Conformer-based radiation source identification method and system

Abstract

The invention provides a Conformer-based radiation source identification method and system, and the method comprises the steps: collecting a radio frequency signal of a radiation source, and obtaining an in-phase part of an original signal as an input signal; the method comprises the following steps: carrying out Patch segmentation and Embedding on an original in-phase signal; further extracting signal features by using the WLC block and the SWLC block to output a feature map; using a full connection layer as a classifier, and outputting the probability that the current signal belongs to each transmitter; and realizing a defense mechanism by using inverse Sigmoid, disturbing the probability value of the model, manipulating the perception of an attacker on a loss surface, and hindering the attacker from extracting the model. The security of a wireless equipment authentication system in scenes such as Internet of Things and marine traffic is ensured, and invasion of illegal equipment is prevented. In combination with sliding window linear attention and a convolution mechanism, local and global features in a radiation source signal are effectively modeled.

Description

Technical Field

[0001] The present invention belongs to the technical field of radiation source identification, and specifically relates to a Conformer-based radiation source identification method and system, and is particularly suitable for a wireless device authentication system in scenarios such as the Internet of Things and maritime traffic. Background Art

[0002] With the widespread application of the new generation of WI-FI technology, the speed and reliability of data transmission have been greatly improved, which has promoted the development and popularization of various IoT applications. However, the inherent broadcast characteristics of these wireless networks make real-time data easy to be forged, which may lead to the leakage of sensitive information. In order to solve these security issues, various device authentication methods have been developed, including key-based authentication protocols and static device identifiers such as MAC and IP addresses. However, these methods often have limitations in solving these security problems. For example, although key authentication protocols provide mutual verification, the leakage of a single key may endanger the security of the entire system. In addition, due to the low cost and wide distribution of many wireless devices, key distribution and management have become extremely difficult. On the other hand, software addresses are usually not encrypted and are therefore vulnerable to spoofing attacks, where attackers can spoof devices by copying the device's identity information.

[0003] As a critical system for maritime safety, the Automatic Identification System (AIS) is another example of a wireless authentication system that is susceptible to manipulation. While AIS does not rely on traditional authentication methods such as MAC and IP addresses, it uses a broadcast mechanism to transmit key data such as ship name, position and speed. This makes AIS vulnerable to spoofing attacks, where attackers can tamper with this information to cover up illegal activities such as illegal fishing.

[0004] As a hardware-level authentication system, source identification (SEI) technology provides another effective solution to these security issues in wireless network communication scenarios. It uses the radio frequency fingerprint (RFF) in the RF circuit to distinguish different radiation sources. These RF fingerprints mainly come from the inherent manufacturing tolerances and non-idealities of the electronic components inside the RF circuit, such as filter distortion, oscillator frequency and phase noise, I / Q imbalance in the mixer, and nonlinear distortion of the power amplifier. Currently, SEI has been widely used in military and civilian communications, and is expected to enhance communication security by detecting and mitigating the threat of illegal radiation sources.

[0005] Although traditional SEI systems have demonstrated some promising results, they often face challenges in resource-constrained IoT applications. These challenges stem from two important limitations: (1) High computational cost: Extracting features from long, high-resolution signals is computationally expensive for devices with limited processing power. (2) Lack of generalizability: Traditional SEI systems often rely on manually designed scene-specific features, limiting their applicability in different communication scenarios. In addition, traditional SEI systems focus mainly on feature extraction and classification, often ignoring the defense of the model itself. Recent studies have shown that deep learning models are vulnerable to theft or inversion attacks, allowing attackers to infer sensitive information or even reconstruct the data used to train the model.

[0006] In view of this, it is very meaningful to propose a radiation source identification method and system based on Conformer. Summary of the invention

[0007] The present invention provides a Conformer-based radiation source identification method and system to ensure the security of wireless device authentication systems in scenarios such as the Internet of Things and maritime traffic, and prevent the intrusion of illegal devices. Combining sliding window linear attention and convolution mechanism, local and global features in radiation source signals are effectively modeled. This combination enables the method to achieve powerful feature extraction capabilities without relying on artificial features based on specific signal modulation patterns or environmental information, making the method particularly suitable for communication scenarios with different protocols and environments. In addition, in response to the problem that the SEI system may be attacked by theft or inversion, the inverse sigmoid function is used to introduce controlled perturbations in the final activation layer to maximize the loss value of the stolen model to limit the attacker's ability to steal the model.

[0008] In a first aspect, the present invention proposes a radiation source identification method based on Conformer, the method comprising the following steps:

[0009] Collect the radio frequency signal of the radiation source and obtain the in-phase part of the original signal as the input signal;

[0010] Perform patch segmentation and embedding on the original in-phase signal, divide the input signal into patches through a one-dimensional convolution operation with preset parameters, and project it into a preset dimensional space to obtain an embedding vector, and then use Dropout to regularize the feature vector;

[0011] Further extract signal features using WLC blocks and SWLC blocks, which include feedforward networks, normal window linear attention WLA or sliding window linear attention SWLA, point convolution and downsampling convolution operations to output feature maps;

[0012] Use the fully connected layer as a classifier to output the probability that the current signal belongs to each transmitter; and

[0013] The inverse sigmoid is used to implement the defense mechanism. By perturbing the model probability value, the attacker's perception of the loss surface is manipulated, thus hindering the attacker from extracting the model.

[0014] Preferably, when performing patch segmentation and embedding on the original in-phase signal, the input original in-phase signal S=[s 1 ,s 2 ,…,s k ] 1×k , whose length is k, and the patch token with a patch size of p is defined as Where n = k / p, and the one-dimensional convolution with a kernel size of p slides on the input signal with a step size of p to obtain the embedding vector E = [Conv1d(S)] T n×d , where d is the Embedding dimension.

[0015] Further preferably, the further extracting signal features using the WLC block and the SWLC block includes: given an input vector X of the WLC block and / or the SWLC block, the calculation process of the output feature map F is as follows:

[0016]

[0017] F 2 =F 1 +WLA / SWLA(F 1 )

[0018] F 3 =PC(F 2 )+DC(F 2 )

[0019]

[0020] Among them, FN 1 、FN 2 is a feedforward network, WLA is a normal window linear attention, SWLA is a sliding window linear attention, LN is layer normalization, and PC is used The expansion factor is used to project the number of channels of the point convolution, D F is the output dimension of the block, D X is the input dimension of the block, and DC is the downsampling convolution used to reduce the length of the feature map sequence.

[0021] Further preferably, the calculation process of the downsampling convolution DC is expressed as follows:

[0022] Y 1 =GLU(PC 1 (LN(F 2 )))

[0023] Y 2 =Swish(BN(SDC(Y 1 )))

[0024] DC 输出 =Drop(PC 2 (Y 2 ))

[0025] Among them, GLU stands for gated linear unit; Swish and Drop refer to Swish activation function and Dropout respectively; BN is batch normalization, and SDC stands for strided depthwise convolution.

[0026] Preferably, the calculation steps of the common window linear attention WLA are:

[0027] The input token is normalized and split into non-overlapping windows. The self-attention mechanism is performed in each window. After the attention mechanism, a pre-normalized residual unit with Dropout is used to help the model converge. The self-attention mechanism decouples the Softmax operation through an approximate function and calculates To change the order of calculation:

[0028] O h =φ(Q h )Q(K h +B h ) T V h +DWC(V h )

[0029] Among them, φ(·) is the mapping function, represents position encoding, and DWC(·) is a deep convolutional module.

[0030] Further preferably, the calculation steps of the sliding window linear attention SWLA are:

[0031] After the input token is normalized and split into non-overlapping windows, the token is cyclically slid by half a window, and the self-attention mechanism is executed in each window. After the attention mechanism, a pre-normalized residual unit with Dropout is used to help the model converge.

[0032] Preferably, the classifier comprises a one-dimensional adaptive average pooling layer, a dimensional flattening layer, and a fully connected layer of preset input dimension and output dimension, which are connected in sequence.

[0033] Preferably, the use of inverse Sigmoid to implement a defense mechanism, by perturbing the model probability value, manipulating the attacker's perception of the loss surface, and hindering the attacker from extracting the model includes:

[0034] Defense mechanism model probability value Y = {y i} is disturbed to obtain the model probability value after adding noise disturbance: The disturbance mode is where y i represents the probability value of the i-th radiation source, is a disturbance with controllable amplitude, α i is the normalization parameter: Using the perturbation with inverse sigmoid, the perturbation function is almost irreversible: r(y i )≈β(s -1 (y i )-1 / 2), where β is a positive amplitude adjustment parameter, s -1 (·) is the inverse sigmoid function: s -1 (y i )=s(γg(y i )), where s(·) is the sigmoid function, γ is a positive dataset and model-specific convergence parameter, and g(y i ) is y i Pseudo-logical value of:

[0035] In a second aspect, an embodiment of the present invention provides a radiation source identification system based on Conformer, including:

[0036] A signal acquisition module is configured to acquire a radio frequency signal from a radiation source and obtain an in-phase portion of the original signal as an input signal;

[0037] A preprocessing module is configured to perform patch segmentation and embedding on the original in-phase signal, divide the input signal into patches through a one-dimensional convolution operation with preset parameters, and project it into a preset dimensional space to obtain an embedding vector, and then use Dropout to regularize the feature vector;

[0038] A feature extraction module configured to further extract signal features using a WLC block and a SWLC block, including a feed-forward network, a normal window linear attention WLA or a sliding window linear attention SWLA, a point convolution, and a down-sampling convolution operation to output a feature map;

[0039] A classifier module configured to use the fully connected layer as a classifier and output a probability that the current signal belongs to each transmitter;

[0040] The defense module is configured to implement a defense mechanism using inverse sigmoid, which perturbs the model probability value to manipulate the attacker's perception of the loss surface and hinder the attacker from extracting the model.

[0041] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any implementation manner in the first aspect.

[0042] Compared with the prior art, the beneficial results of the present invention are:

[0043] The purpose of the present invention is to provide a Conformer-based radiation source identification method to ensure the security of wireless device authentication systems in scenarios such as the Internet of Things and maritime traffic, and prevent the intrusion of illegal devices. Combined with sliding window linear attention and convolution mechanism, local and global features in radiation source signals are effectively modeled. This combination enables the method to achieve powerful feature extraction capabilities without relying on artificial features based on specific signal modulation patterns or environmental information, making the method particularly suitable for communication scenarios with different protocols and environments. In addition, in response to the problem that the SEI system may be attacked by theft or inversion, the inverse sigmoid function is used to introduce controlled perturbations in the final activation layer to maximize the loss value of the stolen model to limit the attacker's ability to steal the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated into and constitute a part of this specification. The accompanying drawings illustrate the embodiments and are used together with the description to explain the principles of the present invention. It will be easy to recognize other embodiments and many expected advantages of the embodiments because they become better understood by reference to the following detailed description. The elements of the drawings are not necessarily to scale with each other. The same reference numerals refer to corresponding similar parts.

[0045] Figure 1 Schematic diagram of the overall process of the radiation source identification method based on Conformer according to an embodiment of the present invention;

[0046] Figure 2 is a flow chart of a method for identifying radiation sources based on Conformer according to an embodiment of the present invention;

[0047] Figure 3 A structural diagram of a normal window linear attention WLA and a sliding window linear attention SWLA provided for an embodiment of the present invention;

[0048] Figure 4The accuracy of the test set samples in different training rounds during the training process of the Conformer-based radiation source identification method provided by the embodiment of the present invention;

[0049] Figure 5 An experiment comparing the accuracy of stolen models with and without a defense layer when the attacker has different proportions of the original training data set provided by the embodiment of the present invention;

[0050] Figure 6 Schematic diagram of a Conformer-based radiation source identification system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0052] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0053] First, Figure 1 It is shown that the embodiment of the present invention discloses a radiation source identification method based on Conformer, such as Figure 1 As shown, the method comprises the following steps:

[0054] S1. Perform patch segmentation and embedding on the original in-phase signal. The input signal is divided into patches through a one-dimensional convolution operation with preset parameters, and projected into a preset dimensional space to obtain an embedding vector. Then, Dropout is used to regularize the feature vector.

[0055] S2, further extracting signal features using the WLC block and the SWLC block, which include a feedforward network, a normal window linear attention WLA or a sliding window linear attention SWLA, a point convolution, and a downsampling convolution operation to output a feature map;

[0056] S3, using the fully connected layer as a classifier, outputting the probability that the current signal belongs to each transmitter; and

[0057] S4. Use inverse sigmoid to implement a defense mechanism. By perturbing the model probability value, it manipulates the attacker's perception of the loss surface and hinders the attacker from extracting the model.

[0058] Specifically, Figure 2 As shown, in a specific embodiment, the method includes the following steps:

[0059] (1) Collect the RF signal from the communication radiation source transmitter and obtain the in-phase part of the original signal:

[0060] use The EB500 monitoring receiver collects AIS signals transmitted by 100 ships at the **Ocean Management Public Pier to construct a steady-state signal. The sampling rate is set to 48KHZ, so 1 bit contains 48000 / 9600=5 sampling points, so a complete AIS signal segment contains 256×5=1280 sampling points. Autocorrelation detection is performed on the received signal to intercept possible signal segments, and GMSK demodulation and NRZI decoding are performed to intercept valid signal segments. 260 AIS signals are collected for each ship, of which 80 samples are used as training sets and the remaining 180 samples are used as test sets. Each AIS signal uses 689-1200 sampling points as the steady-state part to construct a steady-state data set for testing.

[0061] (2) Patch segmentation and embedding of the original in-phase signal:

[0062] Let S = [s 1 ,s 2 ,…,s k ] 1×k Representing the original in-phase signal of the input, its length is k, and the patch token with patch size p can be defined as:

[0063]

[0064] Where n = k / p. Patch segmentation and embedding are achieved by applying a one-dimensional convolution with a kernel size of p, which slides over the input signal with a step size of p. This extracts a patch of size p and projects it into a d-dimensional space. The resulting embedding vector E is as follows:

[0065] E = [Conv1d(S)] T n×d

[0066] Where d is the Embedding dimension.

[0067] In this example, the original in-phase signal length k = 512, the patch size p = 4, and the Embedding dimension d = 96;

[0068] After embedding, Dropout with a zero probability of 0.1 is used for regularization to enhance the robustness of the model.

[0069] (3) Use WLC and SWLC blocks to further extract signal features: In theory, given the input vector X of the WLC / SWLC block, the output feature map F is calculated as follows:

[0070]

[0071] F 2 =F 1 +WLA / SWLA(F 1 )

[0072] F 3 =PC(F 2 )+DC(F 2 )

[0073]

[0074] where FN 1 、FN 2 is a feedforward network, WLA is a normal window linear attention, SWLA is a sliding window linear attention, LN is layer normalization, and PC is used The expansion factor is used to project the point convolution of the number of channels, where D F is the output dimension of the block, D X is the input dimension of the block, and DC is the downsampling convolution used to reduce the length of the feature map sequence.

[0075] In this example, FN 1 The feed-forward network includes the following connected in sequence: layer normalization module, linear layer with input dimension of 96 and output dimension of 384, Swish activation function, Dropout with zero probability of 0.1, linear layer with input dimension of 384 and output dimension of 96, Dropout with zero probability of 0.1; FN 2 The feedforward network includes the following connected in sequence: layer normalization module, linear layer with input dimension of 256 and output dimension of 1024, Swish activation function, Dropout with zero probability of 0.1, linear layer with input dimension of 1024 and output dimension of 256, Dropout with zero probability of 0.1; point convolution PC is implemented using one-dimensional convolution with input dimension of 96, output dimension of 256, kernel size of 1, and stride of 2.

[0076] (4) For the downsampling convolution DC used to reduce the length of the feature map sequence in step (3), in a possible implementation, its calculation process is expressed as follows:

[0077] Y 1 =GLU(PC 1 (LN(F 2 )))

[0078] Y 2 =Swish(BN(SDC(Y 1 )))

[0079] DC 输出 =Drop(PC 2 (Y 2 ))

[0080] Among them, GLU stands for gated linear unit, which is used to enhance the extraction of feature dependencies in the sequence. Swish and Drop refer to the Swish activation function and Dropout, respectively, which are used to help regularize the network. BN is batch normalization, and SDC stands for strided depthwise convolution, which downsamples the length of the feature map.

[0081] In this example, the point convolution PC 1 It is implemented using a one-dimensional convolution with an input dimension of 96, an output dimension of 512, a kernel size of 1, and a stride of 1; the strided depth convolution SDC is implemented using a one-dimensional convolution with an input dimension of 256, an output dimension of 256, a kernel size of 15, and a stride of 2; the point convolution PC 2 It is implemented using a one-dimensional convolution with an input dimension of 256, an output dimension of 256, a kernel size of 1, and a stride of 1; the Dropout zero probability is set to 0.1.

[0082] (5) Figure 3 As shown in (a), for the ordinary window linear attention WLA described in step (3), in a possible implementation, the calculation steps are as follows: the input token is normalized and split into non-overlapping windows, and the self-attention mechanism is executed in each window instead of the entire input token. After the attention mechanism, a pre-normalized residual unit with Dropout is used to help the model converge.

[0083] In this example, the window size is set to 32, the number of attention heads is 4, and the Dropout zero probability is set to 0.1.

[0084] (6) For the self-attention mechanism described in step (5), in one possible implementation, it is a focused linear attention mechanism that decouples the Softmax operation through an approximate function and first calculates To change the order of calculation:

[0085] O h =φ(Q h )φ(K h +B h ) T V h +DWC(V h )

[0086] Among them, φ(·) is a mapping function used to rearrange Q while maintaining the sharp distribution of the original Softmax attention mechanism h ,K h and V h The multiplication order of . represents the position encoding, and DWC(·) is a deep convolutional module used to restore feature diversity. In this case, the computational complexity of the attention mechanism is ,Generally speaking, the number of tokens N is greater than the channel dimension D, which makes the computational complexity linearly related to the number of patches N, and is more suitable for scenarios with a large number of patches.

[0087] In this example, the mapping function where x **p It means to raise x to the power of p bit by bit. The deep convolution module DWC(·) is implemented using 4 groups of one-dimensional convolutions with input dimension of 24, output dimension of 24, kernel size of 5, stride of 1, and padding of 2.

[0088] (7) Figure 3 As shown in (b), for the sliding window linear attention SWLA described in step (3), in a possible implementation, the calculation steps are as follows: after the input token is normalized and split into non-overlapping windows, the token is cyclically slid by half a window, and the self-attention mechanism is performed within each window instead of on the entire input token. After the attention mechanism, a pre-normalized residual unit with Dropout is used to help the model converge.

[0089] In this example, the window size is set to 32, the number of attention heads is 4, and the Dropout zero probability is set to 0.1.

[0090] (8) Use the fully connected layer as a classifier; output the probability that the current signal belongs to each transmitter.

[0091] In this example, the classifier includes a one-dimensional adaptive average pooling layer, a dimension flattening layer, and a fully connected layer with an input dimension of 256 and an output dimension of 100, which are connected in sequence.

[0092] (9) Inverse Sigmoid is used to implement a defense mechanism that can manipulate the attacker’s perception of the loss surface, moving it away from the optimal parameters and hindering it from extracting the model smoothly.

[0093] Assume that the model probability value is Y = {y i}, where y i Represents the probability value of the i-th radiation source; the model probability value after adding noise disturbance is

[0094] For the model probability value Y = {yi}Perform the following perturbations:

[0095]

[0096] in is a disturbance with controllable amplitude, α i is the normalization parameter:

[0097]

[0098] In order to maintain the accuracy of the model and retain the original probability value as much as possible, while increasing the difficulty of inversion for the attacker, a perturbation containing inverse Sigmoid is used. This perturbation function is almost irreversible:

[0099] r(y i )≈β(s -1 (y i )-1 / 2)

[0100] Among them, β is a positive amplitude adjustment parameter, s -1 (·) is the inverse sigmoid function:

[0101] s -1 (y i )=s(γg(y i ))

[0102] where s(·) is the sigmoid function, γ is a convergence parameter specific to the positive dataset and model, and g(y i ) is y i The pseudo-logical value of The behavior of and makes the perturbation comparable to the original probability:

[0103]

[0104] In this example, the amplitude adjustment parameter β=1.0; the specific convergence parameter γ=0.7.

[0105] Furthermore, in a specific simulation experiment embodiment provided by the present invention, the details are as follows:

[0106] 1. Simulation conditions: The hardware platform of the simulation experiment of the present invention is: the processor is Montage Jintide(R)C5218R, the main frequency is 2.10GHZ, the memory is 96GB, and the GPU is NVIDIA GeForce RTX 4090.

[0107] The software platform of the simulation experiment of the present invention is: Ubuntu 20.04.6LTS operating system, python=3.10, pytorch=2.1.0.

[0108] 2. Simulation content and result analysis:

[0109] The batch size is set to 128, the AdamW optimizer is used, the OneCycleLR learning rate is used, the initial learning rate is set to 0.001, the maximum learning rate reaches 0.005 in the 30th round, the final learning rate is set to 0.001, the minimum training round is 400, the maximum training round is 500, and the training is stopped when the accuracy of the test set does not increase after 40 consecutive rounds. During the training process, the accuracy of the test set samples at different training rounds is as follows Figure 4 As shown in the figure, the model reaches a maximum accuracy of 95.45% on the test set at the 464th round. The floating point calculation amount of the model tested using the thop library is only 93.52M, and the parameter amount is only 0.75M. It can be seen that the radiation source identification method based on Conformer provided by the present invention achieves high recognition accuracy while maintaining low computational cost, has strong feature extraction capability, and is suitable for deployment in actual environments.

[0110] For an attacker, the most limited resource may be real data samples that query the stolen model, so consider the case where the attacker has different proportions of the original training data set. Figure 5 As shown, after adding the defense layer, the method provided by the present invention can reduce the accuracy of the attacker's model to less than 10%, indicating that the adopted defense mechanism can effectively limit the attacker's ability to steal the model and can be used to enhance the security of the radiation source identification system.

[0111] Further references Figure 6 As an implementation of the methods shown in the above figures, the present application provides an embodiment of a system, which is similar to Figure 1 Corresponding to the method embodiment shown, the system can be specifically applied to various electronic devices.

[0112] In a second aspect, this embodiment also discloses a radiation source identification system based on Conformer, such as Figure 6 As shown, the system includes: a signal acquisition module 61, a preprocessing module 62, a feature extraction module 63, a classifier module 64 and a defense module 65.

[0113] In a specific embodiment, the signal acquisition module 61 is configured to collect the radio frequency signal of the radiation source and obtain the in-phase part of the original signal as the input signal; the preprocessing module 62 is configured to perform patch segmentation and embedding on the original in-phase signal, divide the input signal into patches through a one-dimensional convolution operation with preset parameters, and project it into a preset dimensional space to obtain an embedding vector, and then use Dropout to regularize the feature vector; the feature extraction module 63 is configured to use a WLC block and a SWLC block to further extract signal features, including a feedforward network, a normal window linear attention WLA or a sliding window linear attention SWLA, a point convolution and a downsampling convolution operation to output a feature map;

[0114] The classifier module 64 is configured to use the fully connected layer as a classifier to output the probability that the current signal belongs to each transmitter; the defense module 65 is configured to use the inverse sigmoid to implement a defense mechanism, which manipulates the attacker's perception of the loss surface by perturbing the model probability value, thereby hindering the attacker from extracting the model.

[0115] The functions of the above modules correspond to the methods and will not be repeated here.

[0116] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer readable medium, and the computer program includes a program code for executing the method shown in the flowchart.

[0117] It should be noted that the computer-readable medium described in the present invention may be a computer-readable signal medium or a computer-readable medium or any combination of the above two. The computer-readable medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor device, apparatus or device, or any combination of the above. More specific examples of computer-readable media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution device, apparatus or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer-readable signal medium may also be any computer-readable medium other than a computer-readable medium that can send, propagate or transmit a program for use by or in conjunction with an instruction execution device, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0118] Computer program code for performing the operations of the present invention may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0119] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the device, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based device that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0120] The modules involved in the embodiments of the present invention may be implemented in software or hardware, and the modules described may also be arranged in a processor.

[0121] As another aspect, the present invention also provides a computer-readable medium, which may be included in the electronic device described in the above embodiment; or it may exist alone and not be assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by the electronic device, the electronic device: collects the radio frequency signal of the radiation source, obtains the in-phase part of the original signal as the input signal; performs patch segmentation and embedding on the original in-phase signal, divides the input signal into patches through a one-dimensional convolution operation with preset parameters, and projects it into a preset dimensional space to obtain an embedding vector, and then uses Dropout to regularize the feature vector; uses WLC blocks and SWLC blocks to further extract signal features, including a feedforward network, a normal window linear attention WLA or a sliding window linear attention SWLA, a point convolution and a downsampling convolution operation to output a feature map; uses a fully connected layer as a classifier to output the probability that the current signal belongs to each transmitter; and uses an inverse Sigmoid to implement a defense mechanism, which perturbs the model probability value, manipulates the attacker's perception of the loss surface, and hinders the attacker from extracting the model.

[0122] The above description is only a preferred embodiment of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present invention is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the above features are replaced with the technical features with similar functions disclosed in the present invention (but not limited to) to form a technical solution.

Claims

1. A radiation source identification method based on Conformer, characterized in that: The method comprises the following steps: Collect the radio frequency signal of the radiation source and obtain the in-phase part of the original signal as the input signal; Perform patch segmentation and embedding on the original in-phase signal, divide the input signal into patches through a one-dimensional convolution operation with preset parameters, and project it into a preset dimensional space to obtain an embedding vector, and then use Dropout to regularize the feature vector; Further extract signal features using WLC blocks and SWLC blocks, which include feedforward networks, normal window linear attention WLA or sliding window linear attention SWLA, point convolution and downsampling convolution operations to output feature maps; Use the fully connected layer as a classifier to output the probability that the current signal belongs to each transmitter; as well as The inverse sigmoid is used to implement the defense mechanism. By perturbing the model probability value, the attacker's perception of the loss surface is manipulated, thus hindering the attacker from extracting the model.

2. The method according to claim 1 based on Conforme r The radiation source identification method is characterized in that: When performing patch segmentation and embedding on the original in-phase signal, the input original in-phase signal S=[s1, s2, ..., s k ] 1×k , whose length is k, and the patch token with a patch size of p is defined as Where n = k / p, and the one-dimensional convolution with a kernel size of p slides on the input signal with a step size of p to obtain the embedding vector E = [Conv1d(S)] T n×d , where d is the Embedding dimension.

3. The method according to claim 1 based on Conforme r The radiation source identification method is characterized in that: The method of further extracting signal features using the WLC block and the SWLC block includes: given an input vector X of the WLC block and / or the SWLC block, the calculation process of calculating the output feature map F is as follows: F2=F1+WLA / SWLA(F1) F3=PC(F2)+DC(F2) Among them, FN1 and FN2 are feedforward networks, WLA is normal window linear attention, SWLA is sliding window linear attention, LN is layer normalization, and PC is used The expansion factor is used to project the number of channels of the point convolution, D F is the output dimension of the block, D X is the input dimension of the block, and DC is the downsampling convolution used to reduce the length of the feature map sequence.

4. The method for identifying radiation sources based on Conformer according to claim 3, characterized in that: The calculation process of the down-sampling convolution DC is expressed as follows: Y1=GLU(PC1(LN(F2))) Y2=Swish(BN(SDC(Y1))) DC 输出 =Drop(PC2(Y2)) Among them, GLU stands for gated linear unit; Swish and Drop refer to Swish activation function and Dropout respectively; BN is batch normalization, and SDC stands for strided depthwise convolution.

5. The method for radiation source identification based on Conformer according to claim 3, characterized in that: The calculation steps of the ordinary window linear attention WLA are: The input token is normalized and split into non-overlapping windows. The self-attention mechanism is performed in each window. After the attention mechanism, a pre-normalized residual unit with Dropout is used to help the model converge. The self-attention mechanism decouples the Softmax operation through an approximate function and calculates To change the order of calculation: ABOUT h =φ(Q h )φ(K h +B h ) T V h +DWC(V h ) Among them, φ(·) is the mapping function, represents position encoding, and DWC(·) is a deep convolutional module.

6. The method for identifying radiation sources based on Conformer according to claim 5, characterized in that: The calculation steps of the sliding window linear attention SWLA are: After the input token is normalized and split into non-overlapping windows, the token is cyclically slid by half a window, and the self-attention mechanism is executed in each window. After the attention mechanism, a pre-normalized residual unit with Dropout is used to help the model converge.

7. The method for radiation source identification based on Conformer according to claim 1, characterized in that: The classifier includes a one-dimensional adaptive average pooling layer, a dimension flattening layer, and a fully connected layer with preset input and output dimensions, which are connected in sequence.

8. The method for identifying radiation sources based on Conformer according to claim 1, characterized in that: The defense mechanism implemented by using inverse sigmoid perturbs the model probability value, manipulates the attacker's perception of the loss surface, and hinders the attacker from extracting the model, including: The defense mechanism perturbs the model probability value Y = {yi} to obtain the model probability value after adding noise perturbation: The disturbance mode is Where yi represents the probability value of the i-th radiation source, is a disturbance with controllable amplitude, α i is the normalization parameter: Using the perturbation with inverse sigmoid, the perturbation function is almost irreversible: r(y i )≈β(S -1 (y i )-1 / 2), where β is a positive amplitude adjustment parameter, S -1 (·) is the inverse sigmoid function: s -1 (y i )=s(γg(y i )), where s(·) is the sigmoid function, γ is a positive dataset and model-specific convergence parameter, and g(y i ) is y i Pseudo-logical value of:

9. A radiation source identification system based on Conformer, characterized in that: include: A signal acquisition module is configured to acquire a radio frequency signal from a radiation source and obtain an in-phase portion of the original signal as an input signal; A preprocessing module is configured to perform patch segmentation and embedding on the original in-phase signal, divide the input signal into patches through a one-dimensional convolution operation with preset parameters, and project it into a preset dimensional space to obtain an embedding vector, and then use Dropout to regularize the feature vector; A feature extraction module configured to further extract signal features using a WLC block and a SWLC block, including a feed-forward network, a normal window linear attention WLA or a sliding window linear attention SWLA, a point convolution, and a down-sampling convolution operation to output a feature map; A classifier module configured to use the fully connected layer as a classifier and output a probability that the current signal belongs to each transmitter; The defense module is configured to implement a defense mechanism using inverse sigmoid, which perturbs the model probability value to manipulate the attacker's perception of the loss surface and hinder the attacker from extracting the model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images