Non-cooperative unknown Internet of Things equipment radio frequency sensing method based on feedforward neural network

Abstract

The invention discloses a non-cooperative unknown Internet of Things equipment radio frequency sensing method based on a feedforward neural network. The method focuses on communication signals transmitted by non-cooperative unknown Internet of Things equipment, and since the equipment must transmit the communication signals once transmitting information collected by the equipment to the outside, the detection probability can be effectively improved by capturing the communication signals of the equipment. According to the method, the communication signal features are extracted through the neural network, the features do not need to be obtained from an equipment supplier, and the deployment feasibility and convenience are effectively improved. The feed-forward neural network is one of the simplest neural network structures, the characterization capability is relatively weak, and complex tasks are generally difficult to complete, a specific loss function is adopted in the method, so that the feed-forward neural network obtained through training has the capability of identifying non-cooperative unknown Internet of Things equipment, and therefore the method has the advantages that in similar methods, the non-cooperative unknown Internet of Things equipment can be identified; and the requirement on hardware performance is relatively low, so that the deployment potential of the method is improved, and the deployment cost is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of signal processing technology, and specifically relates to feature processing and pattern recognition of communication signals from unknown devices, and can be used for radio frequency sensing of non-cooperative unknown Internet of Things devices. Background Art

[0002] With the development of IoT technology, IoT devices are widely used in smart homes, industrial production, smart logistics and other fields. The application of IoT technology has greatly facilitated people's lives, empowered traditional industries and created new economic growth points.

[0003] Although the application of IoT technology has brought great value, some criminals have taken advantage of the opportunity to use technology for evil, which has also brought new challenges to the application scenarios of IoT technology. The intrusion of non-cooperative unknown IoT devices is one of the threats that needs to be addressed urgently. For example, intrusion IoT devices in smart homes can easily steal the privacy and living habits of family members; intrusion IoT devices in smart business places can cause the leakage of commercial secrets. Such intrusion IoT devices usually try to disguise themselves and will not actively respond to the perception and communication requests of the detection device, so they are called non-cooperative devices. The existence of such threats has triggered an urgent need for non-cooperative unknown IoT device perception technology.

[0004] For such non-cooperative devices, radio frequency-based sensing is currently one of the most effective means. Because although such devices will not actively respond to the requests of sensing devices, they need to actively transmit communication signals to communicate with the outside world in order to transmit the information they collect to the outside world. Once this communication signal can be sensed and identified, the existence of the intrusion IoT device can be determined. The existing signal detection methods of this type are usually based on the classic hypothesis testing theory to detect the presence or absence of the target signal from the background noise. However, with the development of IoT technology, the IoT systems actually deployed are becoming increasingly complex, and the space is full of electromagnetic waves for communication between various devices. The complex electromagnetic environment makes the existing detection methods gradually unable to cope with the needs, and the detection effect cannot meet the use requirements; at the same time, in order to avoid the legitimate devices in the scene from being accidentally detected as intrusion IoT devices, it is necessary to know the communication protocol or signal characteristics of the legitimate devices from the suppliers of the legitimate devices, which is not realistic in the current situation where there are many types of IoT devices, from different suppliers, and suppliers are generally unwilling to disclose their trade secrets.

[0005] In summary, the introduction of feedforward neural network technology to process the communication signals emitted by non-cooperative unknown IoT devices is an effective way to solve the existing difficulties in detecting intrusion IoT devices. By developing this method, we can effectively deal with the threat of intrusion IoT devices and promote the safe and efficient development of IoT technology and industry.

[0006] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0007] The purpose of the present invention is to propose a non-cooperative unknown IoT device radio frequency sensing method based on a feedforward neural network in view of the shortcomings of the existing non-cooperative unknown IoT device sensing method. By utilizing the robustness of the neural network, it can work in a complex electromagnetic environment; then, by using the neural network to extract the communication signal characteristics emitted by the IoT device, and based on this feature, the legitimacy of the device is judged, thereby achieving the communication protocol or signal characteristics of the legal device without knowing the supplier of the legal device; finally, semantic space segmentation is performed based on the semantic center, thereby achieving efficient detection of non-cooperative unknown intrusion IoT devices, providing technical support for the safe and rapid development of IoT technology.

[0008] The objective of the present invention is to achieve the following technical solution: a non-cooperative unknown IoT device radio frequency sensing method based on a feedforward neural network, comprising the following steps:

[0009] (1) Deployment of communication signal monitoring equipment: Several communication signal monitoring devices are deployed in the scene where the intrusion IoT devices need to be monitored. The monitored RF signal waveform x(t) is expressed as:

[0010] x(t)=s(t)+n(t), t=1, 2,...,T

[0011] Where T represents the time length of the RF signal, s(t) represents the monitored communication signal waveform emitted by the IoT device, and n(t) represents the received noise;

[0012] (2) Feature extraction of communication signals: Different feature extraction strategies are selected according to different scenarios. All features of the extracted communication signals are expressed in vector form as follows:

[0013] X=[x1,x2,…,x G ]

[0014] Among them, x g expresses the g-th feature extracted from x(t), and G expresses the total number of features extracted from x(t);

[0015] (3) Feature compression representation based on feedforward neural network: The communication signal features are compressed and represented using a feedforward neural network. The specific expression is as follows:

[0016] R i =ReLu(W i ×Ri-1 +b i ), i=1,2,…,I

[0017] Among them, ReLu() expresses the nonlinear activation function, R i Represents the output of the i-th layer feedforward neural network, R0 = X, W i expresses the learnable weights of the i-th layer of the feedforward neural network, b i Expresses the learnable bias of the i-th layer of the feedforward neural network, I expresses the number of layers of the feedforward neural network; the output R of the last layer of the feedforward neural network I as a compressed representation of communication signal characteristics;

[0018] (4) Semantic space segmentation based on semantic center: In the training phase, M training data from N legitimate IoT devices are used, and multiple loss functions are introduced to learn the parameters W. i 、b i Perform training; the mth, m=1,2,...,Mth training data X m The output of the feedforward neural network is recorded as The device number to which the mth training data belongs is recorded as n m ∈{1, 2, ..., N}, n m At the same time, as the label of the mth training data, the center vector of the device numbered n in the semantic space is recorded as c n , n=1,2,...,N;

[0019] The specific expression of the loss function is as follows:

[0020]

[0021] in:

[0022]

[0023] k, r are hyperparameters; L1 is the classification loss function, Softmax() expresses the Softmax function, and Classifier() represents a classification neural network; L2 is the reconstruction loss function, and Decoder() represents a decoding neural network; the loss function L3 makes the distance between the representations of all communication signals belonging to the same device in the training set as close as possible in the semantic space, and the distance between the representations of all communication signals belonging to different devices as far as possible, and this distance can be controlled by the hyperparameter r; the loss function L3 simultaneously constrains the difference and absolute distance between the representations of different classes in the semantic space, and the weight between these two constraints can be controlled by the hyperparameter k;

[0024] (5) Device classification and identification of non-cooperative unknown intrusion IoT devices: Collect the radio frequency signals of all legitimate IoT devices in the scene and calculate the semantic center. The calculation formula of the semantic center of the legitimate IoT device numbered n is expressed as:

[0025]

[0026] Among them, δ n () is the judgment function. If X m is the feature extracted from the radio frequency signal emitted by the IoT device numbered n, then δ n (X m )=1, otherwise δ n (X m )=0; After this step, the characteristics of the communication signal emitted by the legitimate IoT device are identified and classified;

[0027] Each time a communication signal waveform x is received from an IoT device u (t), calculate the low-dimensional compressed representation of signal features Use the following formula to determine whether it is an unknown intrusion IoT device:

[0028]

[0029] Among them, distance() expresses the Minkowski distance between vectors, and φ expresses the set distance threshold; when y u When it is equal to -1, the signal waveform x is judged u (t) From an unknown intrusion IoT device, otherwise the signal waveform x u (t) From a known legitimate IoT device, whose number is n = y u .

[0030] Furthermore, the deployment of the communication signal monitoring equipment is specifically as follows: construct a virtual two-dimensional coordinate system in the scenario where the intrusion Internet of Things device needs to be monitored; select any point as the origin of the coordinate system, select any direction as the positive direction of the x-axis, and its perpendicular direction as the positive direction of the y-axis; starting from the origin of the coordinate system, a number of communication signal monitoring devices are evenly arranged along the positive and negative directions of the x and y axes, with the same distance d between them; each communication signal monitoring device can monitor the radio frequency signal waveform within a certain range centered on it, and can save and transmit it to a computer for processing and analysis.

[0031] Furthermore, the feature extraction of the communication signal is specifically: directly selecting the collected radio frequency signal waveform x(t) itself at all times t as the signal feature.

[0032] Furthermore, the feature extraction of the communication signal is specifically as follows: in some scenarios, if the characteristics of legitimate IoT devices and intrusive IoT devices are known, these characteristics are used to extract more efficient features from x(t), including: for devices with a stable bias in the RF signal level, an unbiased estimator of the mean X1 of the first-order statistic of x(t) is calculated as a feature; for devices whose RF signal AC power is significantly different from that of other devices, an unbiased estimator of the variance X2 of the second-order statistic of x(t) is calculated as a feature.

[0033] Furthermore, the feature extraction of the communication signal is specifically as follows: considering the non-cooperative IoT device perception scenario, since most IoT devices adopt the MQTT protocol or comply with the OPC-UA standard, the packet structure and communication timing of the device's communication data packet are particular, and thus the communication signal is also particular, and proprietary features are extracted from them as communication signal features; the expression of the proprietary features is:

[0034]

[0035] Where x represents the calculated proprietary feature, elemax[·] represents the maximum value of all elements in a vector or matrix, ⊙ represents the element-wise multiplication of a vector or matrix, and l(t) is a waveform designed according to the OPC-UA standard or the MQTT protocol.

[0036] Furthermore, the specific design method of l(t) is:

[0037] Take the fixed content specified in the OPC-UA standard or MQTT protocol, process these contents according to the Internet protocol stack to obtain the corresponding baseband signal, and then modulate the baseband signal to obtain the modulated waveform l′(t). Then, in order to make the dimension of l′(t) the same as x(t) and satisfy the operation rule of vector element-wise multiplication, l′(t) is copied into a column vector l′(t). The obtained l′(t) can be used as l(t).

[0038] For legitimate IoT devices, the communication signal is measured directly through the measurement method when the legitimate device transmits a communication signal that complies with the OPC-UA standard or the MQTT protocol, and it is used as l(t).

[0039] Furthermore, the classification loss function makes the input X mm Can be correctly classified to the device it belongs to m , ensuring smooth classification; the reconstruction loss function makes the output of the feedforward neural network can be restored to its input X m , thus ensuring Keep enough information about X m In the loss function L3, It constrains the differences between representations of different classes in the semantic space, making representations of the same class as close together as possible and representations of different classes as far apart as possible; in a high-dimensional space that is difficult for humans to understand and imagine, the distance between representations is It is a constraint method that conforms to the characteristics of the data; It is a direct constraint on the distance between representations in the semantic space, making the distance between similar representations less than r and the distance between heterogeneous representations greater than r; Acting directly on absolute distance is a constraint method that is more in line with human intuition; The function plays a smoothing role, keeping the gradient calculation results in a reasonable range during training.

[0040] Compared with the prior art, the present invention has the following advantages:

[0041] (1) The present invention focuses on the communication signals emitted by non-cooperative unknown IoT devices. Since such devices must emit communication signals once they transmit the information they collect to the outside world, by capturing their communication signals, the probability of detection can be effectively improved. In addition, the present invention extracts the features of its communication signals through a neural network without obtaining such features from the device supplier, which effectively improves the feasibility and convenience of the deployment of the present invention.

[0042] (2) The feedforward neural network is one of the most concise neural network structures, but its representation ability is relatively weak and it is generally difficult to complete complex tasks. The present invention adopts a specific loss function so that the trained feedforward neural network has the ability to identify non-cooperative unknown IoT devices. Therefore, the present invention has the advantage of lower hardware performance requirements among similar methods, thereby improving the deployment potential of the present invention and reducing deployment costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is the overall flow chart of the present invention.

[0044] Figure 2 It is a schematic diagram of the present invention utilizing the proprietary characteristics of the communication signal of the Internet of Things device. DETAILED DESCRIPTION

[0045] The following is the attached Figure 1 , the technical solution of the present invention is further described in detail.

[0046] For the detection of non-cooperative unknown intruding IoT devices, existing methods based on classical hypothesis testing theory face many challenges such as complex electromagnetic environment and unknown communication protocols of legitimate IoT devices, resulting in low detection rate and difficulty in meeting practical application requirements. In practical applications, the computing performance of IoT devices is often poor, making it difficult to run complex neural networks. In order to address the above challenges, the present invention introduces a feedforward neural network and proposes a non-cooperative unknown IoT device radio frequency sensing method based on a feedforward neural network. The implementation steps of the present invention are as follows:

[0047] (1) Deployment of communication signal monitoring equipment. In the scenario where the intrusion of IoT devices needs to be monitored, the sensor array is evenly arranged. Specifically, a virtual two-dimensional coordinate system is constructed in the scenario. Any point is selected as the origin of the coordinate system, any direction is selected as the positive direction of the x-axis, and its perpendicular direction is selected as the positive direction of the y-axis. Starting from the origin of the coordinate system, a number of communication signal monitoring devices are evenly arranged along the positive and negative directions of the x-axis and the y-axis at the same distance d. Among them, each communication signal monitoring device can monitor the RF signal waveform within a certain range centered on it, and can save and transmit it to a computer for processing and analysis. The monitored RF signal waveform can be expressed as:

[0048] x(t)=s(t)+n(t), t=1, 2,...,T

[0049] Where x(t) is a column vector with a dimension equal to the number of communication signal monitoring devices, T expresses the time length of the RF signal, and s(t) represents the monitored communication signal waveform emitted by the IoT device. At different times t, s(t) may come from a legitimate IoT device or an intruding IoT device. When no IoT device is communicating in the scene, s(t) = 0. n(t) is the received noise, such as additive white Gaussian noise.

[0050] (2) Feature extraction of communication signals: Different strategies can be selected according to different scenarios.

[0051] In general scenarios, the collected RF signal waveform x(t) at all times t can be directly selected as the signal feature.

[0052] In some scenarios, if the characteristics of legitimate IoT devices and intrusion IoT devices are known, these characteristics can be used to extract more efficient features from x(t). For example, if the RF signal level of some devices has a stable bias, the unbiased estimator of the first-order statistic mean X1 of x(t) can be calculated; if the AC power of the RF signal of some devices is significantly different from that of other devices, the unbiased estimator of the second-order statistic variance X2 of x(t) can be calculated, and the expression is as follows:

[0053]

[0054] in,[·] .2 The expression replaces all elements of a vector or matrix with their squares in sequence.

[0055] In particular, in the non-cooperative IoT device perception scenario considered by the present invention: since most IoT devices use the MQTT protocol or comply with the OPC-UA standard, the packet structure and communication timing of the device's communication data packet are unique, and thus its communication signal is also unique, from which a class of efficient proprietary features can be extracted. Figure 2 The schematic diagram of the present invention utilizing the proprietary features of the communication signal of the IoT device is shown. Specifically, the communication data packet specified by the MQTT protocol contains fixed content, such as the first byte of the data packet header field expressing the name of the data packet, the data packet protocol name field containing the fixed four letters 'MQTT', and the data packet version field containing a limited number of MQTT version numbers; devices that comply with the OPC-UA standard also have uniform fixed content in their communication signals according to the standard, such as: connection opening paragraph (OpenConnection), greeting paragraph (Hello), and response paragraph (Acknowledge). Based on these fixed contents, the calculation method of the proprietary features is:

[0056]

[0057] Among them, x expresses the calculated proprietary feature, elemax[·] expresses the maximum value of all elements in a vector or matrix, and ⊙ expresses the element-wise multiplication of a vector or matrix. l(t) is a waveform designed according to the OPC-UA standard or MQTT protocol. The specific design method is:

[0058] Take the fixed content specified in the OPC-UA standard or MQTT protocol, process these contents according to the Internet protocol stack such as TCP / IP, and obtain the baseband signal corresponding to these fixed contents. Then, according to the common modulation method of IoT devices, obtain the modulated waveform l′(t). Then, in order to make the dimension of l′(t) the same as x(t) and satisfy the operation rule of vector element-wise multiplication, copy l′(t) into a column vector l′(t) (each element of which is l′(t), and the dimension is the number of communication signal monitoring devices). At this time, the obtained l′(t) can be used as l(t). In particular, for legitimate IoT devices, when the legitimate device transmits a communication signal that complies with the OPC-UA standard or MQTT protocol, the communication signal can be directly measured by measurement and used as l(t).

[0059] In summary, all the features of the communication signal extracted in this step are uniformly expressed in vector form as follows:

[0060] X=[x1,x2,…,x G ]

[0061] Among them, x g Expresses the gth feature extracted from x(t), and G expresses the total number of features extracted from x(t). In X, the proprietary features extracted from the characteristics of IoT devices can effectively improve the detection effect of intrusion IoT devices.

[0062] (3) Feature compression representation based on feedforward neural network. Since the feature dimension extracted in step (2) is high, the representation of the communication signal itself is not concise enough. Therefore, through the compression representation of the features, an efficient representation of the communication signal is extracted to classify the source of the RF signal. The feedforward neural network is used to compress the communication signal features, which is specifically expressed as follows:

[0063] R i =ReLu(W i ×R i-1 +b i ), i=1,2,…,I

[0064] Among them, ReLu() expresses the nonlinear activation function, R i represents the output of the i-th layer feedforward neural network. In particular, R0 = X, W i expresses the learnable weights of the i-th layer of the feedforward neural network, b i It expresses the learnable bias of the i-th layer of the feedforward neural network, and I expresses the number of layers of the feedforward neural network. The output R of the last layer of the feedforward neural network is I As a compressed representation of the communication signal features. By inputting X into a feedforward neural network, the features in its communication signal can be more refined. At the same time, through this method, the features of the communication signal emitted by the legitimate IoT device can also be extracted, so that the legitimate IoT device can be identified without obtaining relevant information about the communication signal of the legitimate IoT device from the manufacturer, thereby avoiding misidentification of the legitimate IoT device as an intrusion IoT device.

[0065] (4) Semantic space segmentation based on semantic center. In order to make R an efficient representation of communication signal features, in the training phase, M training data from N legitimate IoT devices are used to introduce multiple loss functions to learn the parameters W. i 、b i Perform training. Set the mth, m=1, 2, ..., Mth training data X m The output of the feedforward neural network is recorded as The device number to which the mth training data belongs is recorded as n m ∈{1,2,...,N},n mAt the same time, as the label of the mth training data, the center vector of the device numbered n in the semantic space is recorded as c n , n = 1, 2, ..., N. The introduced loss function is specifically expressed as follows:

[0066]

[0067] in:

[0068]

[0069] k, r are hyper parameters.

[0070] L1 is called the "classification loss function", Softmax() expresses the Softmax function, and Classifier() represents a classification neural network. The classification loss function makes the input X m Can be correctly classified to the device it belongs to m , to ensure the smooth progress of classification.

[0071] L2 is called the "reconstruction loss function", Decoder() represents a decoding neural network, and the reconstruction loss function makes the output of the feedforward neural network can be restored to its input X m , thus ensuring Keep enough information about X m feature information.

[0072] For L3, in general, L3 plays the following role: in the semantic space, the distance between the representations of all communication signals belonging to the same device in the training set is as close as possible, and the distance between the representations of all communication signals belonging to different devices is as far as possible, and this distance can be controlled by the hyperparameter r. Compared with the existing technology, the loss function L3 constrains the difference and absolute distance between the representations of different classes in the semantic space at the same time, taking into account the advantages of both, and the weight between these two constraints can be controlled by the hyperparameter k; for each part of L3 specifically: It constrains the differences between representations of different classes in the semantic space, making representations of the same class as close together as possible and representations of different classes as far apart as possible. In a high-dimensional space that is difficult for humans to understand and imagine, the distance between representations is It is a constraint method that conforms to the characteristics of the data; It is a direct constraint on the distance between representations in the semantic space, making the distance between similar representations less than r and the distance between heterogeneous representations greater than r. Directly acting on the absolute distance is a constraint method that is more in line with human intuition; in addition, The function plays a smoothing role, keeping the gradient calculation results in a reasonable range during training.

[0073] By using the above loss function to train the learnable parameters W in the feedforward neural network i 、b i , so that the output of the feedforward neural network Keep enough information about the input X m feature information, and The dimension is significantly lower than the input X m ,therefore, It can be used as a low-dimensional compressed representation of signal features. At the same time, due to the low-dimensional compressed representation of each type of signal All gathered in their category center c n Nearby, the semantic space segmentation based on the semantic center is completed.

[0074] (5) Device classification and identification of non-cooperative unknown intrusion IoT devices. First, collect the RF signals of all legitimate IoT devices in the scene and calculate their semantic centers. Specifically, the calculation formula for the semantic center of the legitimate IoT device numbered n is expressed as:

[0075]

[0076] Among them, δ n () is the judgment function. If X m is the feature extracted from the radio frequency signal emitted by the IoT device numbered n, then δ n (X m )=1, otherwise δ n (X m )=0. After this step, the characteristics of the communication signal emitted by the legitimate IoT device are identified and classified, so there is no need to obtain the communication signal characteristics of the legitimate IoT device from the device manufacturer. On this basis, each time a communication signal waveform x is received from the IoT device u (t), by following the above steps to calculate the low-dimensional compressed representation of its signal features The formula can be used to determine whether it is an unknown intrusion IoT device. The specific expression of the formula is:

[0077]

[0078] Where distance() expresses the Minkowski distance between vectors, and φ expresses the set distance threshold. For the setting of this threshold, y can be calculated based on the training data. u The value of is then used as a reference and set in combination with experience. u When it is equal to -1, the signal waveform x is judged u (t) From unknown intrusion IoT devices. At other times, the signal waveform x u(t) From a known legitimate IoT device, whose number is n = y u . Thus, RF perception of non-cooperative unknown IoT devices is achieved.

[0079] The above is only a preferred embodiment of the present invention. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Any technician familiar with the art can make many possible changes and modifications to the technical solution of the present invention by using the above disclosed methods and technical contents without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A non-cooperative unknown IoT device radio frequency sensing method based on a feedforward neural network, characterized in that: The following steps are involved: (1) Deployment of communication signal monitoring equipment: Several communication signal monitoring devices are deployed in the scene where the intrusion IoT devices need to be monitored. The monitored RF signal waveform x(t) is expressed as: x(t)=s(t)+n(t),t=1,2,…,T Where T represents the time length of the RF signal, s(t) represents the monitored communication signal waveform emitted by the IoT device, and n(t) represents the received noise; (2) Feature extraction of communication signals: Different feature extraction strategies are selected according to different scenarios. All features of the extracted communication signals are expressed in vector form as follows: X=[x1,x2,…,x G ] Among them, x g expresses the g-th feature extracted from x(t), and G expresses the total number of features extracted from x(t); (3) Feature compression representation based on feedforward neural network: The communication signal features are compressed and represented using a feedforward neural network. The specific expression is as follows: R i =ReLu(W i ×R i-1 +b i ),i=1,2,…,I Among them, ReLu() expresses the nonlinear activation function, R i Represents the output of the i-th layer feedforward neural network, R0 = X, W i expresses the learnable weights of the i-th layer of the feedforward neural network, b i expresses the learnable bias of the i-th layer of the feedforward neural network, I expresses the number of layers of the feedforward neural network; the output r of the last layer of the feedforward neural network I as a compressed representation of communication signal characteristics; (4) Semantic space segmentation based on semantic center: In the training phase, M training data from N legitimate IoT devices are used, and multiple loss functions are introduced to learn the parameters W. i , b i Perform training; the mth, m=1, 2, ..., Mth training data X m The output of the feedforward neural network is recorded as The device number to which the mth training data belongs is recorded as n m ∈{1,2,…,N},n m At the same time, as the label of the mth training data, the center vector of the device numbered n in the semantic space is recorded as c n ,n=1,2,…,N; The specific expression of the loss function is as follows: in: k, r are hyperparameters; L1 is the classification loss function, Softmax() expresses the Softmax function, and Classifier() represents a classification neural network; L2 is the reconstruction loss function, and Decoder() represents a decoding neural network; the loss function L3 makes the distance between the representations of all communication signals belonging to the same device in the training set as close as possible in the semantic space, and the distance between the representations of all communication signals belonging to different devices as far as possible, and this distance can be controlled by the hyperparameter r; the loss function L3 simultaneously constrains the difference and absolute distance between the representations of different classes in the semantic space, and the weight between these two constraints can be controlled by the hyperparameter k; (5) Device classification and identification of non-cooperative unknown intrusion IoT devices: Collect the radio frequency signals of all legitimate IoT devices in the scene and calculate the semantic center. The calculation formula of the semantic center of the legitimate IoT device numbered n is expressed as: Among them, δ n () is the judgment function. If X m is the feature extracted from the radio frequency signal emitted by the IoT device numbered n, then δ n (X m )=1, otherwise δ n (X m )=0; After this step, the characteristics of the communication signal emitted by the legitimate IoT device are identified and classified; Each time a communication signal waveform x is received from an IoT device u (t), calculate the low-dimensional compressed representation of signal features Use the following formula to determine whether it is an unknown intrusion IoT device: Among them, distance() expresses the Minkowski distance between vectors, and φ expresses the set distance threshold; when y u When it is equal to -1, the signal waveform x is judged u (t) From an unknown intrusion IoT device, otherwise the signal waveform x u (t) From a known legitimate IoT device, whose number is n = y u .

2. According to the feedforward neural network-based non-cooperative unknown Internet of Things device radio frequency sensing method according to claim 1, it is characterized in that: The deployment of the communication signal monitoring equipment is specifically as follows: construct a virtual two-dimensional coordinate system in the scene where the intrusion of the Internet of Things device needs to be monitored; select any point as the origin of the coordinate system, select any direction as the positive direction of the x-axis, and its perpendicular direction as the positive direction of the y-axis; starting from the origin of the coordinate system, along the positive and negative directions of the x-axis and y-axis, a number of communication signal monitoring devices are evenly arranged at the same distance d; each communication signal monitoring device can monitor the radio frequency signal waveform within a certain range centered on it, and can save and transmit it to a computer for processing and analysis.

3. The non-cooperative unknown IoT device radio frequency sensing method based on a feedforward neural network according to claim 1 is characterized in that: The feature extraction of the communication signal is specifically: directly selecting the collected radio frequency signal waveform x(t) itself at all times t as the signal feature.

4. The non-cooperative unknown IoT device radio frequency sensing method based on a feedforward neural network according to claim 1 is characterized in that: The feature extraction of the communication signal is specifically as follows: in some scenarios, if the characteristics of legitimate IoT devices and intrusive IoT devices are known, these characteristics are used to extract more efficient features from x(t), including: for devices with a stable bias in the RF signal level, an unbiased estimator of the mean X1 of the first-order statistic of x(t) is calculated as a feature; for devices whose RF signal AC power is significantly different from that of other devices, an unbiased estimator of the variance X2 of the second-order statistic of x(t) is calculated as a feature.

5. The non-cooperative unknown IoT device radio frequency sensing method based on a feedforward neural network according to claim 1 is characterized in that: The feature extraction of the communication signal is specifically as follows: considering the non-cooperative IoT device perception scenario, since most IoT devices use the MQTT protocol or comply with the OPC-UA standard, the packet structure and communication timing of the device's communication data packet are particular, and thus the communication signal is also particular, from which proprietary features are extracted as communication signal features; the expression of the proprietary feature is: Where x represents the calculated proprietary feature, elemax[·] represents the maximum value of all elements in a vector or matrix, ⊙ represents the element-wise multiplication of a vector or matrix, and l(t) is a waveform designed according to the OPC-UA standard or the MQTT protocol.

6. The non-cooperative unknown IoT device radio frequency sensing method based on a feedforward neural network according to claim 5 is characterized in that: The specific design method of l(t) is: Take the fixed content specified in the OPC-UA standard or MQTT protocol, process these contents according to the Internet protocol stack to obtain the corresponding baseband signal, and then modulate the baseband signal to obtain the modulated waveform l′(t). Then, in order to make the dimension of l′(t) the same as x(t) and satisfy the operation rule of vector element-wise multiplication, l′(t) is copied into a column vector l′(t). The obtained l′(t) can be used as l(t). For legitimate IoT devices, the communication signal is measured directly through the measurement method when the legitimate device transmits a communication signal that complies with the OPC-UA standard or the MQTT protocol, and it is used as l(t).

7. The non-cooperative unknown IoT device radio frequency sensing method based on a feedforward neural network according to claim 1 is characterized in that: The classification loss function makes the input X m Can be correctly classified to the device it belongs to m , ensuring smooth classification; the reconstruction loss function makes the output of the feedforward neural network can be restored to its input X m , thus ensuring Keep enough information about X m feature information.

8. The non-cooperative unknown IoT device radio frequency sensing method based on a feedforward neural network according to claim 1 is characterized in that: In the loss function L3, It constrains the differences between representations of different classes in the semantic space, making representations of the same class as close together as possible and representations of different classes as far apart as possible; in a high-dimensional space that is difficult for humans to understand and imagine, the distance between representations is It is a constraint method that conforms to the characteristics of the data; It is a direct constraint on the distance between representations in the semantic space, making the distance between similar representations less than r and the distance between heterogeneous representations greater than r; Acting directly on absolute distance is a constraint method that is more in line with human intuition; The function plays a smoothing role, keeping the gradient calculation results in a reasonable range during training.

Images