CN110069969B - An Authentication Fingerprint Recognition Method Based on Pseudo-Random Integration - Google Patents

Abstract

The invention discloses an authentication fingerprint identification method based on pseudorandom integration, which comprises the following steps: s1, a receiving device carries out signal acquisition and storage on a sending device to obtain N signal samples and stores the N signal samples in a sample library; s2, enhancing N signal samples in a sample library into K signal samples by adopting pseudo-random integration; s3, performing feature extraction on the generated signal sample; s4, repeating the operations from S1 to S3 for a plurality of different transmitting devices, and taking the device number as feedback; training the signals processed by each sending device by using a classification algorithm in machine learning to obtain a classifier; and S5, classifying the waveform to be detected by using a classifier, and judging the equipment to which the waveform belongs. The invention uses a data enhancement mode based on pseudo-random integration, and increases the training data amount for machine learning; meanwhile, machine learning is combined, and instability of the radio frequency fingerprint identification technology based on random integration and machine learning in the identification rate is reduced.

Description

An Authentication Fingerprint Recognition Method Based on Pseudo-Random Integration

technical field

The invention relates to the field of wireless device physical layer access authentication, in particular to an authentication fingerprint identification method based on pseudo-random integration.

Background technique

Access authentication of end nodes is an important and challenging issue for IoT security. Radio frequency fingerprinting is a promising solution to this problem, extracting fingerprints to perform end node access authentication by extracting signal changes due to hardware manufacturing defects. Machine learning can complete radio frequency fingerprinting very well through the training of a large amount of data.

The machine learning performance of RF fingerprints is greatly affected by the amount of training data, and a large amount of RF fingerprint data is required to obtain better training and recognition results. In some cases, data is so precious that it is not easy to obtain. Data augmentation is the most common way to get better results with limited data. Data augmentation has a large impact on the final classification performance and generalization ability of the system, which generates additional pseudo-instances by applying a "label-invariant" transformation to the training instances. Current research provides a variety of different data augmentation strategies, and most of them work in the field of image recognition. Commonly used image enhancement strategies are briefly summarized as follows:

1. Scale transformation: Modify the size of the image or scale the image to a certain ratio.

2. Position transformation: flip, rotate and translate, that is, flip the image in the horizontal or vertical direction. Rotate the image by an angle. Move the image to a certain distance.

3. Intercept or delete. Cut out part of the image, delete or keep this part.

4. Color jitter. Change the brightness, saturation or contrast of the image.

5. PCA jitter. Extract the principal components of the image by principal component analysis, and then add a (0,0.1) Gaussian perturbation.

6. Noise. Pixel points are randomly obtained and then set to high brightness and low gray level to simulate additive noise.

Strategies 1 to 4 are suitable for image enhancement, which cannot be used directly for one-dimensional sequences such as sound or signals. Strategies 5 and 6 can be used for signal classification, but they add extra noise to the signal samples with high signal-to-noise ratio, which will reduce the final classification accuracy. Existing studies have shown that when the signal-to-noise ratio decreases, the accuracy of RF fingerprinting will drop rapidly. Therefore, the current mainstream data enhancement methods are not completely feasible in the main signal processing, and cannot be directly used for RF fingerprint identification.

Salamon and Bello's 2017 study provides a series of data augmentation strategies for sound classification, which are closely related to the signal, as follows: time extension: keep frequency constant, increase or decrease the rate of samples; frequency change: keep The rate remains the same, increasing or decreasing the frequency of the samples; dynamic range compression: some small signals are amplified as needed, while large signals remain unchanged. These strategies can indeed be used for signal processing. However, the approximation of the signals in RFID fingerprinting is very high. A signal of the same type will differ from its transformed signal more than that signal differs from another signal. A key concept of data augmentation is that deformations applied to the data cannot change the meaning of the labels, so these strategies are also not suitable for RFID fingerprinting.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the deficiencies of the prior art and provide an authentication fingerprint identification method based on pseudo-random integration, which increases the amount of training data available for machine learning by using the data enhancement method based on pseudo-random integration; Learning reduces the instability of the recognition rate of the radio frequency fingerprint recognition technology based on random integration and machine learning.

The object of the present invention is to be achieved through the following technical solutions: a kind of authentication fingerprint identification method based on pseudo-random integration, comprising the following steps:

S1. The receiving device collects and stores signals from the sending device, and obtains N signal samples and stores them in the sample library;

S2. Use pseudo-random integration to enhance the N signal samples in the sample library into K signal samples;

S3. Feature extraction is performed on the generated signal samples;

S4. For a plurality of different transmitting devices, repeat the above operations S1 to S3, and use the device number as feedback; and use the classification algorithm in the machine learning algorithm to train the signals processed by each transmitting device to obtain a classifier;

S5. Use the classifier obtained by training to classify the waveform to be detected, and determine the equipment to which it belongs.

Specifically, the step S1 includes:

S101. The receiving device detects the position of the starting point of the received power-on transient signal;

S102. Starting from the starting point, collect M power-on transient signal sample points of a sending device as a signal sample;

S103. In the signal samples, number the power-on transient signal sample points, and define the power-on transient signal sample point amplitude function f _i with the amplitude corresponding to each power-on transient signal sample point:

f _i ={amp:(1,2,...,M)}

={amp ₁ ,amp ₂ ,...,amp _M };

S104. Collect N signal samples according to steps S101-S103 and store them in the sample library.

Specifically, the step S2 performs coherent or non-coherent accumulation of the amplitude functions of multiple power-on transient signal sample points of the same sending device by pseudo-random selection, and uses limited signal samples to generate a large number of training machines for machine learning. data, including:

S201. Randomly select n signal samples from the sample library, n<N;

S202. Starting from the second selection of signal samples, check whether the selected n samples are exactly the same as the previous selection: if so, go to step S204; if not, go to step S203;

S203. The n samples are accumulated coherently or non-coherently, and stored in the database:

where _fr is the original signal sample in the library, and f _a is the generated signal sample;

S204. Remove the signal samples participating in the integration from the sample library;

S205. Repeat steps S201 to S204 until the number of samples in the signal library is less than n;

S206. Initialize the signal library and reset it to the original state;

S207. Repeat S201 to S206 until the number of signal samples in the database reaches K;

S208. Output the signal samples in the database.

Specifically, the step S3 includes:

S301. For each signal sample in the database, firstly perform the folding process, that is, take the absolute value of the amplitude value:

f _avi = | f _i |;

S302. Normalize the folded signal samples:

In the formula, amp _i represents the amplitude of the i-th sample point in the signal sample, i=1,2,...,M, amp _max represents the amplitude of the largest sample point in the signal sample, and amp _min represents the smallest sample point in the signal sample. The sample point amplitude of ;

S303. Perform feature extraction on the normalized signal samples;

S304. Repeat steps S301 to S303 until the processing of the K signal samples in the database is completed, and the characteristic data of the K signal samples is obtained.

Preferably, in step S1, the subsequent running rate can be improved by means of vectorization, that is, the collected signal samples are uniformly stored in the form of a matrix:

Preferably, in the step S4, in order to improve the running rate, the feedback value can also be vectorized, that is:

Among them, _yi is the feedback value corresponding to the sample f _i , and Y is the feedback storage matrix.

Preferably, further, in the step S4, in order to facilitate machine learning, the characteristic data and the feedback value can be stored correspondingly in a unified manner, that is:

Among them, train is the machine learning training matrix.

Preferably, the classification algorithm in the machine learning algorithm includes k-nearest neighbor algorithm, naive Bayes algorithm, SVM algorithm and decision tree algorithm.

Preferably, in the step S101, the method for detecting the position of the starting point includes absolute amplitude value detection and slope detection.

Preferably, the feature extraction method in the step S3 includes feature extraction based on amplitude, feature extraction based on amplitude and phase combination, feature extraction based on frequency, feature extraction based on phase shift, feature extraction based on wavelet transform, and Feature extraction based on multi-resolution analysis; wherein the selection of mother wavelets in the multi-resolution analysis includes

haar, dB2, bior and morl.

The beneficial effects of the present invention are as follows: (1) The present invention utilizes the method of random integration to enhance the data of the radio frequency fingerprint, and utilizes limited signal samples to generate a large number of signal samples, thereby improving the training data available for machine learning and reducing overfitting. , which improves the recognition accuracy to a certain extent. (2) The present invention uses a pseudo-random selection method, which reduces the randomness of the RF fingerprint sample selection and reduces the instability of the recognition accuracy. (3) The present invention performs data collection and feature extraction on multiple transmitting devices, and uses the device number as feedback to realize the training of the classifier and the judgment of the waveform to be detected, so that each device can maintain a high recognition accuracy rate , which improves the reliability of the recognition results.

Description of drawings

Fig. 1 is the method flow chart of the present invention;

2 is a flowchart of data augmentation based on pseudo-random integration;

3 is a comparison diagram of the recognition rates of pseudo-random integration, random integration and signals without data enhancement under each signal-to-noise ratio in the present invention;

FIG. 4 is a comparison diagram of the recognition rate of a single device by pseudo-random integration, random integration, and signals without data enhancement in the present invention.

Detailed ways

The technical solutions of the present invention are further described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the following.

As shown in Figure 1, an authentication fingerprint identification method based on pseudo-random integration includes the following steps:

S1. The receiving device collects and stores signals from the sending device, and obtains N signal samples and stores them in the sample library;

S2. Use pseudo-random integration to enhance the N signal samples in the sample library into K signal samples;

S3. Feature extraction is performed on the generated signal samples;

S4. For a plurality of different transmitting devices, repeat the above operations S1 to S3, and use the device number as feedback; and use the classification algorithm in the machine learning algorithm to train the signals processed by each transmitting device to obtain a classifier;

S5. Use the classifier obtained by training to classify the waveform to be detected, and determine the equipment to which it belongs.

Specifically, the step S1 includes:

S101. The receiving device detects the position of the starting point of the received power-on transient signal;

S102. Starting from the starting point, collect M power-on transient signal sample points of a sending device as a signal sample;

For example, when M is 800, 800 power-on transient signal sample points can be collected from the starting point as one signal sample; in other embodiments, the starting point can also be collected starting from 100 positions before the starting point The first 100 power-on transient signal sample points and the 700 power-on transient signal sample points after the starting point are taken as a signal sample.

S103. In the signal samples, number the power-on transient signal sample points, and define the power-on transient signal sample point amplitude function f _i with the amplitude corresponding to each power-on transient signal sample point:

f _i ={amp:(1,2,...,M)}

={amp ₁ ,amp ₂ ,...,amp _M };

S104. Collect N signal samples according to steps S101-S103 and store them in the sample library.

As shown in FIG. 2 , the step S2 performs coherent or non-coherent accumulation of the amplitude functions of multiple power-on transient signal sample points of the same sending device through pseudo-random selection, and uses limited signal samples to generate a large number of available machines. Learned training data, including:

S201. Randomly select n signal samples from the sample library, n<N;

S202. Starting from the second selection of signal samples, check whether the selected n samples are exactly the same as the previous selection: if so, go to step S204; if not, go to step S203;

S203. The n samples are accumulated coherently or non-coherently, and stored in the database:

where _fr is the original signal sample in the library, and f _a is the generated signal sample;

S204. Remove the signal samples participating in the integration from the sample library;

S205. Repeat steps S201 to S204 until the number of samples in the signal library is less than n;

S206. Initialize the signal library and reset it to the original state;

S207. Repeat S201 to S206 until the number of signal samples in the database reaches K;

S208. Output the signal samples in the database.

Specifically, the step S3 includes:

S301. For each signal sample in the database, firstly perform the folding process, that is, take the absolute value of the amplitude value:

f _avi = | f _i |;

S302. Normalize the folded signal samples:

In the formula, amp _i represents the amplitude of the i-th sample point in the signal sample, i=1,2,...,M, amp _max represents the amplitude of the largest sample point in the signal sample, and amp _min represents the smallest sample point in the signal sample. The sample point amplitude of ;

S303. Perform feature extraction on the normalized signal samples;

S304. Repeat steps S301 to S303 until the processing of the K signal samples in the database is completed, and the characteristic data of the K signal samples is obtained.

In the step S101, the method for detecting the starting point position includes absolute amplitude value detection and slope detection; in the embodiment of the present application, absolute amplitude value detection is adopted: the threshold value is set to 0.003, and when the absolute value of the signal amplitude is greater than 0.003, A total of 3000 sample points (M=3000) are sampled respectively 800 before and 2200 after the point. In this embodiment, the subsequent running rate can be improved by means of vectorization, that is, the collected signal samples are stored uniformly in the form of a matrix:

The feature extraction methods in the step S3 include amplitude-based feature extraction, amplitude- and phase-based feature extraction, frequency-based feature extraction, phase-shift-based feature extraction, wavelet transform-based feature extraction, and multiresolution-based feature extraction. Feature extraction of rate analysis; wherein the choice of mother wavelet in the multi-resolution analysis includes haar, dB2, bior and morl.

In the embodiment of this patent, two-level multi-resolution analysis is adopted, and the dB2 waveform function is used as the mother wavelet, and two discrete wavelet transforms are sequentially performed on the signal waveform:

f _a1i =DWT(f _avi ,dB2)

f _a2i =DWT(f _a1i ,dB2)

specifically,

In the embodiment of the present application, in order to improve the running rate, the feedback value of the step S4 can also be vectorized, that is:

Among them, y _i is the feedback value corresponding to the sample f _i , and Y is the feedback storage matrix; further, in order to facilitate machine learning, the feature data and the feedback value can be stored in a unified manner, that is:

Among them, train is the machine learning training matrix.

The classification algorithms in the machine learning algorithm include k-nearest neighbor algorithm, naive Bayes algorithm, SVM algorithm and decision tree algorithm.

In the embodiment of this patent, step S4 adopts Gaussian kernel SVM to generate a classifier, specifically: the first step is to input a training matrix with feedback values

In the second step, the first l _train elements of each row of the train matrix are characterized, and the last element of each row is the feedback value, and the Gaussian kernel SVM is used for training, specifically:

st,α _i ≥0,i=1,...,n

为高斯核函数，通过调整高斯核尺度σ的大小，使得SVM与训练数据相匹配，防止产生欠拟合或过拟合。in,

For the Gaussian kernel function, by adjusting the size of the Gaussian kernel scale σ, the SVM is matched with the training data to prevent under-fitting or over-fitting.

In order to verify the difference between the pseudo-random integration of the present invention and the traditional random integration, the following four groups of samples were generated to carry out radio frequency fingerprint identification experiments:

(a) Each sending device collects 20 samples (N=20), selects 3 samples each time for integration (n=3), and uses random integration to generate 500 samples (K=500);

(b) Each sending device collects 20 samples (N=20), selects 3 samples each time for integration (n=3), and uses pseudo-random integration to generate 500 samples (K=500);

(c) 20 samples were collected from each sending device for comparison. In order to ensure the reliability of the experiment, the comparison sample adopts ordered coherent accumulation (n=3). This is because coherent accumulation has a noise reduction function, and the above approach will improve the signal-to-noise ratio to the same extent;

(d) Each sending device collects 100 samples, and uses ordered coherent accumulation (n=3) to generate 100 samples for comparison.

The experimental results are shown in Figure 3 and Figure 4, specifically:

1) At low SNR, the recognition result of random integration (a) is better than the ordered coherent accumulation of 20 samples (c), similar to the ordered coherent accumulation of 100 samples (d);

2) Under high signal-to-noise ratio, the classification accuracy of random integration (a) becomes unstable, and cannot achieve the accuracy of orderly integration (c) (d);

3) Although pseudo-random integration (b) appears jittery in the classification accuracy curve, it works better. Under high signal-to-noise ratio, its classification accuracy is stable above 99%;

4) For a single device, the pseudo-random integration (b) has a recognition rate of more than 98% for 10 devices, and the stability is better than the other three methods (a) (c) (d).

To sum up, the present invention collects the signals of the sending device, and generates a large amount of training data by using a small number of collected signal samples. Use the generated training data to train based on the classification algorithm in machine learning to obtain a classifier, and then use the test set to test the performance of the classifier. It is suitable for RF fingerprint recognition scenarios with insufficient training data, and has the advantages of high recognition accuracy, stability and reliability.

The above are preferred embodiments of the present invention, and it should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as an exclusion of other embodiments, but can be used in other combinations, modifications and environments, and can be used herein. Within the scope of the stated concept, modifications can be made through the above teachings or skill or knowledge in the relevant field. However, modifications and changes made by those skilled in the art do not depart from the spirit and scope of the present invention, and should all fall within the protection scope of the appended claims of the present invention.

Claims (7)

1. An authentication fingerprint identification method based on pseudo-random integration is characterized in that: the method comprises the following steps:

s1, a receiving device carries out signal acquisition and storage on a sending device to obtain N signal samples and stores the N signal samples in a sample library;

s2, enhancing N signal samples in a sample library into K signal samples by adopting pseudo-random integration;

in step S2, performing coherent or non-coherent accumulation on multiple start-up transient signal sample point amplitude functions of the same transmission device through pseudo-random selection, and generating a large amount of training data for machine learning by using limited signal samples, including:

s201, randomly selecting N signal samples from a sample library, wherein N is less than N;

s202. starting from the second selection of signal samples, it is detected whether the selected n samples are identical to the previous selection:

if yes, jumping to step S204;

if not, go to step S203;

s203, the n samples are subjected to coherent accumulation or non-coherent accumulation and stored in a database:

wherein f is_rIs a sample of the original signal in the library, f_aTo the generated signal samples;

s204, removing the signal samples participating in integration from a sample library;

s205, repeatedly executing the steps S201 to S204 until the number of samples in the signal library is less than n;

s206, initializing a signal library, and resetting the signal library to be in an original state;

s207, repeatedly executing S201 to S206 until the number of signal samples in the database reaches K;

s208, outputting a signal sample in the database;

s3, performing feature extraction on the generated signal sample;

s4, for a plurality of different transmitting devices, repeating the operations from S1 to S3, and taking the device number as feedback; training the signals processed by each sending device by using a classification algorithm in a machine learning algorithm to obtain a classifier;

and S5, classifying the waveform to be detected by using the classifier obtained by training, and judging the equipment to which the waveform belongs.

2. The authentication fingerprint identification method based on the pseudo-random integration according to claim 1, wherein: the step S1 includes:

s101, detecting the starting point position of a received starting transient signal by receiving equipment;

s102, collecting M starting transient signal sample points of a sending device from a starting point position as a signal sample;

s103, numbering starting transient signal sample points in the signal samples, and defining a starting transient signal sample point amplitude function f by using the corresponding amplitude of each starting transient signal sample point_i，：

f_i＝{amp:(1,2,……,M)}

＝{amp₁,amp₂,……,amp_M}；

S104, collecting N signal samples according to the steps S101-S103 and storing the N signal samples in a sample library.

3. The authentication fingerprint identification method based on the pseudo-random integration according to claim 2, wherein: the step S3 includes:

s301, for each signal sample in the database, firstly carrying out folding processing, namely taking an absolute value of an amplitude value:

f_avi＝|f_i|；

s302, normalization processing is carried out on the folded signal sample:

in the formula, amp_iRepresents the ith sample point amplitude in the signal sample, i 1,2_maxRepresenting the maximum sample point amplitude, amp, in the signal sample_minRepresenting the smallest sample point amplitude in the signal sample;

s303, performing feature extraction on the signal sample after normalization processing;

s304, repeating the steps S301 to S303 until the K signal samples in the database are processed, and obtaining the characteristic data of the K signal samples.

4. The authentication fingerprint identification method based on the pseudo-random integration according to claim 1, wherein: the classification algorithm in the machine learning algorithm comprises a k-nearest neighbor algorithm, a naive Bayes algorithm, an SVM algorithm and a decision tree algorithm.

5. The authentication fingerprint identification method based on the pseudo-random integration according to claim 2, wherein: in step S101, the method for detecting the start point position includes absolute amplitude value detection and slope detection.

6. The authentication fingerprint identification method based on the pseudo-random integration according to claim 1, wherein: the feature extraction method in step S3 includes amplitude-based feature extraction, amplitude and phase combination-based feature extraction, frequency-based feature extraction, phase shift-based feature extraction, wavelet transform-based feature extraction, and multi-resolution analysis-based feature extraction.

7. The authentication fingerprint identification method based on the pseudo-random integration according to claim 6, wherein: the selection of the mother wavelet in the multiresolution analysis includes haar, dB2, bior, and morl.

Images