Passive perception and non-inductive detection method based on multi-source wireless signal fusion

Abstract

The invention discloses a passive perception and non-inductive detection method based on multi-source wireless signal fusion, and the method comprises the steps: detecting each type of wireless signals of each type of wireless equipment through a wireless interface, extracting feature data as a feature sample, calculating the Cartesian product combination of the corresponding feature sample, and carrying out the detection of each type of wireless signals. The method comprises the following steps: acquiring a wireless signal fusion sample of each type of wireless equipment in different application scenarios, and training a wireless equipment detection model by using the wireless signal fusion sample and the corresponding wireless equipment type as training samples; wireless signals existing at the current moment are detected, feature samples are extracted, wireless signal fusion samples corresponding to all possible combinations of the current feature samples are generated, the trained wireless equipment detection model is input, and the wireless equipment model corresponding to the maximum confidence coefficient value is screened out to serve as a final detection result. According to the invention, multi-source wireless signals are passively sensed, and non-inductive monitoring under a dynamic scene of multiple wireless devices is realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of wireless communication, and more specifically relates to a passive sensing and non-sensing detection method based on multi-source wireless signal fusion. BACKGROUND

[0002] With the rapid development of the Internet of Things and smart homes, there is an increasing demand for passive monitoring of mobile targets (such as devices, terminals, etc.) in the environment. However, existing device fingerprinting techniques have the following problems in practical applications: active signal interaction is prone to privacy leakage risk, and may cause perceptible interference to the target device; single signal source dependency makes the system vulnerable to interference in complex dynamic scenarios, resulting in decreased recognition accuracy and robustness; traditional models have significant feature confusion problems when processing multi-target scenarios, and have limited generalization ability; existing technologies cannot effectively fuse multi-source signals, lack adaptability to dynamic environments, and have many other shortcomings, making it difficult to meet the needs of complex dynamic scenarios.

[0003] Firstly, traditional device fingerprinting techniques mainly rely on active signal transmission or extraction of specific hardware features (such as MAC addresses, signal strength, etc.). These methods often require additional hardware support or active signal interaction, which not only easily violates privacy, but also significantly increases hardware costs. At the same time, active signal transmission may cause perceptible interference to the monitored objects, which violates the original design of non-sensing monitoring.

[0004] Secondly, existing technologies do not fully utilize spectral features, and usually rely only on a single signal source (such as Wi-Fi or Bluetooth) for feature extraction. The dependency on a single signal source makes the system vulnerable to interference in complex multi-target scenarios, and the feature discrimination ability is insufficient, resulting in poor accuracy and robustness of device recognition.

[0005] In addition, traditional models have feature confusion problems when processing multi-target dynamic scenarios, especially when the number of targets increases or the target features are similar, the recognition performance decreases significantly. At the same time, these models have limited generalization ability, and when applied to different environments or new devices, they often need to be retrained or adjusted, making it difficult to adapt to the dynamic changes of actual scenarios.

[0006] Finally, existing technologies also have significant defects in privacy protection. Some methods need to collect sensitive device information or user behavior features, which may cause privacy leakage risk, especially in public scenarios, where this privacy risk is particularly prominent. SUMMARY

[0007] The application aims to overcome the shortcomings of the prior art and provide a passive sensing and non-sensing detection method based on multi-source wireless signal fusion.

[0008] To achieve the above-mentioned application purposes, the wireless device non-sensing detection method based on multi-source wireless signal fusion of the application comprises the following steps:

[0009] S1: According to the actual needs, determine the number N of wireless device models and the number M of wireless signal types in the application environment. For each type of wireless device, detect each type of wireless signal through the wireless interface, and extract the corresponding feature data as the feature sample s n,m , n = 1, 2, …, N, m = 1, 2, …, M;

[0010] S2: Generate all possible application scenarios for each type of wireless device. For each application scenario, perform Cartesian product combination on the involved feature samples s n,m to obtain the wireless signal fusion sample S n,k of the application scenario, k = 1, 2, …, K n , K n represents the number of application scenarios of the nth type of wireless device.

[0011] S3: According to the actual needs, set up a wireless device detection model, whose input is the wireless signal fusion sample and whose output is the confidence degree of the wireless device for each type. Take the wireless signal fusion sample obtained in step S2 as the input and the corresponding wireless device model as the expected output, train the wireless device detection model, and obtain the trained wireless device control model.

[0012] S4: When wireless device detection of the application environment is needed, detect the existing wireless signal at the current time through the wireless interface of each type of wireless signal, extract the feature data of each type of wireless signal as the feature sample s of the wireless signal, m' = 1, 2, …, M', M' represents the number of currently detected wireless signal types; then obtain all possible combinations of M' types of wireless signal feature samples K' = 2 M′ -1, for each possible combination, perform Cartesian product combination on the feature data of the corresponding wireless signal to obtain the to-be-detected wireless signal fusion sample k' = 1, 2, …, K'; take each to-be-detected wireless signal fusion sample The trained wireless device detection model is inputted to obtain a predicted wireless device model and a corresponding confidence, and a wireless device model corresponding to a maximum confidence is selected as a final detection result.

[0013] The application is based on a passive sensing and non-sensing detection method of multi-source wireless signal fusion, each type of wireless signal of each model of wireless device is detected through a wireless interface, feature data is extracted as a feature sample, a Cartesian product combination of the corresponding feature sample is calculated to obtain a wireless signal fusion sample of each model of wireless device in different application scenarios, the wireless signal fusion sample and the corresponding wireless device model are used as training samples to train a wireless device detection model, when wireless device detection of an application environment is needed, the existing wireless signal at the current time is detected and the feature sample is extracted, the wireless signal fusion sample corresponding to all possible combinations of the current feature sample is generated, the trained wireless device detection model is inputted, and the wireless device model corresponding to the maximum confidence is selected as the final detection result.

[0014] The application has the following beneficial effects:

[0015] 1) The application adopts a passive sensing mode, does not need to actively send signals, avoids interference to target wireless devices in an application scenario, and effectively solves problems of insufficient privacy protection, high hardware cost, feature confusion, and insufficient generalization ability;

[0016] 2) The application obtains wireless signal fusion samples of wireless devices in different application scenarios by combining various types of wireless signal feature samples, and then screens the detection result through the confidence, which can better adapt to the actual engineering application requirements of wireless devices. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a specific embodiment flowchart of the passive sensing and non-sensing detection method of multi-source wireless signal fusion of the application;

[0018] Figure 2 is a structure diagram of the wireless device detection model in the embodiment;

[0019] Figure 3 is a flowchart of the online detection of the application in the embodiment;

[0020] Figure 4 is a flowchart of the offline detection of the application in the embodiment. DETAILED DESCRIPTION

[0021] The specific embodiments of the present application based on the passive sensing and non-sensing detection method of multi-source wireless signal fusion are described below in conjunction with the drawings, so that those skilled in the art can better understand the present application. It needs to be specially reminded that in the following description, when the detailed description of the known functions and designs may weaken the main content of the present application, these descriptions will be omitted here.

[0022] Embodiments

[0023] Figure 1 is a specific embodiment flowchart of the passive sensing and non-sensing detection method of the present application based on multi-source wireless signal fusion. As shown in Figure 1 The passive sensing and non-sensing detection method based on multi-source wireless signal fusion of the present application includes the following steps:

[0024] S101: Collecting wireless signal feature samples:

[0025] According to the actual needs, the number N of wireless device models and the number M of wireless signal types in the application environment are determined. For each type of wireless signal device, the wireless interface is used to detect each type of wireless signal, and the corresponding feature data is extracted to form the feature sample s n,m , n = 1, 2, …, N, m = 1, 2, …, M.

[0026] In this embodiment, the wireless signal includes three types, which are Wi-Fi signal, classic Bluetooth signal and low-power Bluetooth (BLE) signal. The feature data of the Wi-Fi signal is extracted from the Beacon frame, including BSSID (Basic Service Set Identifier), SSID (Service Set Identifier), channel and RSSI (Received Signal Strength Indication); the feature data of the classic Bluetooth signal is extracted from the device broadcast signal, including device name, MAC address, service category, alias, unique service identifier UUID, broadcast period, RSSI; the feature data of the low-power Bluetooth signal is also extracted from the device broadcast signal, including device unique service identifier UUID, MAC address, device name, signal strength, broadcast period, characteristic broadcast, RSSI.

[0027] The feature sample can be stored as a format file. In this embodiment, JSON or Parquet format file is used, and the name of each format file is prefixed with the type of wireless signal and suffixed with the model of wireless device. For example, the format file name of Wi-Fi data is wifi_xxx, the format file name of classic Bluetooth data is bredr_xxx, and the low-power Bluetooth data is stored as BLE_xxx.

[0028] To improve the accuracy of subsequent wireless device non-inductive detection, data augmentation can be performed on the original feature samples to simulate the changes of wireless signals in diversified scenarios and improve the generalization ability of the subsequent wireless device detection model. The specific method of feature sample data augmentation in this embodiment is: a plurality of data augmentation operations are pre-set, and for each feature sample, one or more data augmentation operations are randomly selected for processing. The data augmentation operations include string mutation operation, MAC randomization operation, channel and frequency switching operation, wherein:

[0029] The string mutation operation generates variant data that retains core features and has differences by randomly inserting, deleting, replacing, adjusting case, and field order of the unique fields (such as device name, etc.) in the feature sample, which is used for data augmentation.

[0030] The MAC randomization operation randomizes the Organizationally Unique Identifier (OUI) part of the MAC address, thereby simulating a device randomization scenario.

[0031] The channel and frequency switching operation randomly switches the channel or frequency in the feature sample within a selectable range. This operation can simulate the spectrum interference scenario of multiple Wi-Fi and Bluetooth coexistence in the actual environment, i.e., simulate the automatic frequency hopping scenario in channel congestion, thereby enhancing the robustness of the wireless device detection model to channel congestion.

[0032] S102: Generate wireless signal fusion samples:

[0033] In actual scenarios, a wireless device may turn off part of the signal (such as turning off Wi-Fi or classic Bluetooth, low-power Bluetooth), resulting in incomplete data. To simulate this situation, based on the format file f n,m Generate wireless signal fusion samples to construct fusion feature samples of multiple source wireless signals. The specific method is:

[0034] Generate all possible application scenarios for each model of wireless device. For each application scenario, the wireless signal format text f n,m is combined by Cartesian product to obtain the wireless signal fusion sample S n,k of the application scenario, k = 1, 2, …, K n , K n represents the number of application scenarios in the nth model of wireless device. Generally, the number of application scenarios can be calculated according to all possible combinations of wireless signals, i.e., K n = 2 M -1.

[0035] S103: Construct and train wireless device detection model:

[0036] The wireless device detection model is set according to actual needs, and the input of the wireless device detection model is a wireless signal fusion sample, and the output of the wireless device detection model is the probability of each model of the wireless device. The wireless signal fusion sample obtained in step S102 is taken as the input, and the corresponding wireless device model is taken as the expected output, and the wireless device control model is trained to obtain the trained wireless device detection model.

[0037] In this embodiment, the wireless device detection model adopts a plurality of technology fusion models. Figure 2 FIG. 1 is a structural diagram of the wireless device detection model in this embodiment. As shown in FIG. 1, the wireless device detection model in this embodiment includes a random forest model, a deep neural network, and a fusion module, wherein: Figure 2

[0038] The random forest model is used to perform label prediction on the input wireless signal fusion sample by using a plurality of decision trees, to obtain the probability that the input sample belongs to each wireless device model, to constitute a first probability vector and to send the first probability vector to the fusion module. The random forest model is a commonly used classification model, which takes a decision tree as a basic unit, integrates a large number of decision trees to form a random forest, and then votes the results of each decision tree to obtain the classification result of the random forest model.

[0039] The deep neural network is used to perform label prediction on the input wireless signal fusion sample, to obtain the probability that the input sample belongs to each wireless device model, to constitute a second probability vector and to send the second probability vector to the fusion module. The deep neural network is a commonly used machine learning model, which can automatically extract complex nonlinear features to predict the classification result.

[0040] The fusion module is used to perform weighted average on the first probability vector and the second probability vector, and then to predict a final probability vector according to the average probability vector.

[0041] When training the wireless device detection model, this embodiment uses stratified shuffle split and randomized search (RandomizedSearchCV) to search and optimize the hyperparameters of the random forest classification model, and simultaneously optimizes the model parameters of the deep neural network.

[0042] In addition, before training the wireless device detection model, the wireless signal fusion sample can also be preprocessed to improve the training efficiency and the performance of the wireless device detection model obtained by training. The specific method is as follows: the wireless signal fusion sample is feature cleaned, that is, fields with high missing rate or irrelevant to the target are removed to ensure data quality, and missing values are filled, then the numerical features in the feature data are standardized, and the category features are one-hot encoded or ordinal encoded.

[0043] ​It can be seen that the wireless device detection model proposed in the embodiment can quickly label and train set samples by means of the advantages of the random forest model and the deep neural network, train the model through the preprocessed high-quality wireless signal fusion samples, and realize real-time identification of the specified individual by adjusting the confidence threshold value in combination with the weighted average and stacking integration method.

[0044] S104: Wireless device passive detection

[0045] When wireless device detection needs to be performed on an application environment, the wireless signals existing at the current time are detected through the wireless interfaces of various wireless signals, and the feature data of each wireless signal is extracted as the feature sample of the wireless signal m' = 1, 2,..., M', M' represents the number of currently detected wireless signal types. Then all possible combinations K' = 2 of M' types of wireless signal feature samples are obtained M′ -1, for each possible combination, the feature data of the corresponding wireless signal is combined by Cartesian product to obtain a to-be-detected wireless signal fusion sample Each to-be-detected wireless signal fusion sample is preprocessed The trained wireless device detection model is inputted to obtain a predicted wireless device model and a corresponding confidence, and the wireless device model corresponding to the maximum confidence value is selected as the final detection result. In terms of feature data, the same preprocessing method as when the training sample is obtained can be used for preprocessing.

[0046] It can be seen that the passive detection process of the wireless device in the present application adopts passive perception, and does not generate any interference to the target wireless device, and has concealment and high efficiency. In actual engineering application, the present application can be trained and detected online, or trained and detected offline. Figure 3 is the flowchart of the online detection of the present application in the embodiment. As shown in Figure 3 The specific steps of the online detection of the present application in the embodiment include:

[0047] S301: Load wireless device detection model

[0048] The trained wireless device detection model is loaded.

[0049] S302: Real-time acquisition of wireless signal feature sample

[0050] The wireless signals existing at the current time are detected through the wireless interfaces of various wireless signals, and the to-be-detected wireless signal fusion sample of each possible combination is obtained by the method in step S104.

[0051] S303: Wireless device model prediction

[0052] Each to-be-detected wireless signal fusion sample is input into the loaded wireless device detection model, and a final wireless device model detection result is obtained according to confidence screening.

[0053] Figure 4 is a flowchart of the offline detection of the present application in the embodiment. As shown in the figure, the specific steps of the offline detection of the present application in the embodiment include: Figure 4

[0054] S401: Offline wireless signal collection and storage:

[0055] The wireless signals existing at each time in a preset historical time period are detected through wireless interfaces of various wireless signals, and the feature data of each wireless signal at each time is extracted by using the method in step S104.

[0056] S402: Loading of a wireless device detection model:

[0057] The trained wireless device detection model is loaded.

[0058] S403: Wireless device model prediction:

[0059] The stored wireless signal feature data at each time is read, the to-be-detected wireless signal fusion sample of each possible combination at each time is obtained by using the method in step S104, and then the loaded wireless device detection model is input, and a final wireless device model detection result at the time is obtained according to confidence screening.

[0060] Although the above describes the specific embodiments of the present application in a demonstrative manner, so as to facilitate the understanding of the present application by the person skilled in the art, it should be clear that the present application is not limited to the scope of the specific embodiments, and for the person skilled in the art, all kinds of changes are obvious within the spirit and scope of the present application limited and determined by the appended claims, and all kinds of changes utilizing the concept of the present application are included in the protection.​

Claims

1. A non-intrusive detection method for wireless devices based on multi-source wireless signal fusion, characterized in that, The method comprises the following steps: S1: Determine the number N of wireless device models and the number M of wireless signal types in the application environment according to actual needs, for each model of wireless device, respectively detect each type of wireless signal through the wireless interface, and extract the corresponding feature data as the feature sample s of the type of wireless signal in the model of wireless device n,m , n = 1, 2, …, N, m = 1, 2, …, M; S2: generating all possible application scenarios for each model of wireless device, for each application scenario, the involved feature samples s n,m Performing Cartesian product combination to obtain the wireless signal fusion sample S of the application scenario n,k , k = 1, 2, …, K n , K n Indicates the number of application scenarios in the nth model of wireless device; S3: setting a wireless device detection model according to actual needs, wherein the input of the wireless device detection model is a wireless signal fusion sample, and the output of the wireless device detection model is a confidence degree of each model of a wireless device; training the wireless device detection model by taking the wireless signal fusion sample obtained in step S2 as the input and corresponding wireless device models as the expected output, to obtain a trained wireless device control model; S4: When wireless device detection of application environment is needed, the existing wireless signals at the current time are detected through the wireless interfaces of various wireless signals, and the feature data of each wireless signal is extracted as the feature sample of the wireless signal M' represents the number of currently detected wireless signal types; then all possible combinations K' = 2 of M' types of wireless signal feature samples are obtained M′ -1, for each possible combination, the feature data of the corresponding wireless signal is combined by Cartesian product to obtain a to-be-detected wireless signal fusion sample Each to-be-detected wireless signal fusion sample is input into the trained wireless device detection model to obtain a predicted wireless device model and a corresponding confidence value, and a wireless device model corresponding to a maximum confidence value is selected as a final detection result. The trained wireless device detection model is input to obtain a predicted wireless device model and a corresponding confidence value, and a wireless device model corresponding to a maximum confidence value is selected as a final detection result.

2. The method of claim 1, wherein, The wireless signal in step S1 includes three types, namely, a Wi-Fi signal, a classic Bluetooth signal and a low-power Bluetooth signal. 3.The wireless device non-invasive detection method based on multi-source wireless signal fusion of claim 2, wherein, The characteristic data of the Wi-Fi signal is extracted from a Beacon frame, including a basic service set identifier (BSSID), a service set identifier (SSID), a channel and a received signal strength indicator (RSSI); the characteristic data of the classic Bluetooth signal is extracted from a device broadcast signal, including a device name, a MAC address, a service category, an alias, a unique service identifier (UUID), a broadcast period and an RSSI; and the characteristic data of the low-power Bluetooth signal is extracted from a device broadcast signal, including a device unique service identifier (UUID), a MAC address, a device name, a signal strength, a broadcast period, a characteristic broadcast and an RSSI.

4. The method of claim 1, wherein, In step S1, the characteristic sample can also be subjected to data enhancement, and the specific method is as follows: A plurality of data enhancement operations are set in advance, and for each characteristic sample, one or more data enhancement operations are randomly selected for processing, wherein the data enhancement operations include a string mutation operation, a MAC randomization operation and a channel and frequency switching operation, wherein: The string mutation operation is to randomly insert, delete, replace, adjust the case and field order of the unique field in the characteristic sample; The MAC randomization operation is to randomize the non-organization unique identifier (OUI) part of the MAC address; The channel and frequency switching operation is to randomly switch the channel or frequency in the characteristic sample within a selectable range.

5. The method of claim 1, wherein, The number K of application scenarios in the step S2 n = 2 M - 1.

6. The method of claim 1, wherein, The wireless device detection model in step S3 includes a random forest model, a deep neural network and a fusion module, wherein: The random forest model is used to predict the label of the input wireless signal fusion sample by using a plurality of decision trees, to obtain the probability that the input sample belongs to each wireless device model, to form a first probability vector and to send the first probability vector to the fusion module; The deep neural network is used to predict the label of the input wireless signal fusion sample, to obtain the probability that the input sample belongs to each wireless device model, to form a second probability vector and to send the second probability vector to the fusion module; The fusion module is used to perform weighted averaging on the first probability vector and the second probability vector, and then to predict a final probability vector according to the average probability vector.

Images