A radio frequency fingerprinting method based on homodyne quadrature signals and contrastive learning

Abstract

A radio frequency fingerprint identification method based on homophase orthogonal signal and contrast learning, data enhancement is performed on selected data, and the data is cropped to a fixed length to give positive and negative sample pairs; the contrast learning model is based on a twin network, each branch network includes a feature extraction network and a feature projection network, during training, the same signal is taken as the positive sample and the other signal is taken as the negative sample on the contrast learning model, the two branch networks are respectively input into the positive and negative sample pairs, and the feature projection network is used for feature transformation projection to the representation space, in application, the structure and parameters of the feature extraction network of branch network A are taken, and are encapsulated into an application scene as a radio frequency fingerprint extractor, radio frequency signals of access equipment are collected, radio frequency fingerprints are extracted and compared with radio frequency fingerprints in a database, equipment authentication or intrusion detection is completed; it is not dependent on a specific data processing method, and the hardware and time overhead of preprocessing is reduced.

Description

Technical Field

[0001] This invention relates to the field of information security and the application of artificial intelligence in this field, and in particular to a radio frequency fingerprint recognition method based on orthogonal signals in the same direction and contrastive learning. Background Technology

[0002] Radio frequency fingerprints refer to unique and distinctive subtle features present in radio frequency signals from different radiation sources. These features persist in radio frequency devices from the same batch and using the same signal. By extracting these characteristic fingerprints, radio frequency devices can be identified, authenticated, or authorized.

[0003] Current radio frequency fingerprint recognition methods are mainly divided into two categories: feature engineering methods and deep learning methods.

[0004] Feature engineering methods rely on domain-specific expertise. They utilize prior knowledge from experts to exploratoryly design features manually through data analysis, then employ support vector machines, decision trees, or shallow neural networks to perform recognition. These expert-designed features typically have clear physical meaning, good interpretability, and can achieve high recognition accuracy in specific systems. However, feature engineering methods depend on expert knowledge, are time-consuming and labor-intensive, are sensitive to noise, have poor generalization ability, and the designed features are difficult to apply to other scenarios.

[0005] Existing deep learning methods, represented by supervised learning, rely on massive amounts of labeled data and automatically extract features through deep neural networks to complete RFID fingerprint identification. Deep learning methods typically possess good feature extraction capabilities without requiring domain expert knowledge. However, RFID data acquisition and simulation are difficult, and data labeling is time-consuming and labor-intensive, making it difficult to meet the massive data requirements for model training. Some deep learning-based RFID fingerprint extraction methods use time-domain and frequency-domain transformations to convert in-phase / orthogonal signals into images, which is difficult to meet real-time processing requirements. Furthermore, deep learning methods are closely tied to training data, making it difficult to apply models trained on specific datasets to other cross-domain scenarios.

[0006] For the reasons mentioned above, a radio frequency fingerprint recognition method based on orthogonal signals in the same direction and contrastive learning has been developed. Summary of the Invention

[0007] The purpose of this invention is to provide a radio frequency fingerprint recognition method based on co-directional orthogonal signals and contrastive learning to solve the above-mentioned problems. This method directly extracts radio frequency fingerprints using co-directional and orthogonal signals, without relying on specific data processing methods, thus reducing the hardware and time overhead of preprocessing. It trains a deep learning model through contrastive learning, without relying on specific datasets or requiring data annotation. This invention has good transferability, can adapt to different radio frequency fingerprint application requirements, and can achieve good recognition results in all cases.

[0008] This invention provides a radio frequency fingerprinting method based on co-directional orthogonal signals and contrastive learning, which constructs a dataset: using several publicly available or self-collected and simulated datasets based on co-directional orthogonal signals from the Internet, a contrastive learning dataset is constructed;

[0009] Building a data loader: Randomly select data from several datasets, add noise, jitter, phase shift, and frequency shift to the selected data, and then randomly prune it to a fixed length. The data loader can provide positive and negative sample pairs.

[0010] Constructing a contrastive learning model: The contrastive learning model is based on Siamese networks, which consist of branch networks A and B. The two branch networks have identical network structures. Each branch network contains a feature extraction network and a feature projection network. The feature extraction network consists of module components for processing one-dimensional time series signals, and the feature projection network consists of two fully connected layers.

[0011] During training, in the contrastive learning model, signal samples from the same signal data but processed by different data augmentation methods are used as positive samples, and signal samples from different signal data are used as negative samples. During forward propagation, the two branch networks are input with positive and negative sample pairs respectively, and then the feature projection network performs feature transformation and projection onto the representation space. The contrastive loss is calculated in the representation space. Through the constraint of the contrastive loss, the distance between different signals in the representation space is made as far as possible. During backpropagation, the parameters of branch network A are updated, and the parameters of branch network B are indirectly updated from the parameters of branch network A using a parameter sharing method with momentum.

[0012] In application, the structure and parameters of the branch network A feature extraction network are taken and encapsulated into the application scenario as an RF fingerprint extractor. During the deployment phase, the RF signal of the access device is collected, and the RF fingerprint is extracted and stored in the database by the RF fingerprint extractor. During the usage phase, when the RF signal is detected, the RF fingerprint is extracted and compared with the RF fingerprint in the database to complete device authentication or intrusion detection.

[0013] Preferably, the module components for processing one-dimensional time series signals include a one-dimensional convolutional neural network, a recurrent neural network, and a Transformer.

[0014] The beneficial effects are as follows: This invention directly extracts radio frequency fingerprints using co-directional and orthogonal signals, without relying on specific data processing methods, thus reducing the hardware and time overhead of preprocessing; it trains deep learning models through comparative learning, without relying on specific datasets or requiring data annotation; this invention has good transferability, can adapt to different radio frequency fingerprint application needs, and can achieve good recognition results in all of them; the parts not described in detail in this invention are commonly used existing technologies. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a diagram of a contrastive learning model using Siamese networks;

[0017] Figure 2 This is a case study of electronic device identity authentication based on radio frequency fingerprinting;

[0018] Figure 3 This is a diagram illustrating an application scenario for electronic device identity authentication. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] See Figures 1-3 A radio frequency fingerprint recognition method based on orthogonal signals in the same direction and contrastive learning.

[0021] Construct a dataset; use several publicly available or self-collected and simulated datasets based on in-direction and orthogonal signals from the Internet to construct a contrastive learning dataset;

[0022] Construct a data loader; randomly select data from several datasets according to a certain proportion; perform data augmentation on the selected data, adding noise, jitter, phase shift, and frequency shift, and then randomly prune it to a fixed length; the data loader can provide positive and negative sample pairs; a positive sample comes from two signal samples from the same signal data processed by different data augmentation methods; a negative sample comes from two signal samples from different signal data.

[0023] Construct a contrastive learning model; the contrastive learning model is based on Siamese networks, which consist of branch networks A and B, and the two branch networks have identical network structures; each branch network contains a feature extraction network and a feature projection network; the feature extraction network consists of modules that can process one-dimensional time series signals, including but not limited to one-dimensional convolutional neural networks, recurrent neural networks, and Transformers, and the feature projection network consists of two fully connected layers;

[0024] Training process: Use a data loader to extract N data points from the unlabeled dataset to form a batch of training data; for any data point x in this batch, use two data augmentation methods to obtain xi and xj, xi and xj form a positive sample pair; xi and the other 2N-2 data points form a negative sample pair;

[0025] Positive and negative sample pairs are input into a feature extraction network to extract features, which are then transformed and projected into a representation space via a feature projection network. The distance between positive and negative sample pairs is then calculated within the representation space. A loss function is designed to ensure that the projected distances between positive samples are closer and those between negative samples are farther. Backpropagation is then used to update the parameters of branch network A, while the parameters of branch network B are indirectly updated from the parameters of branch network A using a parameter sharing method with driving forces. The loss function... As shown below:

[0026] ,

[0027] .

[0028] in, This is a metric function used to measure the distance between two features in the representation space. Commonly used metric methods include Euclidean distance, cosine similarity, and Mahalanobis distance. For temperature parameters, the measured values ​​are scaled, with experimental values ​​ranging from 0.05 to 0.1. For the indicator function, The value of this item is 1 if it is true, and 0 otherwise.

[0029] Specific applications: Taking electronic device authentication based on radio frequency fingerprints as an example, a feature extraction network of a twin network is used as the radio frequency fingerprint extractor. For a given access device, radio frequency signals under different states are collected multiple times, input into the radio frequency fingerprint extractor, and its output is stored as the device's radio frequency fingerprint in the legitimate fingerprint database. During the device authentication phase, the video signal of the device to be authenticated is collected, and its radio frequency fingerprint is extracted using the radio frequency fingerprint extractor and compared with the radio frequency fingerprint in the access device fingerprint database. If the device's radio frequency fingerprint meets the tolerance, authentication is successful; otherwise, authentication fails. The tolerance can be obtained during training or adjusted through multiple tests in specific application scenarios. Training is performed on the RadioML2018.01A dataset and the COSINE dataset, and testing is conducted on a self-built communication signal dataset, with a test accuracy of no less than 97.5%.

[0030] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A radio frequency fingerprint recognition method based on orthogonal signals in the same direction and contrastive learning, characterized in that: Dataset construction: Construct a contrastive learning dataset using several publicly available or self-collected and simulated datasets based on co-directional and orthogonal signals; Building a data loader: Randomly select data from several datasets, add noise, jitter, phase shift, and frequency shift to the selected data, and then randomly prune it to a fixed length. The data loader can provide positive and negative sample pairs. Constructing a contrastive learning model: The contrastive learning model is based on Siamese networks, which consist of branch networks A and B. The two branch networks have identical network structures. Each branch network contains a feature extraction network and a feature projection network. The feature extraction network consists of module components for processing one-dimensional time series signals, and the feature projection network consists of two fully connected layers. During training, in the contrastive learning model, signal samples from the same signal data but processed by different data augmentation methods are used as positive samples, and signal samples from different signal data are used as negative samples. During forward propagation, the two branch networks are input with positive and negative sample pairs respectively, and then the feature projection network performs feature transformation and projection onto the representation space. The contrast loss is calculated in the representation space. Through the constraint of the contrast loss, the distance between different signals in the representation space is made as far as possible. During backpropagation, the parameters of branch network A are updated normally, while the parameters of branch network B are indirectly updated from the parameters of branch network A using a parameter sharing method with driving momentum. In application, the structure and parameters of the branch network A feature extraction network are taken and encapsulated into the application scenario as an RF fingerprint extractor. During the deployment phase, the RF signal of the access device is collected, and the RF fingerprint is extracted and stored in the database by the RF fingerprint extractor. During the usage phase, when the RF signal is detected, the RF fingerprint is extracted and compared with the RF fingerprint in the database to complete device authentication or intrusion detection.

2. The radio frequency fingerprint recognition method based on orthogonal signals and contrastive learning according to claim 1, characterized in that: The module components for processing one-dimensional time series signals include a one-dimensional convolutional neural network, a recurrent neural network, and a Transformer.

Images