A radio frequency fingerprint recognition method based on semi-supervised prototype network

Abstract

The present invention provides a radio frequency fingerprint recognition method based on a semi-supervised prototype network, which belongs to the field of Internet of Things security information technology. In a limited training set, it is impossible to include all categories of radiation source device signals, and unknown class samples may appear in the test set. When this situation occurs in a real scene, the trained closed set model does not have the ability to recognize unknown classes when recognizing unknown signals. Especially in some special environments, such as when a drone pretends to be an authenticated device, recognition errors may occur when radio frequency fingerprint recognition is performed without obtaining the signal of the device. In the radio frequency fingerprint recognition method based on a semi-supervised prototype network, the convolution layer is combined with the attention mechanism for effective feature extraction, and the improved semi-supervised method of time series integrated prediction is used to strengthen the prototype; the soft threshold selection method is used as the basis for sample classification, so that the model can effectively distinguish the differences between I / Q signal samples, thereby ensuring the accuracy of model recognition.

Description

Technical Field

[0001] The invention belongs to the technical field of Internet of Things security, and in particular relates to a radio frequency fingerprint recognition method based on a semi-supervised prototype network. Background Art

[0002] With the rapid development of technologies such as big data and artificial intelligence, various types of IoT devices are widely used, and the total number of communication equipment is also increasing rapidly, and information security issues are gradually arising. Radio frequency fingerprint recognition came into being to solve the complexity of various encryption and decryption technologies. It aims to identify communication devices from the inherent characteristics of components. At present, radio frequency fingerprint recognition has been used as an additional security layer for wireless devices. By analyzing the characteristics of electromagnetic waves emitted from the transmitter, radio frequency fingerprint technology can uniquely identify each wireless device, thereby avoiding deception or counterfeit attacks.

[0003] Biometric fingerprints are unique to the human body and can be used as biological features for identity authentication, such as fingerprints, irises, faces, and voices. These features are unique to each person and will not change or be lost. Among them, fingerprints are one of the most common and reliable biometric fingerprints, which are easy to collect and identify. Fingerprint recognition technology is an identity authentication technology. After extracting the target fingerprint, the target fingerprint is compared with the fingerprints still retained in the system to determine whether the target fingerprint matches any one in the fingerprint library. Similar to fingerprint recognition technology, radio frequency fingerprint recognition technology is based on the individual inherent characteristics of the hardware components of the wireless device terminal for identification. Just like a person's fingerprint, each wireless device terminal has its own unique hardware characteristics. During the manufacturing process, each device will have slight differences in hardware due to factors such as process and materials. These differences will directly affect the device's wireless signal transmission and reception capabilities, resulting in the unique representation of the radio frequency fingerprint of the device's transmitted signal, and these differences cannot be copied by other devices. In order to efficiently and safely identify and authenticate devices, researchers collect device transmission signals for feature extraction and compare them with these device fingerprint features to accurately identify the identity of the device. For example, radio frequency fingerprint recognition technology has been used in the ADS-B system for aircraft identification and classification. In addition, other wireless devices, such as Bluetooth devices, mobile phones, etc., can also be identified using the radio frequency fingerprint recognition method.

[0004] At present, the existing RF fingerprint recognition methods mainly adopt supervised learning methods, and their model training is a closed-world setting, that is, it is assumed that the unlabeled data only contains categories that have been encountered in the labeled training data. However, the real application environment is open and dynamic, and unknown categories may appear during test data or model deployment. The existing RF fingerprint recognition algorithms in closed-set scenarios are only suitable for the recognition of known types of devices and cannot recognize the signals of devices that did not appear in the training process. In the open set recognition problem, the existing RF fingerprint recognition algorithms have low robustness and low recognition performance. Summary of the invention

[0005] In view of the above problems, the present invention provides a radio frequency fingerprint recognition method based on a semi-supervised prototype network, which solves the problem that the existing radio frequency fingerprint recognition algorithm has low robustness and low recognition performance in the open set recognition problem.

[0006] The technical solution of the present invention is:

[0007] A radio frequency fingerprint recognition method based on a semi-supervised prototype network comprises the following steps:

[0008] S1. Acquire I / Q signal data and perform preprocessing to obtain training data, wherein the preprocessing method is to mark data of known categories, thereby dividing the training data into training data of known categories and training data of unknown categories;

[0009] S2. Construct a semi-supervised prototype network and use known category training data to obtain the initial prototype of each category of samples:

[0010]

[0011] Among them, n represents the total number of samples in a single category, y′ h is the original vector y of the sample h Multiply it by the weight of the sample point and the calculation method is:

[0012]

[0013] in, is the sum of the distances between the hth sample and all other samples in the same category. It is the sum of the distances between the i-th sample and all other samples in the same category.

[0014] Update the initial prototype using unknown category training data:

[0015]

[0016] Among them, z mrepresents the total number of samples of the mth category in training; S L Indicates a labeled dataset; S U represents an unlabeled dataset; y i represents the true label of the labeled data, Represents the pseudo-label obtained by integrating the predictions of unlabeled data 1(·) is an indicator function that outputs 1 when the input is judged to be true and outputs 0 when the input is judged to be false.

[0017] The temporal consistency loss function is used to train the entire model:

[0018]

[0019] Among them, w(t) is a function that changes with time and is used to control the weight influence of consistency loss. All samples participate in the calculation of consistency loss. During the training process, the task goal is to minimize the difference between the integrated prediction value and the current prediction, so that the network can obtain the prediction value as smoothly as possible, and the integrated prediction value will not fluctuate greatly due to a certain abnormal prediction value. At the same time, it can guide the model to learn more feature representations, and finally use the loss function to obtain the classifier.

[0020] S3. Use the obtained classifier to classify and identify the obtained I / Q signal data.

[0021] The beneficial effect of the present invention is that the present invention relates to a radio frequency fingerprint recognition method based on a semi-supervised prototype network. First, the calculation method of the traditional prototype initialization method is too simple and does not have good stability, which will reduce the recognition accuracy of the model. To solve this problem, the present invention designs a prototype initialization method with neighbor weighting to calculate a small part of labeled data, and designs a semi-supervised training method based on time series integrated prediction to further utilize a large amount of unlabeled data to strengthen prototype generation. The present invention integrates the historical information and current predictions of unlabeled samples, takes advantage of semi-supervised training, effectively reduces the cost of excessive manual labeling of samples, and reduces the impact of incorrectly labeled samples during model training, thereby improving the recognition accuracy and stability of the model. The semi-supervised method is applied to radio frequency fingerprint recognition technology, and a radio frequency fingerprint recognition method based on a semi-supervised prototype network is studied. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the main framework of the semi-supervised prototype network of the present invention.

[0023] Figure 2 This is a flow chart of network task training of the present invention.

[0024] Figure 3 It is a structural diagram of the feature extractor of the present invention.

[0025] Figure 4 Flowchart for initializing the neighbor weighted prototype of the present invention. DETAILED DESCRIPTION

[0026] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0027] The present invention is based on a semi-supervised prototype network for radio frequency fingerprint recognition. The main framework of the semi-supervised prototype network is as follows: Figure 1 shown.

[0028] Example

[0029] like Figure 2 As shown, the present invention provides a radio frequency fingerprint recognition method based on a semi-supervised prototype network, characterized in that it includes the following steps:

[0030] The I / Q signal data of a task is divided into labeled data and unlabeled data as input, which is fed into the model for training, and a feature extractor is used to calculate the feature vector of each sample.

[0031] A single I / Q sample is divided into multiple consecutive 2×512-sized I / Q data (the training set contains m categories) as input, and the number of channels is 1. The initial input size of the convolutional layer is 2×512×channels, where channels represents the number of channels. Figure 3 The convolutional neural network shown extracts the RF fingerprint features of the signal and finally obtains the feature map of the signal in the high-dimensional feature space.

[0032] In summary, during the training process, for each training sample, after being processed by the convolutional feature extraction layer, an effective feature vector of the sample can be obtained; after all features are extracted, the feature representation of all training samples is obtained. The specific convolution kernel sizes used are shown in the following table.

[0033] CNN convolution parameter table

[0034]

[0035] like Figure 4 As shown in , in order to obtain the weight of the nearest neighbor weighting, it is necessary to calculate the distance between samples belonging to the same category and samples belonging to other categories. The sum of the distances between data points can be obtained by summing them up. If the sum of the distances between a data point and other data points is large, it means that the position of the data point is far away from the other data points, indicating that the data point is located in an area with sparse distribution of the same type, and its weight should be appropriately reduced. On the contrary, if the sum of the distances is small, it means that the position of the data point is close to the other data points, and its weight needs to be increased.

[0036] The formula for calculating the distance between similar samples is as follows.

[0037] s ij =||y i -y j ||,i,j∈[1,n]

[0038] And the formula for summing the distances to all other sample points is as follows.

[0039]

[0040] For S h The normalization process is shown in the formula.

[0041]

[0042] After that, the weighted value is calculated, and its calculation expression is as follows.

[0043]

[0044] Finally, the initial prototype of each type of sample is obtained, and its mathematical expression is shown below.

[0045]

[0046] The key to using semi-supervised training to strengthen the initial prototype is to use unlabeled data to improve the classifier's prediction of the prototype for each category. Since unlabeled data provides more information, it can help estimate the location of the prototype more accurately. Therefore, unlabeled data can be used to update the prototype to improve the performance of the classifier. In order to further improve the prediction accuracy of unlabeled data, the model can accumulate prediction information of unlabeled data during semi-supervised training.

[0047] The specific update calculation of integrated prediction is shown in the following formula.

[0048]

[0049] While using labeled samples to calculate the class prototype, a large amount of unlabeled data is used to update the class prototype by using the integrated predicted values ​​as pseudo labels for the unlabeled data. The specific updates are shown below.

[0050]

[0051] In this embodiment, considering the stability of the semi-supervised training process, a temporal consistency loss function is defined as follows.

[0052]

[0053] w(t) represents a function that changes over time, as shown in formula (3-20), and its purpose is to control the weight impact of consistency loss.

[0054]

[0055] At the beginning of training, the network parameters are random and cannot be well fitted. A w(t) value that is too large will affect the normal training of the network. Therefore, w(t) is designed to increase slowly over time.

[0056] All samples participate in the consistency loss calculation. During the training process, the task goal is to minimize the difference between the integrated prediction value and the current prediction, so that the network can obtain the prediction value as smoothly as possible, without causing large fluctuations in the integrated prediction value due to a certain abnormal prediction value. At the same time, it can guide the model to learn more feature representations. Unlabeled data can not only be used to improve the refinement process of category prototypes, but the model can also incorporate and update the prediction results of unlabeled samples according to the category distribution obtained during the training process to complete integration and updating; this can effectively reduce the time cost of manual labeling and improve the anti-interference ability of prototype generation and reduce the impact of incorrect labeling.

Claims

1. A radio frequency fingerprint recognition method based on a semi-supervised prototype network, characterized in that: The following steps are involved: S1. Acquire I / Q signal data and perform preprocessing to obtain training data, wherein the preprocessing method is to mark data of known categories, thereby dividing the training data into training data of known categories and training data of unknown categories; S2. Construct a semi-supervised prototype network and use known category training data to obtain the initial prototype of each category of samples: Among them, n represents the total number of samples in a single category, y′ h is the original vector y of the sample h Multiply it by the weight of the sample point and the calculation method is: in, is the sum of the distances between the hth sample and all other samples in the same category. It is the sum of the distances between the i-th sample and all other samples in the same category; Update the initial prototype using unknown category training data: Among them, z m represents the total number of samples of the mth category in training; S L Indicates a labeled dataset; S U represents an unlabeled dataset; y i represents the true label of the labeled data, Represents the pseudo-label obtained by integrating the predictions of unlabeled data 1(·) is an indicator function that outputs 1 when the input is judged to be true and outputs 0 when the input is judged to be false; The temporal consistency loss function is used to train the entire model: Among them, w(t) is a function that changes with time and is used to control the weight influence of consistency loss. All samples participate in the calculation of consistency loss. During the training process, the task goal is to minimize the difference between the integrated prediction value and the current prediction, so that the network can obtain the prediction value as smoothly as possible, and the integrated prediction value will not fluctuate greatly due to a certain abnormal prediction value. At the same time, the model is guided to learn more feature representations, and finally the classifier is obtained using the loss function; S3. Use the obtained classifier to classify and identify the obtained I / Q signal data.

Images