A radio frequency fingerprint identification method for receiver migration problem

Abstract

This invention relates to a radio frequency fingerprinting method for receiver migration problems, belonging to the field of radiation source individual identification technology. The invention introduces a demodulation and reconstruction method in data processing, which, compared with existing methods, enhances key features in the signal without losing information. This makes the training data superior in terms of features, allowing the network to extract more useful information, thus improving recognition performance and noise resistance. The invention incorporates a custom-designed Gaussian encoder into the network structure, enabling the discovery of more hidden information during feature extraction and reducing the impact of data spatial distribution on feature learning. The random variables introduced in the Gaussian encoder also enhance the network's noise resistance.

Description

Technical Field

[0001] This invention relates to a radio frequency fingerprinting method for receiver migration problems, belonging to the field of radiation source individual identification technology. Background Technology

[0002] In recent years, with the emergence of various standards such as Bluetooth Low Energy (BLE), WiFi, LoRa, and Sigfox, the number of Internet of Things (IoT) devices has surged. Authentication of IoT devices is crucial to ensuring that received messages are sent from authorized and legitimate devices. Traditional device authentication schemes typically rely on cryptographic encryption and use software addresses (such as MAC addresses) as the basis for device identification, but these are highly vulnerable to tampering, leading to potential spoofing attacks. In fact, establishing secure and efficient symmetric keys in the IoT is an extremely challenging task.

[0003] Radio Frequency Fingerprint Identification (RFFI) is an unencrypted device authentication technology. Like human fingerprints, wireless devices possess unique radio frequency fingerprints due to physical defects in the transmitter's front-end hardware components. These defects are difficult to mimic and typically include oscillator drift, mixer defects, and power amplifier nonlinearities. These defects have been modeled in previous research and considered suitable for device identification. RFFI systems extract unique features from the wireless signals transmitted by IoT devices to infer their identity. RFFI can be formulated as a multi-class classification problem, and numerous examples demonstrate that deep learning has been used to improve RFFI performance.

[0004] Despite significant efforts to improve RFFI performance, receiver effects have often been overlooked in previous research. Variations in receiver hardware characteristics can severely impact RFFI performance. Most existing RFFI work assumes the use of the same receiver during training and inference, and that receiver characteristics do not change over time. However, this assumption is not always applicable to real-world IoT applications. For example, mobile IoT devices will be served by different access points / gateways depending on their coverage. Furthermore, even using the same receiver, the hardware characteristics of low-cost receivers can change over time. While retraining on new receivers is a possible solution, long training times are often prohibitive when many end nodes are running in an IoT network. Therefore, there is an urgent need for a receiver-agnostic RFFI system that can be deployed in a highly practical manner. Collaborative RFFI methods addressing the receiver agnostic problem use multiple receivers to collect sufficient data packets from the device under test, employing adversarial learning to train a neural network. During inference, the receivers are equipped with the trained neural network model. Once a signal is captured, the receivers first perform independent inferences, then multi-receiver fusion is performed to achieve better classification performance. However, this method has high requirements for signal-to-noise ratio and lacks the ability to extract radio frequency fingerprint features, which limits the recognition effect, especially at low signal-to-noise ratios.

[0005] When receiver relocation is involved in practical applications, the performance of RF fingerprinting methods will be significantly affected due to changes in receiver hardware. Existing methods for receiver relocation rely too heavily on signal quality, lack RF fingerprint feature extraction capabilities, and are poorly adapted to low signal-to-noise ratio scenarios. Summary of the Invention

[0006] The technical problem solved by this invention is to overcome the shortcomings of the prior art and, based on the consideration of in-depth mining of signal features, propose a deep learning recognition method for radio frequency fingerprints based on signal reconstruction and Gaussian encoders. This addresses the problems of weak noise resistance, weak radio frequency fingerprint feature extraction capability, and insufficient accuracy of existing methods.

[0007] The technical solution of this invention is:

[0008] A radio frequency fingerprinting method for receiver migration problems, the method comprising the following steps:

[0009] The first step is to use two different receivers to collect the radio frequency signals of the authorized device, and store the collected radio frequency signals in a database as a training library for deep learning neural networks.

[0010] The second step is to demodulate and idealize the digital reconstruction of the radio frequency signals in the database from the first step to obtain the original signals without transmitter and receiver characteristics, and then splice them with the corresponding received signals to form training samples.

[0011] The third step is to design a neural network based on the characteristics and spatial distribution of the signal data, and introduce a Gaussian encoder to improve the network's anti-interference ability and deep feature extraction ability.

[0012] The fourth step involves using the training samples stitched together in the second step to train the network designed in the third step. Adversarial learning is employed to iteratively optimize the network parameters. The task objective and loss function are selected based on the actual situation. The resulting RF fingerprinting network is then directly applied to a known receiver or a new, unknown receiver.

[0013] In the first step, the radio frequency signal includes a receiver tag and an authorized device tag, and the radio frequency signal includes an I-channel signal and a Q-channel signal. The I-channel signal and the Q-channel signal are concatenated as follows:

[0014] x = [I rec Q rec ] T ∈R 2×N

[0015] Where x represents the received radio frequency signal, I rec Indicates the received I-channel signal, Q rec This represents the received Q-channel signal, and N represents the length of the radio frequency signal;

[0016] In the second step, to demodulate and reconstruct the signal, it is necessary to estimate the carrier frequency, code rate, time difference, and symbol decision of the acquired signal. After these steps, the basic information of the acquired radio frequency signal can be obtained, thereby digitally reconstructing the distortion-free modulated signal under ideal transmitter conditions.

[0017] The methods for signal demodulation and idealized digital reconstruction are as follows:

[0018] Step 2.1: Estimate the carrier frequency of the radio frequency signals in the database using the frequency center method. The method is described as follows:

[0019]

[0020] Where x(n) represents I in x rec The nth value and Q rec The vector [I(n); Q(n)] composed of the nth value T N represents the length of the radio frequency signal, f s Indicates the sampling rate. It is the estimated carrier frequency.

[0021] Step 2.2: Estimate the code rate of the radio frequency signal using the envelope squared spectrum method. First, perform a Hilbert transform on the radio frequency signal x(n) to convert it into an analyzable form:

[0022]

[0023] Where T s Indicates the sampling time interval, where the next After modulus calculation and sum of squares Fast Fourier Transform (FSFT), the envelope squared spectrum can be obtained. The point with the maximum amplitude in the spectrum is the estimated code rate. The operation process is as follows:

[0024]

[0025] This is the estimated code rate.

[0026] Step 2.3 estimates the time difference between the sampling start position and the symbol start position in the RF signal. This step uses a non-data-aided squared estimation method to obtain the time difference estimate. It is assumed that the time error is within... L It remains unchanged within a symbol period, and there will be within each symbol period. G Subsequent sampling, then time difference It can be obtained by the following method:

[0027]

[0028] Where L is the number of symbol periods, and G is the number of samples within a single symbol period;

[0029] Step 2.4: Perform symbol determination on the symbols in the radio frequency signal. This step can rely on the constellation diagram obtained by optimal sampling to determine the symbol corresponding to each symbol.

[0030] Step 2.5: After completing the parameter estimation above, the ideal modulation signal can be reconstructed. The ideal signal is then concatenated with the received signal to form a new sample for network training in subsequent steps. This sample has more obvious signal features than the original sample, and its format is as follows:

[0031] x conc =[I rec ;I ideal Q rec Q ideal ] T ∈R 4×N

[0032] Among them, I ideal Q represents the reconstructed ideal I-channel signal. ideal This represents the ideal Q-path signal after reconstruction.

[0033] In the third step, the designed neural network is based on a deep residual shrinking network, in which a Gaussian encoder is embedded to obtain the deep hidden features of the RF fingerprint, reducing the impact of noise interference and receiver characteristics on RF fingerprint recognition. The Gaussian encoder ensures that the feature extraction output conforms to a Gaussian distribution N(μ,σ). It contains two independent sub-networks, each consisting of several fully connected layers. These two networks extract the mean and variance characteristics from the features, respectively, and unify them into z before outputting:

[0034] z = μ + β·exp(σ)

[0035] Where μ represents the mean characteristic extracted by one subnetwork, σ represents the variance characteristic extracted by another subnetwork, and β is an N(0,1) random number that conforms to a Gaussian distribution.

[0036] In the fourth step, adversarial learning is used for training, with two classifiers: the first classifier distinguishes authorized targets, and the second classifier distinguishes receivers. To improve the performance of the first classifier and achieve the goal of distinguishing the target to be identified, we minimize the loss function L. tx To achieve this goal, the network needs to be able to perform receiver transfer, making it impossible for the second classifier to distinguish the receivers. This can be achieved by introducing a gradient inversion layer. The gradient inversion layer ensures that the forward propagation of gradients is unaffected, while during backward propagation, the network parameters are updated in the opposite direction. The gradient update method is as follows:

[0037]

[0038] Where θ represents the network parameters, α is the learning rate, -λ is the gradient update coefficient caused by the gradient reversal layer, and L... tx Let L represent the loss function of the first classifier. rx Let L represent the loss function of the second classifier. tx and L rx Using the cross-entropy function, it can be expressed in the following form:

[0039]

[0040] Where y itx p represents the tag of the i-th transmitter. itx y represents the probability that the network prediction result is the i-th transmitter. jrx p represents the tag of the j-th receiver. jrx This represents the probability that the network prediction result is the j-th receiver.

[0041] Beneficial effects

[0042] 1. The present invention provides an RF fingerprinting method for receiver migration problems. This method introduces a demodulation and reconstruction approach during data processing, significantly enhancing the signal characteristics of the transmitter and receiver. This facilitates targeted extraction and deep feature mining of the transmitter's RF fingerprint in subsequent processing and also improves noise immunity.

[0043] 2. This invention introduces a Gaussian encoder after the feature extraction network, which makes it easier to discover latent features in the signal and eliminates interference from factors such as data distribution on feature extraction. At the same time, the random quantity introduced by the Gaussian encoder also enhances the network's noise resistance during training.

[0044] 3. The network trained by this invention can be directly transferred to an unknown receiver for radio frequency fingerprint recognition without needing to undergo retraining or other steps.

[0045] 4. This invention applies adversarial learning methods, designs two classification tasks, and uses gradient inversion layers to achieve network transferability.

[0046] 5. This invention introduces a demodulation and reconstruction method in data processing, which enhances key features in the signal without losing information compared to existing methods. This gives the training data superior features and allows the network to extract more useful information, thus improving recognition performance and noise resistance.

[0047] 6. This invention incorporates our designed Gaussian encoder into the network structure, which enables the discovery of more hidden information during feature extraction and reduces the impact of data spatial distribution on feature learning. The randomness introduced in the Gaussian encoder also improves the network's noise resistance. Attached Figure Description

[0048] Figure 1 The RFFI recognition accuracy is given under different signal-to-noise ratios. Detailed Implementation

[0049] The following describes in detail, with reference to the accompanying drawings and embodiments, a radio frequency fingerprinting method for addressing receiver migration problems according to the present invention:

[0050] A radio frequency fingerprinting method for receiver migration problems, the method comprising the following steps:

[0051] The first step involves selecting two receivers as receiving ends and seven transmitters as targets to be identified. A transmission sequence is set, using QPSK modulation, a code rate of 100 kBaud, and a carrier frequency of 350 kHz. Each transmitter transmits signals sequentially. Two receivers are used to receive the signals, with 10x oversampling and 1000 pairs of I / Q data per sample. Tags are then created based on the corresponding relationships to obtain the original database.

[0052] The second step is to demodulate and reconstruct each sample in the database to obtain the corresponding ideal signal, which is then spliced ​​with the original sample to form a new form, thus obtaining the database with enhanced features.

[0053] The third step is to design the network, select a convolution kernel size of (2,3) and a learning rate of 0.0001, add our designed Gaussian encoder after the deep residual shrinking network, and then connect two classifiers, with a gradient inversion layer added before the receiver classifier.

[0054] The fourth step involves applying adversarial learning methods, using the database obtained in the second step as the training set to train the network until it converges and stabilizes. The network model and parameters are then stored.

[0055] The fifth step involves selecting a new receiver to receive radio frequency signals from seven transmitters, and then applying the trained network to classify and identify the received signals. The resulting recognition effect is shown in the image below. Figure 1 ,according to Figure 1 It can be seen that the recognition accuracy of the present invention is above 60% in the signal-to-noise ratio range of 0dB to 20dB, and the performance at low signal-to-noise ratio is still good, while the recognition effect of the comparison method is poor at low signal-to-noise ratio.

[0056] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A radio frequency fingerprinting method for addressing receiver migration problems, characterized in that... The steps of this method include: The first step is to use two different receivers to collect the radio frequency signals of the authorized device, and store the collected radio frequency signals in a database as a training library for deep learning neural networks. The second step is to demodulate and idealize the digital reconstruction of the radio frequency signals in the database from the first step to obtain the original signals without transmitter and receiver characteristics, and then splice them with the corresponding received signals to form training samples. The third step is to design a neural network based on the characteristics and spatial distribution of the signal data. The fourth step involves using the training samples spliced ​​in the second step to train the network designed in the third step, and then using adversarial learning to iteratively optimize the network parameters to complete the radio frequency fingerprint recognition. In the first step, the radio frequency signal includes Road signals and road signal, Road signals and The formula for splicing road signals is: in, This indicates the received radio frequency signal. Indicates received road signal, Indicates received road signal, Indicates the length of the radio frequency signal; In the second step, the method for signal demodulation and idealized digital reconstruction is as follows: Step 2.1: Estimate the carrier frequency of the radio frequency signals in the database using the frequency center method. The method is described as follows: in, express middle The Values ​​and The A vector consisting of values , Indicates the length of the radio frequency signal. Indicates the sampling rate. It is the estimated carrier frequency; Step 2.2: Estimate the code rate of the radio frequency signal using the envelope squared spectrum method. First, analyze the radio frequency signal... Perform a Hilbert transform to convert it into an analyzable form: in Indicates the sampling time interval; That is, the estimated code rate; Step 2.3: Estimate the time difference between the sampling start position and the symbol start position in the radio frequency signal. It can be obtained using the following formula: Where L is the number of symbol periods, and G is the number of samples within a single symbol period; Step 2.4: Perform symbol determination on the symbols in the radio frequency signal; Step 2.5: Reconstruct the modulated signal by concatenating the ideal signal with the received signal to form a new sample for network training. The sample format is as follows: in, Represents the ideal after reconstruction road signal, Represents the ideal after reconstruction road signal; In the third step, the designed neural network is based on a deep residual shrinking network, in which a Gaussian encoder is embedded to obtain the deep hidden features of the RF fingerprint. The Gaussian encoder ensures that the output of the feature extraction conforms to a Gaussian distribution. It comprises two independent sub-networks, each consisting of several fully connected layers. These two networks extract the mean and variance characteristics from the features, respectively, and unify them using the following formula: Post-output: in This represents the mean characteristic extracted by a subnetwork. This represents the variance characteristics extracted by another subnetwork. It conforms to a Gaussian distribution. Random numbers; In the fourth step, the training adopts an adversarial learning method, setting up two classifiers: the first classifier is used to distinguish authorized targets, and the second classifier is used to distinguish receivers. The second classifier includes a gradient inversion layer, where gradient updates are performed as follows: in Represents parameters in the network. It's the learning rate. The gradient update coefficients are caused by the gradient reversal layer. This represents the loss function of the first classifier. This represents the loss function of the second classifier. and Using the cross-entropy function, it can be expressed in the following form: in Indicates the first The transmitter's label, This indicates that the network prediction result is the first... The probability of a transmitter. Indicates the first The tag of the receiver, This indicates that the network prediction result is the first... The probability of a receiver.

Images