Construction method of radio frequency fingerprint authentication dataset based on feature orthogonal decoupling

Abstract

The application discloses a kind of based on feature orthogonal decoupling's radio frequency fingerprint authentication data set construction method, it is related to radio frequency fingerprint identity authentication field, the method of the present application collects multiple equipment radio frequency signal and constructs training data set, trains orthogonal decoupling auto-encoder network, and through the orthogonal separation of mutual information minimization's confrontation training in latent space between baseband semantic information and hardware fingerprint feature;When applying, the fingerprint feature vector of target device and local device is extracted and pre-distortion compensation vector is calculated, the content vector of pure baseband data is spliced after with compensation vector, and then decoding reconstruction is generated, with the radio frequency simulation signal of target device fingerprint feature.The application fundamentally eliminates secondary nonlinear distortion, and can efficiently construct large-scale radio frequency fingerprint authentication data set with high fidelity.

Description

Technical Field

[0001] This invention belongs to the field of radio frequency fingerprint authentication technology, specifically relating to a method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling. Background Technology

[0002] With the rapid development of 6G mobile communication technology and massive IoT devices, Radio Frequency Fingerprint (RFF) authentication technology has become a key means of ensuring the security of the physical layer of wireless communication. RF fingerprints originate from the minute tolerances generated during the manufacturing process of analog hardware components inside the transmitter, such as power amplifiers, mixers, and oscillators. They possess characteristics such as device uniqueness, non-cloning, and tamper-proof properties, and are hailed as the physical DNA of wireless devices. In the research and testing of RF fingerprint authentication systems for next-generation communication base stations, it is necessary to construct extremely large and diverse RF datasets to fully train and verify the performance and defense capabilities of the identification system. However, realistically collecting RF signals from massive amounts of different terminal devices is costly, time-consuming, and limited by device availability. Therefore, efficiently generating high-fidelity RF fingerprint datasets through simulation has become an urgent need for the industry.

[0003] Currently, RF fingerprint simulation dataset construction techniques are mainly divided into two categories: data-driven and physics-driven. Data-driven methods, represented by generative adversarial networks (GANs) and diffusion models, implicitly learn the joint distribution of hardware impairments and channel effects through massive sample training, thereby synthesizing RF waveforms of specific devices. For example, models combining variational autoencoders and GANs, as well as time-frequency domain adversarial generation methods, can expand sample diversity to some extent. However, these purely data-driven methods lack underlying physical constraints and are prone to producing artifacts that do not conform to hardware characteristics. Physics-driven methods attempt to improve physical fidelity by mathematically modeling hardware distortions such as power amplifier nonlinearity and in-phase or quadrature imbalance, constructing a differentiable simulation framework. Recently, a deep learning-based digital baseband cascaded preprocessing architecture has been proposed in the field of RF fingerprint feature simulation. The architecture first inputs the clean baseband signal into a pre-trained inverse feature elimination model, artificially superimposing pre-distortion information that is inversely phase to the distortion of the local hardware; then, the resulting inverse compensation signal is serially fed into the forward feature injection model to apply the RF fingerprint mapping rule of the target device; finally, the cascaded distortion signal is sent into the real physical transmission link, hoping that by offsetting the inherent distortion of the local hardware with the pre-distortion information, the final output synthetic RF signal will retain only the characteristics of the target device, thus serving as a sample for inclusion in the authentication dataset.

[0004] However, the aforementioned cascaded preprocessing architecture has serious flaws in its underlying mathematical logic. Since the forward feature injection model is essentially a highly nonlinear neural network mapping system, when the inverse compensation signal carrying local predistortion information forcibly penetrates this nonlinear model, the originally precisely calculated compensation features undergo uncontrollable nonlinear distortion and deformation. This secondary nonlinear distortion causes the compensation signal that finally reaches the physical hardware layer to no longer be the absolute inverse of the physical distortion, thus triggering cross-contamination of the feature dimensions. The direct consequence is that the local baseline features cannot be completely canceled out; residual features intertwine with the deformed target features, generating a large amount of distortion noise in the time-domain waveform. Simultaneously, the waveform fidelity of the baseband communication signal itself is severely compromised, and the communication bit error rate and normalized mean square error increase sharply. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling, comprising: Radio frequency signals from multiple different terminal devices are collected, each radio frequency signal is labeled with a corresponding device identification identifier, and each radio frequency signal is preprocessed to obtain a corresponding two-dimensional feature matrix, forming a training dataset; Construct an orthogonal decoupled autoencoder network, and use the training dataset to perform adversarial training on the orthogonal decoupled autoencoder network to minimize mutual information, so as to orthogonally separate the baseband semantic information and hardware fingerprint features in the latent space, and obtain the trained orthogonal decoupled autoencoder network. The historical communication signals of the target device and the local device are preprocessed and then input into the fingerprint feature encoder of the trained orthogonal decoupled autoencoder network to extract the target feature vector and the local feature vector. Based on the target feature vector and the local feature vector, the predistortion compensation vector is calculated. The generated clean digital baseband data is preprocessed and then input into the content encoder of the trained orthogonal decoupled autoencoder network to extract the clean baseband content latent vector. The clean baseband content latent vector is concatenated with the predistortion compensation vector and then sequentially passed through the joint decoder of the trained orthogonal decoupled autoencoder network and the simulation base to generate an RF simulation signal with the fingerprint characteristics of the target device, which is then included in the RF fingerprint authentication dataset.

[0006] In one embodiment of the present invention, the preprocessing includes: Extract the in-phase channel data sequence and the positive-phase channel data sequence of the radio frequency signal, and calculate the amplitude and the square of the amplitude of the radio frequency signal at each moment as the feature data of the radio frequency signal; Construct a temporal memory window that includes the current moment and the past M moments; The characteristic data of the radio frequency signal are sampled using the timing memory window to obtain the in-phase channel data, positive channel data, amplitude, and amplitude squared corresponding to the timing memory window. The sampled in-phase channel data, positive-phase channel data, amplitude, and amplitude squared are aligned and concatenated according to the time dimension to obtain the two-dimensional feature matrix.

[0007] In one embodiment of the present invention, the orthogonal decoupled autoencoder network includes a content encoder, a fingerprint feature encoder, a joint decoder, and an adversarial discriminator module; The two-dimensional feature matrix is ​​received in parallel at the input ends of the content encoder and the fingerprint feature encoder; The outputs of both the content encoder and the fingerprint feature encoder are connected to the adversarial discriminator module; The outputs of the content encoder and the fingerprint feature encoder are concatenated and then input into the joint decoder.

[0008] In one embodiment of the present invention, the adversarial discriminator module includes a discriminator and an adversarial discriminator; the output of the fingerprint feature encoder is connected to the discriminator; and the output of the content encoder is connected to the adversarial discriminator.

[0009] In one embodiment of the present invention, during the training process, the orthogonal decoupled autoencoder network... The content encoder is used to map the two-dimensional feature matrix into a baseband content latent vector, and the baseband content latent vector is used to represent the pure communication semantic information after stripping the hardware fingerprint; The fingerprint feature encoder is used to map the two-dimensional feature matrix into a radio frequency fingerprint latent vector, and the radio frequency fingerprint latent vector is used to characterize the transmitter hardware distortion features; The joint decoder is used to reconstruct a time-domain radio frequency signal from the spliced ​​baseband content latent vector and the radio frequency fingerprint latent vector. The discriminator is used to identify the device identity based on the radio frequency fingerprint hidden vector; The adversarial discriminator is used to identify device identity based on the baseband content latent vector.

[0010] In one embodiment of the present invention, the total loss function of the orthogonal decoupled autoencoder network during training includes: the reconstruction loss of the joint decoder, the adversarial loss of the adversarial discriminator, and the classification loss of the discriminator.

[0011] In one embodiment of the present invention, during the training process, when the orthogonal decoupled autoencoder network performs backpropagation using the total loss function, it flips the gradient passed from the adversarial discriminator to the content encoder, so that the content encoder removes all information related to device identity until the probability of the adversarial discriminator identifying the device identity identifier from the baseband content latent vector is reduced to the level of random guessing.

[0012] In one embodiment of the present invention, the historical communication signals of the target device and the historical communication signals of the local device are preprocessed and then input into the fingerprint feature encoder of the trained orthogonal decoupled autoencoder network to extract the target feature vector and the local feature vector, including: The historical communication signals of the target device are preprocessed to obtain two-dimensional feature matrices of multiple time segments. The fingerprint feature encoder of the trained orthogonal decoupled autoencoder network is used to extract features from the two-dimensional feature matrices of multiple time segments corresponding to the historical communication signals of the target device. The extracted feature vectors are averaged in the time dimension to obtain the target feature vector. The historical communication signals of the local device are preprocessed to obtain two-dimensional feature matrices of multiple time segments. The fingerprint feature encoder of the trained orthogonal decoupled autoencoder network is used to extract features from the two-dimensional feature matrices of multiple time segments corresponding to the historical communication signals of the local device. The extracted feature vectors are averaged in the time dimension to obtain the local feature vector.

[0013] In one embodiment of the present invention, the predistortion compensation vector is calculated in the decoupled feature vector space by subtracting the target feature vector from the local feature vector.

[0014] In one embodiment of the present invention, the pure baseband content latent vector and the predistortion compensation vector are concatenated and then sequentially passed through the joint decoder of the trained orthogonal decoupled autoencoder network and the simulation base for simulation to generate an RF simulation signal with the fingerprint characteristics of the target device, including: The pure baseband content latent vector and the predistortion compensation vector are linearly concatenated along the feature channel dimension to obtain a joint vector; The joint vector is input into the joint decoder of the trained orthogonal decoupled autoencoder network for forward decoding and reconstruction to generate a digital baseband drive signal containing pre-distortion features. In a pure digital simulation environment, the digital baseband drive signal is passed through a simulation base, causing the inherent low-level characteristics of the local device to cancel out the inverse compensation characteristics in the predistortion compensation vector, thereby generating a radio frequency simulation signal with the fingerprint characteristics of the target device.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention provides a method for constructing an RF fingerprint authentication dataset based on orthogonal decoupling of features. This method constructs an orthogonal decoupling architecture based on minimizing adversarial mutual information, achieving complete isolation between content and features. It fundamentally separates physical features from communication content, providing an interference-free feature baseline for high-fidelity feature simulation in the digital domain. 2. The method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling of the present invention obtains the pre-distortion compensation vector through simple vector algebraic subtraction calculation in mutually independent feature spaces. The compensation calculation is not affected by nonlinear cross-contamination, ensuring that the synthesized target device fingerprint features have extremely high physical authenticity. 3. The method for constructing an RF fingerprint authentication dataset based on feature orthogonal decoupling of the present invention, when generating large-scale datasets, concatenates the real-time baseband content latent vector with the fixed-stored pre-distortion compensation vector, and completes signal synthesis through a single joint decoder forward inference. This significantly reduces the computational latency of generating massive samples and ensures that the communication semantics are not affected by excessive nonlinear distortion, providing high-quality data support for training a highly robust authentication system.

[0016] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling, provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of a method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling provided in an embodiment of the present invention; Figure 3 This is a simulation architecture diagram of an radio frequency fingerprint based on an orthogonal decoupled autoencoder network provided in an embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following describes in detail, with reference to the accompanying drawings and specific embodiments, a method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling proposed in accordance with the present invention.

[0019] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0020] This invention provides a method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling. Please refer to [link to relevant documentation]. Figure 1 , Figure 1 This is a flowchart illustrating a method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling, as provided in an embodiment of the present invention. Figure 1 As shown, the overall execution flow of the method of the present invention mainly consists of four core technical steps: data preprocessing and multi-channel feature representation, construction and offline training of orthogonal decoupled autoencoders, feature linear hedging calculation in manifold space, and decoding and physical emission.

[0021] Please refer to the above. Figure 2 , Figure 2 This is a schematic diagram of a method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling, provided by an embodiment of the present invention. Figure 2 As shown, specifically, the method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling in this embodiment includes the following steps: Step 1: Collect radio frequency signals from multiple different terminal devices, label each radio frequency signal with the corresponding device identification, preprocess each radio frequency signal to obtain the corresponding two-dimensional feature matrix, and form a training dataset.

[0022] In this embodiment, before performing deep feature extraction, the radio frequency signal (one-dimensional time domain signal) acquired by the radio frequency transceiver must be converted into a multi-dimensional feature matrix that can fully reflect the physical distortions of the underlying hardware (such as the memory effect of the power amplifier and the imbalance of the mixer).

[0023] Specifically, preprocessing includes: S1: Extract the in-phase channel data sequence and the positive-phase channel data sequence of the radio frequency signal, and calculate the amplitude and the square of the amplitude of the radio frequency signal at each moment as the characteristic data of the radio frequency signal.

[0024] In this embodiment, the in-phase channel (I-channel) data sequence and the positive-mode channel (Q-channel) data sequence are extracted from the radio frequency signal. In order to more accurately capture the nonlinear distortion caused by the power amplifier, in addition to the basic I / Q sequence, the amplitude of the signal at the current moment and the square of the amplitude are calculated as independent feature dimensions.

[0025] S2: Construct a time-series memory window that includes the current time and the past M time points.

[0026] Because the power amplifier not only distorts the current RF signal, its internal energy storage components such as capacitors and inductors also cause the RF signal from historical moments to affect the current output. For each current moment n, the sampled values ​​from the past M moments are extracted.

[0027] S3: Use the timing memory window to perform sliding sampling on the characteristic data of the radio frequency signal to obtain the in-phase channel data, positive channel data, amplitude, and amplitude squared corresponding to the timing memory window.

[0028] S4: Align and concatenate the sampled in-phase channel data, positive-mode channel data, amplitude, and amplitude squared according to the time dimension to obtain a two-dimensional feature matrix.

[0029] In this embodiment, the extracted I-path values, Q-path values, amplitude |u|, and amplitude squared |u| from the current time n and the past M times are used. 2 Each of these four channels is treated as a separate one-dimensional channel. These four channels are then aligned and concatenated along the time dimension to form a 4×(M+1) two-dimensional feature matrix (2DFeatureMatrix). This two-dimensional feature matrix will serve as the standard input to the subsequent orthogonal decoupled autoencoder network.

[0030] Step 2: Construct an orthogonal decoupled autoencoder network. Use the training dataset to perform adversarial training on the orthogonal decoupled autoencoder network to minimize mutual information, so as to orthogonally separate the baseband semantic information and hardware fingerprint features in the latent space, and obtain the trained orthogonal decoupled autoencoder network.

[0031] In this embodiment, the orthogonal decoupled autoencoder network includes a content encoder, a fingerprint feature encoder, a joint decoder, and an adversarial discriminator module. The inputs of the content encoder and the fingerprint feature encoder receive two-dimensional feature matrices in parallel; the outputs of both the content encoder and the fingerprint feature encoder are connected to the adversarial discriminator module; the outputs of the content encoder and the fingerprint feature encoder are concatenated and then input to the joint decoder. The adversarial discriminator module includes a discriminator and an adversarial discriminator; the output of the fingerprint feature encoder is connected to the discriminator; the output of the content encoder is connected to the adversarial discriminator.

[0032] In this embodiment, the content encoder (E) cont It consists of multiple stacked one-dimensional and two-dimensional convolutional neural networks, with flattening layers at the ends. Its input is the two-dimensional feature matrix with hardware distortion obtained in step 1, and its output is a 64-dimensional latent vector, denoted as the baseband content latent vector Z. cont The physical meaning of this vector is the pure communication semantic information stripped of all hardware defects.

[0033] In this embodiment, the fingerprint feature encoder (E rff The structure of the RF fingerprint latent vector is similar to that of the content encoder, but the weights are independent. Its input is also a two-dimensional feature matrix with hardware distortion, and its output is another 64-dimensional latent vector, denoted as the RF fingerprint latent vector Z. rff The physical meaning of this vector is purely to characterize the nonlinear distortion features caused by the transmitter's internal hardware (PA nonlinearity, I / Q imbalance).

[0034] In this embodiment, the joint decoder (D) receives Z during the training process. cont and Z rff It first linearly concatenates the two orthogonal latent vectors along the feature channel dimension to obtain a complete 128-dimensional fused vector. Then, through multiple deconvolutional or fully connected layers, it reconstructs the vector into a time-domain radio frequency signal with temporal characteristics.

[0035] Please refer to the figure. Figure 3 This is a simulation architecture diagram of an radio frequency fingerprint based on an orthogonal decoupled autoencoder network provided by an embodiment of the present invention. Figure 3 As shown, specifically, during the training process of the orthogonal decoupled autoencoder network, the content encoder is used to map the two-dimensional feature matrix into a baseband content latent vector, which represents the pure communication semantic information after removing the hardware fingerprint; the fingerprint feature encoder is used to map the two-dimensional feature matrix into an RF fingerprint latent vector, which represents the transmitter hardware distortion features; the joint decoder is used to reconstruct the time-domain RF signal from the spliced ​​baseband content latent vector and the RF fingerprint latent vector; the discriminator is used to identify the device identity based on the RF fingerprint latent vector; and the adversarial discriminator is used to identify the device identity based on the baseband content latent vector.

[0036] In this embodiment, the total loss function of the orthogonal decoupled autoencoder network during training includes: the reconstruction loss of the joint decoder, the adversarial loss of the adversarial discriminator, and the classification loss of the discriminator. The reconstruction loss can employ mean squared error (MSE) to ensure that the joint decoder can reconstruct the signal losslessly.

[0037] In this embodiment, during the training process, the orthogonal decoupled autoencoder network flips the gradient passed from the adversarial discriminator to the content encoder when backpropagating using the total loss function. This causes the content encoder to remove all information related to device identity until the probability of the adversarial discriminator identifying the device identity from the baseband content latent vector is reduced to the level of random guessing.

[0038] Understandably, this is to ensure the latent vector Z of the baseband content. cont and radio frequency fingerprint hidden vector Z rff Since the latent vectors are orthogonal and mutually exclusive in high-dimensional space, an adversarial discriminator module (comprising a discriminator and an adversarial discriminator) is introduced during the training phase. First, the discriminator only receives the latent vector Z of the radio frequency fingerprint. rff It is required to be able to identify the specific transmitter device from which the current signal originates with near 100% accuracy (i.e., identify the device's identity). Then, the adversarial discriminator only receives the baseband content latent vector Z. cont It attempts to deduce the device's identity from this information. During backpropagation to update network parameters, this embodiment deliberately reverses the propagation of the adversarial discriminator to the content encoder E. cont The gradient of E is forced through this game-like training that minimizes mutual information. cont Completely remove all RFID fingerprint information related to device identity. When training converges, the adversarial discriminator extracts data from the baseband content latent vector Z. cont The probability of guessing the device identity will decrease to the level of completely random guessing (e.g., the probability approaches 8.33% with 12 devices). This makes the baseband content latent vector Z... cont It no longer contains any hardware fingerprints, achieving full feature orthogonal decoupling.

[0039] Step 3: After preprocessing the historical communication signals of the target device and the local device, input them into the fingerprint feature encoder of the trained orthogonal decoupled autoencoder network to extract the target feature vector and the local feature vector. Based on the target feature vector and the local feature vector, calculate the predistortion compensation vector.

[0040] It should be noted that after training the orthogonal decoupled autoencoder network, complex radio frequency signals can be deconstructed into independent feature vectors. The originally highly nonlinear hardware distortion has been transformed into linear vectors that can be directly computed algebraically. Performing local feature elimination and target synthesis calculations in this state will avoid secondary distortion.

[0041] Please continue reading Figure 3Specifically, the historical communication signals of the target device and the local device are preprocessed and then input into the fingerprint feature encoder of the trained orthogonal decoupled autoencoder network to extract the target feature vector and the local feature vector, including: Step a: Preprocess the historical communication signals of the target device to obtain two-dimensional feature matrices of multiple time segments. Use the fingerprint feature encoder of the trained orthogonal decoupled autoencoder network to extract features from the two-dimensional feature matrices of multiple time segments corresponding to the historical communication signals of the target device. Average the extracted feature vectors in the time dimension to obtain the target feature vector.

[0042] In this embodiment, the target feature vector represents a static constant vector of the inherent physical characteristics of the target device, denoted as Z. tar .

[0043] Step b: Preprocess the historical communication signals of the local device to obtain two-dimensional feature matrices for multiple time segments. Use the fingerprint feature encoder of the trained orthogonal decoupled autoencoder network to extract features from the two-dimensional feature matrices of the historical communication signals of the local device for multiple time segments. Average the extracted feature vectors in the time dimension to obtain the local feature vector.

[0044] In this embodiment, the local feature vector represents a static constant vector of inherent hardware defects of the local device, denoted as Z. self .

[0045] In this embodiment, the predistortion compensation vector is calculated in the decoupled feature vector space by subtracting the target feature vector from the local feature vector. Specifically, the predistortion compensation vector Z... pre = Z tar - Z self。 It should be noted that existing technologies directly force signal cancellation within a nonlinear network in the time domain, which can easily lead to waveform distortion. In contrast, this embodiment performs a simple subtraction operation on the "fingerprint feature vector" under a decoupled feature state. Since the fingerprint feature vector and the baseband content vector are forcibly isolated by the network and do not interfere with each other, this subtraction operation will not contaminate or distort the baseband communication waveform, eliminating "secondary nonlinear distortion" and ensuring that the RF features ultimately generated and included in the dataset have high physical fidelity.

[0046] Step 4: After preprocessing, the generated clean digital baseband data is input into the content encoder of the trained orthogonal decoupled autoencoder network to extract the clean baseband content latent vector. The clean baseband content latent vector is concatenated with the predistortion compensation vector and then simulated by the joint decoder of the trained orthogonal decoupled autoencoder network and the simulation base to generate an RF simulation signal with the fingerprint characteristics of the target device, which is then included in the RF fingerprint authentication dataset.

[0047] Please continue reading Figure 3 In this embodiment, the pure baseband content latent vector and the predistortion compensation vector are concatenated and then sequentially passed through the joint decoder of a trained orthogonal decoupled autoencoder network and a simulation base for simulation, generating an RF simulation signal with the fingerprint characteristics of the target device, including: Step i: Linearly concatenate the pure baseband content latent vector and the predistortion compensation vector along the feature channel dimension to obtain the joint vector; Step ii: Input the joint vector into the joint decoder of the trained orthogonal decoupled autoencoder network and perform a forward decoding reconstruction to generate a digital baseband drive signal containing predistortion features; Step iii: In a pure digital simulation environment, the digital baseband drive signal is passed through the simulation base, so that the inherent low-level characteristics of the local device cancel each other out with the reverse compensation characteristics in the predistortion compensation vector, generating an RF simulation signal with the fingerprint characteristics of the target device.

[0048] Specifically, the simulation platform generates clean digital baseband data u that will be used as sample content in the dataset. After preprocessing, this data is input into the trained content encoder E. cont A forward propagation is performed to rapidly extract the pure baseband latent vector Z. cont (u). The obtained pure baseband content latent vector Z cont (u), and the pre-calculated static constant vector predistortion compensation vector Z stored in memory. pre Perform direct concatenation. The concatenated joint vector [Z] cont (u),Z pre The signal is fed into the trained joint decoder D for forward decoding to generate a digital baseband drive signal X containing predistortion features.

[0049] In a purely digital simulation environment, the digital baseband drive signal X with the aforementioned pre-distortion characteristics is fed into the simulation platform for final virtual evolution and output. When the signal passes through the system's virtual baseline link (i.e., the underlying mathematical or fundamental network model used to generate data), the inherent underlying features of the baseline model are mathematically equivalent to a baseline feature Z being forcibly superimposed in the feature vector space. selfAt this point, the overall characteristics of the RF simulation signal, which is finally output in the digital domain and saved to disk as a dataset sample, are as follows: Z out = Z pre + Z self = (Z tar - Z self ) + Z self = Z tar .

[0050] Understandably, the inverse compensation feature in the predistortion compensation vector achieves digital-level zero-residual offset cancellation with the baseline features of the simulation system. The resulting RF simulation signal, generated and entered into the dataset, exhibits fingerprint features that fully represent the true physical properties of the target device being simulated, while maintaining extremely high fidelity in its baseband communication waveform (free from secondary distortion contamination, ensuring signal quality for training). This is achieved by rapidly switching between loading different target feature vectors Z on the software side. tar This allows for the efficient and low-cost generation of radio frequency fingerprint data for hundreds or thousands of different terminal devices within the same digital simulation platform, solving the cost and management challenges of large-scale physical device data collection.

[0051] The method for constructing an RF fingerprint authentication dataset based on orthogonal decoupling in this invention constructs an orthogonal decoupling architecture based on adversarial mutual information minimization, achieving complete isolation between content and features. This fundamentally separates physical features from communication content, providing an interference-free feature baseline for high-fidelity feature simulation in the digital domain. Within mutually independent feature spaces, a pre-distortion compensation vector is obtained through simple vector algebraic subtraction. The compensation calculation is not affected by nonlinear cross-contamination, ensuring that the synthesized target device fingerprint features have extremely high physical fidelity. When generating large-scale datasets, the real-time baseband content latent vector is concatenated with the permanently stored pre-distortion compensation vector, and signal synthesis is completed through a single forward inference by a joint decoder. This significantly reduces the computational latency of generating massive samples and ensures that communication semantics are not affected by excessive nonlinear distortion, providing high-quality data support for training a highly robust authentication system.

[0052] Furthermore, specific simulation experiments are used to verify and illustrate the effectiveness of the method for constructing the radio frequency fingerprint authentication dataset based on feature orthogonal decoupling in this invention.

[0053] The experiments were conducted using a digital domain software simulation platform. On the data generation and network operation sides, the generation of the original baseband signal (using 16-QAM modulation, 5MHz bandwidth, 2.4GHz carrier) and the mathematical modeling of various underlying nonlinear physical impairments were implemented using the MATLAB 2024a platform. The construction and training of the orthogonal decoupled autoencoder network, as well as the inference and synthesis of a large-scale high-fidelity target feature dataset, were completed using Python 3.8.18 and the PyTorch deep learning framework. Finally, the diverse simulation datasets efficiently generated by the system in the digital domain, after being superimposed with Gaussian white noise (AWGN) of varying intensities to simulate the actual signal-to-noise ratio (SNR) environment, were batch-fed into a pre-constructed complex-valued convolutional neural network (CVCNN, with 12 classification categories) for recognition and verification, thereby comprehensively evaluating the physical fidelity and effectiveness of the simulation datasets generated by the method of this invention.

[0054] The experiment selected "basic generated state without feature compensation", "traditional cascaded predistortion generation method", and the mainstream "data-driven adversarial generation method" as baseline comparison schemes. The test conditions were set at a signal-to-noise ratio (SNR) of 15 dB, and the dataset quality assessment test results are shown in Table 1.

[0055] Table 1 As can be seen from the data in Table 1, the traditional cascaded predistortion generation method suffers from incomplete baseline feature elimination in the simulation system due to the forced penetration of nonlinear models. The uncancelled baseline residuals are deeply coupled and cross-contaminated with the target features, resulting in a sample realism rating (TSR) of only 42.42% and significant waveform distortion. Using such data to train a recognition system would lead to a severe "simulation-reality" domain shift. In contrast, the method of this invention completely decouples the baseband content from the hardware fingerprint in the positive AC shape space and uses simple linear vector subtraction for local feature elimination and target synthesis, fundamentally avoiding secondary nonlinear distortion. Under the same pure digital simulation environment, the TSR of the data generated by the method of this invention jumps to 92.58%, indicating that the generated fingerprint features almost perfectly replicate the real hardware attributes of the target physical device. Simultaneously, because the decoupling mechanism does not contaminate the baseband communication semantics, its BER is extremely low (0.01%), and its NMSE reaches an optimal level of -18.8dB. This fully demonstrates that the method of the present invention can efficiently construct large-scale and diverse radio frequency fingerprint authentication datasets with extremely high physical fidelity and waveform quality in the pure digital domain.

[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0057] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0058] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling, characterized in that, include: Radio frequency signals from multiple different terminal devices are collected, each radio frequency signal is labeled with a corresponding device identification identifier, and each radio frequency signal is preprocessed to obtain a corresponding two-dimensional feature matrix, forming a training dataset; Construct an orthogonal decoupled autoencoder network, and use the training dataset to perform adversarial training on the orthogonal decoupled autoencoder network to minimize mutual information, so as to orthogonally separate the baseband semantic information and hardware fingerprint features in the latent space, and obtain the trained orthogonal decoupled autoencoder network. The historical communication signals of the target device and the local device are preprocessed and then input into the fingerprint feature encoder of the trained orthogonal decoupled autoencoder network to extract the target feature vector and the local feature vector. Based on the target feature vector and the local feature vector, the predistortion compensation vector is calculated. The generated clean digital baseband data is preprocessed and then input into the content encoder of the trained orthogonal decoupled autoencoder network to extract the clean baseband content latent vector. The clean baseband content latent vector is concatenated with the predistortion compensation vector and then sequentially passed through the joint decoder of the trained orthogonal decoupled autoencoder network and the simulation base to generate an RF simulation signal with the fingerprint characteristics of the target device, which is then included in the RF fingerprint authentication dataset.

2. The method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling according to claim 1, characterized in that, The preprocessing includes: Extract the in-phase channel data sequence and the positive-phase channel data sequence of the radio frequency signal, and calculate the amplitude and the square of the amplitude of the radio frequency signal at each moment as the feature data of the radio frequency signal; Construct a temporal memory window that includes the current moment and the past M moments; The characteristic data of the radio frequency signal are sampled using the timing memory window to obtain the in-phase channel data, positive channel data, amplitude, and amplitude squared corresponding to the timing memory window. The sampled in-phase channel data, positive-phase channel data, amplitude, and amplitude squared are aligned and concatenated according to the time dimension to obtain the two-dimensional feature matrix.

3. The method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling according to claim 1, characterized in that, The orthogonal decoupled autoencoder network includes a content encoder, a fingerprint feature encoder, a joint decoder, and an adversarial discriminator module; The two-dimensional feature matrix is ​​received in parallel at the input ends of the content encoder and the fingerprint feature encoder; The outputs of both the content encoder and the fingerprint feature encoder are connected to the adversarial discriminator module; The outputs of the content encoder and the fingerprint feature encoder are concatenated and then input into the joint decoder.

4. The method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling according to claim 3, characterized in that, The adversarial discriminator module includes a discriminator and an adversarial discriminator; the output of the fingerprint feature encoder is connected to the discriminator; the output of the content encoder is connected to the adversarial discriminator.

5. The method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling according to claim 4, characterized in that, During the training process, the orthogonal decoupled autoencoder network... The content encoder is used to map the two-dimensional feature matrix into a baseband content latent vector, and the baseband content latent vector is used to represent the pure communication semantic information after stripping the hardware fingerprint; The fingerprint feature encoder is used to map the two-dimensional feature matrix into a radio frequency fingerprint latent vector, and the radio frequency fingerprint latent vector is used to characterize the transmitter hardware distortion features; The joint decoder is used to reconstruct a time-domain radio frequency signal from the spliced ​​baseband content latent vector and the radio frequency fingerprint latent vector. The discriminator is used to identify the device identity based on the radio frequency fingerprint hidden vector; The adversarial discriminator is used to identify device identity based on the baseband content latent vector.

6. The method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling according to claim 4, characterized in that, The total loss function of the orthogonal decoupled autoencoder network during training includes: the reconstruction loss of the joint decoder, the adversarial loss of the adversarial discriminator, and the classification loss of the discriminator.

7. The method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling according to claim 6, characterized in that, During training, the orthogonal decoupled autoencoder network flips the gradient passed from the adversarial discriminator to the content encoder when backpropagating using the total loss function. This causes the content encoder to remove all information related to device identity until the probability of the adversarial discriminator identifying the device identity from the baseband content latent vector drops to the level of random guessing.

8. The method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling according to claim 1, characterized in that, The historical communication signals of the target device and the local device are preprocessed and then input into the fingerprint feature encoder of the trained orthogonal decoupled autoencoder network to extract the target feature vector and the local feature vector, including: The historical communication signals of the target device are preprocessed to obtain two-dimensional feature matrices of multiple time segments. The fingerprint feature encoder of the trained orthogonal decoupled autoencoder network is used to extract features from the two-dimensional feature matrices of multiple time segments corresponding to the historical communication signals of the target device. The extracted feature vectors are averaged in the time dimension to obtain the target feature vector. The historical communication signals of the local device are preprocessed to obtain two-dimensional feature matrices of multiple time segments. The fingerprint feature encoder of the trained orthogonal decoupled autoencoder network is used to extract features from the two-dimensional feature matrices of multiple time segments corresponding to the historical communication signals of the local device. The extracted feature vectors are averaged in the time dimension to obtain the local feature vector.

9. The method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling according to claim 1, characterized in that, The pre-distortion compensation vector is calculated in the decoupled feature vector space by subtracting the target feature vector from the local feature vector.

10. The method for constructing a radio frequency fingerprint authentication dataset based on feature orthogonal decoupling according to claim 1, characterized in that, The pure baseband latent vector and the predistortion compensation vector are concatenated and then sequentially passed through the joint decoder of the trained orthogonal decoupled autoencoder network and the simulation base for simulation, generating an RF simulation signal with the fingerprint characteristics of the target device, including: The pure baseband content latent vector and the predistortion compensation vector are linearly concatenated along the feature channel dimension to obtain a joint vector; The joint vector is input into the joint decoder of the trained orthogonal decoupled autoencoder network for forward decoding and reconstruction to generate a digital baseband drive signal containing pre-distortion features. In a pure digital simulation environment, the digital baseband drive signal is passed through a simulation base, causing the inherent low-level characteristics of the local device to cancel out the inverse compensation characteristics in the predistortion compensation vector, thereby generating a radio frequency simulation signal with the fingerprint characteristics of the target device.

Images