close to the true radiomap label. When the number of users > 200, the estimated radiomap is basically indistinguishable from the true radiomap label, indicating that an increase in the number of users can improve the estimation accuracy of the radiomap, and when the number of users reaches more than 200, the estimated radiomap is accurate enough.

[0045] The above is the preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in other combinations, modifications, and environments, and can be changed within the scope of the concept described herein through the above teachings or the technology or knowledge in the relevant field. And the changes and alterations made by those skilled in the art that do not depart from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention.

Claims

1. A distributed radio map perception method with privacy protection in a wireless network, characterized by: The following steps are involved: S1. Each user is assigned an independent local sub-model. The user uses their own local sub-model to extract local data features, and finally generates a deep local feature and uploads it to the server; S2. Configure the decision sub-model for the server. The server receives local features from all users as inputs to the decision sub-model. The decision sub-model first uses the attention mechanism in the model input layer to assign importance to data from different users, and then obtains transmitter information through subsequent neural network layer reasoning. Finally, the transmitter information is visualized to obtain a picture representing the signal strength of the transmitter location, and forms a two-channel picture with the building environment picture. The local sub-model and the decision sub-model are trained by vertical federated learning. The training process includes: A1. For data sample n, each user k uses the local sub-model to extract deep data features and uploads them to the server; the server receives the deep data features from all users as input to the decision sub-model and obtains the predicted value of the transmitter information; A2. Calculate the mean square error between the predicted value of the transmitter information and the true label at the server as the training loss; A3. After traversing all data samples and calculating the training loss, the gradient of the decision sub-model is calculated based on the training loss through back propagation. A4. For each user k, the server calculates the partial derivative of the training loss with respect to the deep data features output by user k and transmits it back to the user; A5. Based on the partial derivatives transmitted above, each user k calculates the gradient of the local sub-model using the back-propagation method. A6. Update all local sub-model and decision sub-model parameters; A7, loop A1-A6 until the set training round is reached; S3. Input the two-channel image into the pre-trained U-Net model to generate a radio map, and finally obtain the radio map estimation result.

2. According to claim 1, a distributed radio map perception method for privacy protection in a wireless network is characterized by: The local data is the signal strength information received by the user end.

3. The distributed radio map perception method for privacy protection in a wireless network according to claim 1, characterized in that: In step S1, each user uses the user position and received signal strength as inputs of a local sub-model to calculate local features; wherein each user's local sub-model is a fully connected neural network, but the parameters are independent and not shared.

4. The distributed radio map perception method for privacy protection in a wireless network according to claim 1, characterized in that: The step S2 comprises: S201. The server uses the local features received from all K users as inputs of a decision sub-model, where the decision sub-model consists of an attention layer located at the input layer and a subsequent fully connected neural network; S202. The attention mechanism of the input layer of the decision sub-model performs adaptive importance detection and weight allocation on user features; S203. Using a fully connected neural network to process user features carrying importance weights, transmitter information including transmitter location and transmission signal strength is obtained; S304. Visualize the transmitter information as a grayscale image of size L1*L2, that is, a picture representing the signal strength at the transmitter location, where the grayscale value corresponding to the coordinates of the transmitter is 255, and the other coordinates are 0, and form a two-channel picture with the geographic grayscale map of the building environment image of size L1*L2 corresponding to the area.

5. The distributed radio map perception method for privacy protection in a wireless network according to claim 1, characterized in that: The sample set used in the vertical federated learning includes multiple samples; The label of each sample is stored in the server, which is the transmitter location and strength; The sample features of each sample include: local features located at each user, namely the user position and the corresponding received signal strength; Each sample is obtained by the server collecting the transmitter position and strength at different historical moments, and each user terminal collecting its own position and the corresponding received signal strength.

6. The distributed radio map perception method for privacy protection in a wireless network according to claim 1, characterized in that: When the U-Net model is pre-trained, it is necessary to first construct a sample set, and the sample features are: a two-channel image, consisting of an image representing the signal strength of the transmitter location and a geographic grayscale image of the building environment image; the sample label is the radio map information; After pre-training the U-Net model with the sample set, the U-Net model can predict the radio map information based on the input two-channel image.

Images