Lightweight radio frequency fingerprint identification method for variable-length signal

Abstract

The invention discloses a lightweight radio frequency fingerprint identification method for a variable-length signal, and relates to the technical field of fingerprint identification. Comprising the following steps: S1, acquiring different Internet of Things devices for training, and collecting labeled signal data packets; s2, preprocessing the signal data packet to form a training data set; s3, constructing a depth separable convolutional neural block-SCU module based on depth separable convolution, and designing and constructing an SCU-Net network based on the SCU module; s4, training the SCU-Net network by using the training data set to obtain a preliminary neural network, expanding data by using a mode of manually adding noise and simulating a multi-path effect, and retraining the neural network to obtain a trained convolutional neural network; and S5, obtaining a Lora signal of a to-be-identified device, preprocessing the Lora signal, and inputting the Lora signal into the trained convolutional neural network to obtain an identification result corresponding to the to-be-identified device. According to the invention, the defects of high cost and high equipment requirement of the existing identification scheme can be solved, and the design requirements of low cost and light weight are met.

Description

Technical Field

[0001] The present invention relates to the field of radio frequency fingerprint technology, and in particular to a lightweight radio frequency fingerprint recognition method for variable-length signals. Background Art

[0002] With the rapid development of the Internet of Things, the number of IoT access devices has increased dramatically, and their security has become increasingly important. In order to prevent illegal devices from accessing, protect device privacy data, and allow legitimate users to access the network, we need to authenticate IoT devices when they are accessed. Traditional authentication methods rely on software addresses, such as MAC addresses, which are easily tampered with. Once malicious users obtain access credentials, they can access user privacy data or launch malicious attacks. In addition, most of the IoT access devices are small and medium-sized devices, which are often low-cost and power-constrained. Therefore, there is an urgent need for lightweight and reliable device authentication schemes to ensure the security of IoT access devices.

[0003] At present, the authentication method using inherent physical layer characteristics has shown great advantages in the communication security authentication of the Internet of Things. Compared with traditional high-level authentication technology, physical layer authentication relies on the inherent hardware damage of the device generated during the manufacturing process (called radio frequency fingerprint) for effective authentication. These hardware characteristic differences caused by manufacturing deviations are very small and will not affect the normal communication of the device. Similar to human biometric fingerprints, their uniqueness cannot be easily tampered with. We call this verification method Radio Frequency Fingerprint Identification (RFFI). This is a very promising non-encryption technology. The use of this technology makes it possible for the authentication of IoT devices to no longer be affected by the limited performance of the IoT devices themselves. Therefore, it is particularly suitable for power-constrained and low-cost IoT devices.

[0004] Traditional RFFI methods are highly dependent on the quality of the designed algorithm and require expert experience as support. They have great limitations in actual application scenarios and are not suitable for scenarios with massive IoT devices. In order to break the limitations of traditional RFFI methods, deep learning (DL) methods have been introduced into RFFI scenarios and have been widely used, which has become a hot topic in the field of device authentication in recent years. Compared with traditional RFFI methods, deep learning-based methods have shown greater advantages in data mining and feature extraction. They can simultaneously process the two major modules of feature extraction and device identification, and have stronger robustness and signal characterization capabilities. For example, Qian et al. used a fixed-length preamble as input and introduced the attention mechanism into a convolutional neural network to extract hidden features of the signal. Cai et al. used a deep separable convolutional neural network to realize lightweight RF fingerprint recognition. Zhang et al. used a knowledge-driven multi-scale attention convolutional network to capture higher-level features for authentication. However, most of the current solutions for RF fingerprint authentication using deep learning use fixed-length inputs, especially neural networks such as CNN and DNN, which are specially designed for fixed-length inputs. To this end, Shen et al. designed a radio frequency fingerprint recognition scheme for scenarios with variable signal lengths and proposed a channel-independent spectrum to mitigate the channel influence effect. Although the radio frequency fingerprint recognition schemes using various neural networks mentioned above have achieved good classification and recognition results to a certain extent, a deeper model means that the network model used will have more parameters. This will not only increase the computational overhead, but also put higher requirements on the performance of the device. The use of high-performance receivers will inevitably increase the implementation cost, resulting in a lack of replicability. This requirement does not meet the overall characteristics of large-scale and cheap IoT devices. Summary of the invention

[0005] In view of the shortcomings of the above-mentioned prior art, the present invention provides a lightweight RF fingerprint recognition method for variable-length signals. Aiming at the actual application needs of the Internet of Things, a deep separable convolutional neural network is used to extract RF fingerprint features and authenticate equipment. A RFFI solution for variable-length signal scenarios based on a lightweight separable convolutional unit (SCU) is designed, and a multi-packet inference strategy based on exponential moving average (EMA) smoothing is used to perform data enhancement, enhance the robustness of the network in low signal-to-noise ratio scenarios, and improve recognition performance. The experimental results show that compared with the existing neural network solutions, the proposed solution can reduce the model parameter quantity and calculation time by 50% or more while ensuring accuracy.

[0006] The specific technical solutions are as follows: A lightweight radio frequency fingerprint recognition method for variable-length signals, comprising: S1. Obtain different IoT devices for training and collect labeled signal data packets; S2. Preprocess the signal data packet to form a training data set; S3. Based on the depthwise separable convolution, the SCU module is constructed. Based on the SCU module, the SCU-Net network is designed and constructed. The SCU module input first passes through the depthwise convolution layer, and the RELU activation function and the BN layer are enabled. Then, the current output and the SCU module input are linearly superimposed. Finally, the pointwise convolution layer is input, and the RELU activation function and the BN layer are also enabled. S4. Use the training data set to train the SCU-Net network to obtain a preliminary neural network, expand the data by artificially adding noise and simulating multipath effects, and retrain the neural network to obtain a trained convolutional neural network; S5. Obtain the Lora signal of the device to be identified, preprocess the Lora signal, input the preprocessed result into the trained convolutional neural network, classify the features of each individual data packet, and obtain the predicted probability distribution of the current packet , apply EMA filtering to the prediction results of multiple packets, and recursively calculate the classification probability of each packet Realize multi-packet fusion according to classification probability , select the category with the highest classification probability as the final prediction result, and obtain the recognition result corresponding to the device to be identified.

[0007] Furthermore, in the above scheme, S1 includes collecting labeled signal data packets from multiple different IoT devices for training.

[0008] In the above scheme, the pre-processing step in S2 includes: (1) Synchronize the signal data packets, expressed as: , Where r*[n+k] represents the conjugate of r[n+k] and L is the symbol length of a LoRa symbol, given as: Wherein, the symbol duration T is an integer multiple of the sampling time Ts, so that L is an integer; it is defined that when M[n]>threshold, the signal data packet arrives; (2) The received signal data is estimated by To compensate, given as: , in, ,B is the instantaneous frequency, is the frequency offset, is a complex exponential term used to compensate for the frequency offset of the signal. j is an imaginary unit; (3) The preamble part of the signal data is normalized by dividing it by its root mean square, and the preprocessed signal is represented as r[n]; , in, It is the leading code part of the signal data. represents its root mean square.

[0009] In the above scheme, in S5, the step of preprocessing the Lora signal is the same as the step of preprocessing the signal data packet in S2.

[0010] In the above scheme, further, in S5, The receiver antenna captures the LoRa signal of the device to be identified ; then down-converted to baseband and sampled by an analog-to-digital converter to obtain a digital baseband signal r base [ n ], the digital baseband signal is expressed as: , in is the transmitted baseband signal, T s is the sampling interval, and are the carrier frequencies of the transmitter and receiver respectively, is the CFO between them; to simplify the notation, T s , the digital baseband signal can be rewritten as: .

[0011] The above scheme is further smoothed using the EMA method, and the calculation formula is as follows: , in is the smoothed value at time point t, is the original observation value at time point t, α is the smoothing coefficient, 0<α≤1.

[0012] Furthermore, in the above scheme, the SCU-Net network can be replaced by MobileNet or ShuffleNet.

[0013] Furthermore, in the above scheme, the EMA method can be replaced by sliding average, weighted moving average or time series analysis.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention provides a radio frequency fingerprint recognition method for scenarios with variable signal lengths to solve the shortcomings of high cost and high equipment requirements of existing recognition solutions, achieves low-cost and lightweight design requirements, and meets the overall demand for large-scale and low-cost Internet of Things devices; specifically, a deep separable convolution neural block-SCU module is constructed based on deep separable convolution, and a SCU-Net network is designed and built based on the SCU module; wherein, the SCU module input first passes through the deep convolution layer, and the RELU activation function and the BN layer are enabled; then, the current output and the SCU module input are linearly superimposed; finally, the point-by-point convolution layer is input, and the RELU activation function and the BN layer are also enabled. In this way, while solving the above problems, the feature extraction efficiency and accuracy of the radio frequency signal are improved, that is, the performance of radio frequency fingerprint recognition is improved, and the computational complexity and the number of model parameters are significantly reduced.

[0015] 2. The present invention is based on a lightweight network and is designed based on a deep separable convolutional neural network. The lightweight features of the network design have a significant effect on reducing the number of parameters and improving the calculation speed, making it better suitable for resource-constrained IoT devices.

[0016] 3. The multi-packet inference strategy based on EMA of the present invention: the EMA method is used to analyze multiple received RF data packets to improve the signal stability and recognition reliability under low signal-to-noise ratio conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the steps of the method of the present invention; Figure 2 It is a schematic diagram of the existing RFFI system; Figure 3 It is a schematic diagram of radio frequency fingerprint recognition in a scenario with variable signal length according to the present invention; Figure 4 It is the SCU-Net model based on the SCU module of the present invention; Figure 5 This is a schematic diagram of the standard convolution principle; Figure 6 It is a schematic diagram of the principle of group convolution operation; Figure 7 This is an example of the effect of using a multi-package inference method based on EMA. DETAILED DESCRIPTION

[0019] The following is a further detailed description of the embodiments of the invention in conjunction with the accompanying drawings to make the objectives, technical solutions and technical effects of the invention more clearly presented.

[0020] like Figure 1 As shown, the present invention discloses a lightweight radio frequency fingerprint recognition method for variable-length signals, comprising: A lightweight radio frequency fingerprint recognition method for variable-length signals, comprising: S1. Obtain different IoT devices for training and collect labeled signal data packets; S2. Preprocess the signal data packet to form a training data set; S3. Based on the depthwise separable convolution, the SCU module is constructed. Based on the SCU module, the SCU-Net network is designed and constructed. The SCU module input first passes through the depthwise convolution layer, and the RELU activation function and the BN layer are enabled. Then, the current output and the SCU module input are linearly superimposed. Finally, the pointwise convolution layer is input, and the RELU activation function and the BN layer are also enabled. S4. Use the training data set to train the SCU-Net network to obtain a preliminary neural network, expand the data by artificially adding noise and simulating multipath effects, and retrain the neural network to obtain a trained convolutional neural network; S5. Obtain the Lora signal of the device to be identified, preprocess the Lora signal, input the preprocessed result into the trained convolutional neural network, classify the features of each individual data packet, and obtain the predicted probability distribution of the current packet , apply EMA filtering to the prediction results of multiple packets, and recursively calculate the classification probability of each packet Realize multi-packet fusion according to classification probability , select the category with the highest classification probability as the final prediction result, and obtain the recognition result corresponding to the device to be identified.

[0021] Among them, the above S1 includes collecting labeled signal data packets from multiple different IoT devices for training.

[0022] Here, the preprocessing steps in S2 include: (1) Synchronize the signal data packets, expressed as: , Where r*[n+k] represents the conjugate of r[n+k] and L is the symbol length of a LoRa symbol, given as: Wherein, the symbol duration T is an integer multiple of the sampling time Ts, so that L is an integer; it is defined that when M[n]>threshold, the signal data packet arrives; (2) The received signal data is estimated by To compensate, given as: , in, , B is the instantaneous frequency, is the frequency offset, is a complex exponential term used to compensate for the frequency offset of the signal. j is an imaginary unit; (3) The preamble part of the signal data is normalized by dividing it by its root mean square, and the preprocessed signal is represented as r[n]; , in, It is the leading code part of the signal data. represents its root mean square.

[0023] In the above scheme, the preprocessing operations of S2 and S5 are the same, that is, the signal data packet of the Lora signal is preprocessed in S5, as shown in the above preprocessing steps.

[0024] In S5, the reception of Lora signals includes: The receiver antenna captures the LoRa signal of the device to be identified ; then down-converted to baseband and sampled by an analog-to-digital converter to obtain a digital baseband signal r base [ n ], the digital baseband signal is expressed as: , in is the transmitted baseband signal, T s is the sampling interval, and are the carrier frequencies of the transmitter and receiver respectively, is the CFO between them; to simplify the notation, T s , the digital baseband signal can be rewritten as: .

[0025] The principle process of the present invention is described in detail as follows: 1. RFFI system model As attached Figure 2 As shown in the figure, the existing RFFI system needs to be composed of multiple modules, including LoRa signal collection, data preprocessing, training model, extracting signal features, and using the RFFI protocol to analyze the signal features to finally identify legitimate devices. This includes multiple terminal devices to be classified and a gateway used to receive and forward signals from network servers and terminal devices. The forwarded signals are then analyzed by the trained RFFI system to infer which device the packet signal comes from.

[0026] For a specific IoT device i, its signal x(i) is distorted due to hardware damage and then transmitted through the channel. Although in the design of communication systems, whether it is the transmitter, receiver, or channel, there will be noise that interferes with the signal, in the analysis model, additive noise is considered to be superimposed on the signal before entering the receiver. Figure 2 As shown, after the transmitted signal passes through the channel, that is, x(i) in the figure, the noise n(t) will be superimposed before entering the receiver. Therefore, the signal xi(t) entering the receiver is:

[0027] in represents the specific hardware distortion of device i, s(t) is the multiplicative interference of the channel, n(t) is the additive noise, and * represents the convolution operation. When the receiver captures the packet from device i, it extracts the preamble part xi(t) Then, the RFFI system is used to map xi(t) to the device tag i to achieve the recognition effect, which can be expressed as:

[0028] in Represents the identification process of the RFFI system.

[0029] 2. Signal acquisition The RFFI system for processing LoRa signals requires real LoRa signal data samples. The LoRa signal is first captured by the receiver antenna, i.e. . Then it is down-converted to baseband. The received baseband signal is sampled by an analog-to-digital converter to obtain a digital baseband signal , the digital baseband signal can be mathematically expressed as

[0030] in is the transmitted baseband signal, Ts is the sampling interval, and are the carrier frequencies of the transmitter and receiver respectively, is the CFO between them. To simplify the notation, Ts is omitted. The digital baseband signal can be rewritten as 3. Preprocessing

[0031] The raw signal data cannot be directly input into the RFFI system for processing, so the received signal needs to be preprocessed to meet the basic requirements of RFFI, including synchronization, CFO estimation and compensation, and normalization. These steps need to be applied to the training set and test set data.

[0032] 3.1 Synchronization Accurate signal synchronization can facilitate the positioning of the starting position of the preamble, while inaccurate positioning will be affected by channel noise, which will greatly affect the performance of the RFFI system. In this regard, there are currently a variety of LoRa signal synchronization methods, which can be expressed as:

[0033] Where r*[n+k] represents the conjugate of r[n+k] and L is the symbol length of a LoRa symbol (in samples), given as: The symbol duration T must be an integer multiple of the sampling time Ts so that L is an integer. A threshold is predefined to detect the arrival of a packet: when M[n]>threshold, the packet arrives.

[0034] 3.2 CFO estimation and compensation: Carrier Frequency Offset (CFO) is also one of the RF impairments used in RFFI. CFO is the difference between the transmitter and receiver carrier frequencies caused by the transmitter oscillator. This shift in the carrier frequency results in a small deviation in the received signal and can be used as an identifier for that specific device. Even after compensating for CFO, it has been observed that the accuracy degrades over time due to the hardware characteristics of the device. RFFI essentially requires CFO compensation to guarantee system stability. It has been proven that crystal oscillators are particularly sensitive to temperature changes and the drift of the oscillator frequency can severely degrade system performance.

[0035] Ideal instantaneous frequency of baseband fundamental chirp It increases linearly from −B2 to B2. However, there is an unavoidable frequency offset Δf in the received baseband signal r[n]. Therefore, the instantaneous frequency f[n] becomes:

[0036] Due to the linear nature of f[n], a rough estimate of CFO can be obtained by calculating the average value of f[n] of the received preamble. is given as: The received signal can be estimated by To compensate, given by 3.3 Normalization Normalization prevents the system from learning non-device specific received powers. The preamble part is normalized by dividing by its RMS. The preprocessed signal is denoted as r[n].

[0037]

[0038] 4. Identification principle The RFFI scheme for variable signal length proposed by the present invention includes two main steps: signal training and signal recognition. The structure of the scheme is shown in the attached figure. Figure 3 As shown in the figure. First, a large number of labeled signal data packets are collected from K different training devices, and these data packets are preprocessed to form a training data set. Next, a neural network model that can adapt to input data of different sizes is designed. Finally, the parameters of this neural network are initialized and trained using the above prepared training data set.

[0039] In order to realize the network set by this scheme, an SCU module is designed. In this network module, the traditional depth-separable convolution is split and reconnected by residual connection. That is, the input of the entire module first passes through a deep convolution layer, the activation function is the RELU function, and the BN layer is used to optimize the model convergence; then, the current output is linearly superimposed with the input of the module; finally, the point-by-point convolution layer is input, the activation function is still the RELU function, and the BN layer is still used to optimize the model convergence. Based on the SCU module, a lightweight SCU-Net network is designed for the processed signal samples, which is used as the basic network model for radio frequency fingerprint recognition. The model is shown in the attached figure. Figure 4 shown.

[0040] For the Lora signal training samples of the input model, they first pass through a common two-dimensional convolution layer, which also uses the RELU activation function and the BN layer; then the SCU module designed above. Next, a two-dimensional global average pooling layer (Global Average Pooling, GAP) is used instead of the traditional fully connected layer to further extract and filter the output features of the convolution layer to further reduce the number of parameters and calculations of the model. Traditional CNN generally passes through one or more fully connected layers before the network layer outputs the classification. This is mainly to reduce the feature dimension of the output and purify the features so that they can be output to the final classifier, but its large number of parameters can easily lead to overfitting. Because the pooling operation can also achieve feature dimensionality reduction, pooling can be used instead of full connection. Unlike general average pooling, GAP 2D averages the entire feature map of the input. The GAP layer calculates the global average of each feature map output by the upper layer, and then uses it as the confidence corresponding to each category, and directly inputs it into the classifier for classification. Compared with the fully connected layer, the GAP 2D layer greatly reduces the number of parameters, can better compress the model, and can effectively solve the overfitting problem caused by the fully connected layer.

[0041] 5. Data enhancement strategy In order to improve the classification accuracy, multi-packet inference is used for data enhancement. The multi-packet inference method is a technology used in network traffic analysis and other signal processing fields. It enhances the accuracy of data interpretation by analyzing and processing information from multiple data packets. It aims to extract features from a series of data packets and analyze them in order to identify specific behaviors or properties of data streams or signals. This method usually needs to consider the temporal continuity and changing trends of the data. Combining exponential moving average (EMA) for smoothing is a common method to improve inference performance. In multi-packet inference, all collected packets can be used to calculate EMA, where the information provided by each packet is used to update the calculation of EMA. EMA helps reduce noise and fluctuations, making trends more obvious.

[0042] EMA gives more weight to recent data and is therefore more sensitive to recent changes. The calculation formula is as follows: in is the smoothed value at time point t, is the original observation at time point t, α is the smoothing coefficient, 0<α≤1, this coefficient determines the importance of the new observation relative to the historical value. The larger α is, the more sensitive it is to the most recent observation.

[0043] The effects of the above scheme are further described below: 1. The depth-wise separable convolutional block of the present invention introduces an effect on reducing the number of parameters.

[0044] like Figure 5 As shown, it represents the existing standard convolution operation. If the input feature map size is , the convolution kernel size is , the output feature map size is , we can get the number of parameters of the standard convolutional layer as A filter is applied to the input feature map If you want to output a value in the range of 1, you need A filter.

[0045] The SCU module set for variable-length signals in the present invention is a depth-separable convolution block, which introduces the effect of reducing the number of parameters, thereby achieving the function of improving recognition efficiency and reducing the number of parameters. Figure 6 Represents the grouped convolution operation. Figure 6 Specifically, the input feature map is divided into g groups according to the number of channels, and the size of each group of input feature maps is , the corresponding convolution kernel size is , the size of each group of output feature maps is . Concatenate the g groups of results to get a final size of The output feature map of the grouped convolutional layer is .

[0046] On the feature map output by the conventional convolution, each point is composed of the input feature map points; and on the feature map output by the grouped convolution, each point is calculated by the input feature map Naturally, the number of parameters for grouped convolution is the same as that for standard convolution. .

[0047] Table 1 shows the performance comparison results of different methods. As shown in Table 1, in terms of parameter quantity, the parameter quantity of the SCU-Net method is 266,524, which is much lower than other methods. Fewer parameters means that the SCU-Net model is more lightweight, easier to train and deploy, and reduces the risk of overfitting. In terms of storage size, the storage size of the SCU-Net method is only 0.4MB, which is much smaller than other methods. The extremely small storage requirement enables the SCU-Net method to run efficiently on resource-constrained devices, which is very suitable for embedded systems and mobile devices. From multiple indicators, the SCU-Net method performs well in terms of parameter quantity, storage size, training time and test time, especially in resource-constrained environments. Although its average accuracy is slightly lower than that of LSTM and Transformer, considering all factors, the SCU-Net method has high practical value and efficiency in practical applications. As an efficient, lightweight and easy-to-deploy option, the SCU-Net method is particularly suitable for application scenarios that require high performance and low resource consumption.

[0048] Table 1 Comparison results of complexity performance of different methods method Parameter quantity Storage Space (MB) Training time (mins) Test time(s) Average precision FCN 655594 8.2MB 381.6 28.59 74.19% LSTM 856586 10.3MB 451.2 26.31 76.59% Transformer 348938 10.3MB 350.2 23.52 75.00% GRU 644618 7.8MB 401.5 27.46 73.48% SCU-Net 266524 0.4MB 270.6 22.62 74.51% 2. The present invention has better data enhancement effect.

[0049] The performance comparison of the EMA-based multi-package inference method and the original multi-package inference method is shown in the figure below. Figure 7As shown in the figure. The blue (normal) is the original curve without data enhancement, the red (packet=3(without EMA)) is the solution introduced in the literature [7], and the yellow (packet=3(EMA)) and green (packet=5(EMA)) are the multi-packet inference methods based on EMA that we use. It can be seen from the figure that the use of the multi-packet inference method based on EMA can effectively improve the data classification accuracy under low signal-to-noise ratio conditions (SNR∈[-5,20]). At high SNR values, this improvement is more significant, reaching an accuracy of nearly 90%. And both are higher than the existing methods. However, from the comparison of the yellow and green curves, it can be seen that the increase in the number of data packets can slightly improve the performance, but it does not have a significant advantage. It can be seen from the figure that EMA can effectively improve the data classification accuracy under low SNR conditions. At high SNR values, this improvement is more significant, reaching an accuracy of nearly 90%.

[0050] The above description is only a preferred feasible embodiment of the present invention and is not intended to limit the scope of the patent application of the present invention. All equivalent changes, equivalent substitutions or modified changes performed within the technical spirit and principles suggested by the present invention should be included in the scope of patent protection covered by the present invention.

Claims

1. A lightweight radio frequency fingerprint recognition method for variable-length signals, characterized in that: include: S1. Obtain different IoT devices for training and collect labeled signal data packets; S2. Preprocess the signal data packet to form a training data set; S3. Based on the depthwise separable convolution, the SCU module is constructed. Based on the SCU module, the SCU-Net network is designed and constructed. The SCU module input first passes through the depthwise convolution layer, and the RELU activation function and the BN layer are enabled. Then, the current output and the SCU module input are linearly superimposed. Finally, the pointwise convolution layer is input, and the RELU activation function and the BN layer are also enabled. S4. Use the training data set to train the SCU-Net network to obtain a preliminary neural network, expand the data by artificially adding noise and simulating multipath effects, and retrain the neural network to obtain a trained convolutional neural network; S5. Obtain the Lora signal of the device to be identified, preprocess the Lora signal, input the preprocessed result into the trained convolutional neural network, classify the features of each individual data packet, and obtain the predicted probability distribution of the current packet , apply EMA filtering to the prediction results of multiple packets, and recursively calculate the classification probability of each packet Realize multi-packet fusion according to classification probability , select the category with the highest classification probability as the final prediction result, and obtain the recognition result corresponding to the device to be identified.

2. According to claim 1, a lightweight radio frequency fingerprint recognition method for variable-length signals is characterized in that: The S1 includes collecting labeled signal data packets from multiple different IoT devices for training.

3. The lightweight radio frequency fingerprint recognition method for variable-length signals according to claim 1, characterized in that: The preprocessing step in S2 includes: (1) Synchronize the signal data packets, expressed as: , where r*[n+k] represents the conjugate of r[n+k] and L is the symbol length of a LoRa symbol, given as: , Wherein, the symbol duration T is an integer multiple of the sampling time Ts, so that L is an integer; it is defined that when M[n]>threshold, the signal data packet arrives; (2) The received signal data is estimated by To compensate, given as: , in, , B is the instantaneous frequency, is the frequency offset, is a complex exponential term used to compensate for the frequency offset of the signal. j is an imaginary unit; (3) The preamble part of the signal data is normalized by dividing it by its root mean square, and the preprocessed signal is represented as r[n]; ， in, It is the leading code part of the signal data. represents its root mean square.

4. The lightweight radio frequency fingerprint recognition method for variable-length signals according to claim 3 is characterized in that: In S5, the step of preprocessing the Lora signal is the same as the step of preprocessing the signal data packet in S2.

5. The lightweight radio frequency fingerprint recognition method for variable-length signals according to claim 1, characterized in that: In S5, The receiver antenna captures the LoRa signal of the device to be identified ; then down-converted to baseband and sampled by an analog-to-digital converter to obtain a digital baseband signal r base [ n ], the digital baseband signal is expressed as: , in is the transmitted baseband signal, T s is the sampling interval, and are the carrier frequencies of the transmitter and receiver respectively, is the CFO between them; to simplify the notation, T s , the digital baseband signal can be rewritten as: .

6. The lightweight radio frequency fingerprint recognition method for variable-length signals according to claim 1, characterized in that: Use the EMA method for smoothing, and the calculation formula is as follows: ， in is the smoothed value at time point t, is the original observation value at time point t, α is the smoothing coefficient, 0<α≤1.

7. The lightweight radio frequency fingerprint recognition method for variable-length signals according to claim 1, characterized in that: The SCU-Net network can be replaced by MobileNet or ShuffleNet.

8. The lightweight radio frequency fingerprint recognition method for variable-length signals according to claim 6, characterized in that: The EMA method may be replaced by sliding average, weighted moving average or time series analysis.

Images