CN109633289A - A kind of red information detecting method of electromagnetism based on cepstrum and convolutional neural networks - Google Patents
A kind of red information detecting method of electromagnetism based on cepstrum and convolutional neural networks Download PDFInfo
- Publication number
- CN109633289A CN109633289A CN201811625057.5A CN201811625057A CN109633289A CN 109633289 A CN109633289 A CN 109633289A CN 201811625057 A CN201811625057 A CN 201811625057A CN 109633289 A CN109633289 A CN 109633289A
- Authority
- CN
- China
- Prior art keywords
- electromagnetism
- red information
- sample
- red
- signal
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R29/00—Arrangements for measuring or indicating electric quantities not covered by groups G01R19/00 - G01R27/00
- G01R29/08—Measuring electromagnetic field characteristics
Landscapes
- Physics & Mathematics (AREA)
- Electromagnetism (AREA)
- General Physics & Mathematics (AREA)
- Investigating Or Analyzing Materials By The Use Of Magnetic Means (AREA)
Abstract
The present invention provides a kind of red information detecting method of the electromagnetism based on cepstrum and convolutional neural networks, the sample of multiple electromagnetic leakage signals is acquired first and is done down-sampled, the leakage of information feature in electromagnetic leakage signal is extracted by cepstral analysis, forming the red information characteristics of electromagnetism based on cepstrum indicates, then a large amount of classification based trainings are carried out to the red information characteristics extracted by convolutional neural networks, obtain the detection model about the red information of electromagnetism, then electromagnetic leakage signal sample to be measured is inputted, the red information characteristics of down-sampled and cepstrum electromagnetism are equally done to extract, identification decision finally is carried out to red information characteristics using trained electromagnetism red infomation detection model, pass through the included priori label of Determination result and electromagnetic leakage signal sample to be measured, assess the Detection accuracy of the red information of electromagnetism.The red information detecting method of electromagnetism provided by the invention can detect electromagnetic leakage signal, high sensitivity and Detection accuracy higher than conventional method in the environment of low signal-to-noise ratio.
Description
Technical field
The present invention relates to field of information security technology, and in particular to a kind of to be let out using cepstrum and convolutional neural networks to electromagnetism
Leak the method that red information is detected.
Background technique
With the deep development and extensive use of information technology and computer technology, computer is led in military affairs, government affairs, business
The effect in domain is also more and more important.A kind of processing equipment of the computer as data information, can be inevitable when handling information
Ground external environment distributes electromagnetic energy.The electromagnetic energy leakage of equipment includes radiating and conducting two kinds of approach: radiation leakage is
Various holes, the connecting cable etc. spuious electromagnetic energy penetrated in the form of an electromagnetic wave on device housings and shell gives off
It goes, and leakage is then to conduct spuious electromagnetic energy by various routes (including power supply line and signal wire etc.).
Above two leak path is interrelated, and under the antenna effect effect of conducting wire, the stray electromagnetic energy along conducting wire conduction has portion
It point is converted into the stray electromagnetic energy that electromagenetic wave radiation goes out, and is radiated space to be partly coupled on outer company's conducting wire, i.e.,
Make that there are exchange energy phenomenons.
Contain complicated spectral components in these electromagnetic radiation energies (i.e. electromagnetic wave signal) and carries bulk information, and
These information being compromised usually are divided into two kinds of " red information " and " poison-pen letter breath " again, wherein " poison-pen letter breath " and equipment are handled or passed
Defeated information is unrelated, and " red information " then with equipment processing or transmit it is information-related, when " red information " is in certain precondition
It is received and restores by particular instrument down, will constitute a serious threat to information security.Currently, for Computer and telephone intergration
Research mainly for " red information ", this not only contributes to the protecting information safety for reinforcing country, and business secret and
Protection in terms of people's privacy also important in inhibiting.
When the red information of electromagnetism is intercepted and captured and is restored, it is desirable that the intensity and signal-to-noise ratio of leakage signal must reach certain
Condition, therefore in order to avoid the red leakage of information of electromagnetism, we can take appropriate measures to weaken the intensity of leakage signal, reduce
The signal-to-noise ratio of leakage signal, and then achieve the purpose that electromagnetic protection.And in order to examine electromagnetic protection method whether effective, Wo Menxu
The red information of the electromagnetism of leakage is detected, therefore, the red information characteristics of electromagnetism extract and identify safe to electromagnetic information is ensured
It is significant.
Since the red information of electromagnetism has the characteristics that signal-to-noise ratio is low, characteristic signal intensity is faint, in complicated electromagnetic environment,
The red information characteristics extracting method of traditional electromagnetism is due to needing to pre-define red information characteristics, and these red information characteristics hold very much
It is easily submerged in complicated electromagnetic environment, it is difficult to separation and Extraction, therefore conventional method can not be extracted and be represented well
The red information characteristics of electromagnetism influence subsequent detection discrimination.Detection discrimination can take target device before and after safeguard procedures
Quantitative evaluation is made in safety, is the quantitative basis of Information Security Evaluation.
Summary of the invention
The purpose of the present invention is to provide it is a kind of can be in the electromagnetic leakage under complex electromagnetic environment, from low signal-to-noise ratio effectively
It extracts red information characteristics and detects knowledge method for distinguishing.
To achieve the above object, the present invention provides a kind of red infomation detections of the electromagnetism based on cepstrum and convolutional neural networks
Method acquires the sample of electromagnetic leakage signal first, and extracts the red letter hidden in electromagnetic leakage signal by cepstral analysis
Feature is ceased, classification based training then is carried out to the red information characteristics extracted by convolutional neural networks, is obtained about the red letter of electromagnetism
The detection model of breath then inputs electromagnetic leakage signal to be measured, and forms the red information characteristics of electromagnetism of the signal based on cepstrum
It indicates, identification decision finally is carried out to red information characteristics using trained electromagnetism red infomation detection model.
Above-mentioned detection method includes obtaining two parts of detection model and the red infomation detection of electromagnetism, and each section specifically includes
Following steps:
Obtain detection model
Step S10) the multiple electromagnetic leakage signal samples of acquisition;
Step S11) it is to solve the problems, such as that different acquisition environment down-sampling precision is inconsistent, down-sampled place is done to all signals
Reason;
Step S12) cepstral analysis is made to each signal after down-sampled and extracts leakage of information feature therein, form base
It is indicated in the red information characteristics of the electromagnetism of cepstrum;
Step S13) the red information characteristics of above-mentioned electromagnetism are indicated to carry out classification based training in input convolutional neural networks;
Step S14) by a large amount of sample training, obtain and save the detection model of the red information of electromagnetism;
The red infomation detection of electromagnetism
Step S20) by electromagnetic leakage signal sample to be measured input detection model;
Step S21) it is down-sampled to electromagnetic leakage signal sample to be measured progress according to acquisition environment down-sampling precision;
Step S22) cepstral analysis is made to the measured signal sample after down-sampled and extracts leakage of information feature therein, shape
It is indicated at the red information characteristics of electromagnetism based on cepstrum;
Step S23) utilize the red infomation detection model of the trained electromagnetism based on convolutional neural networks to red information characteristics
Carry out identification decision.
Preferably, step S11) and step S21) in down-sampled process it is as follows:
(1) the length L of training/test sample is setmAnd the smallest precision S needed for training/test samplem, and acquire
Original signal sample length be LsAnd the precision of acquisition is Ss;
(2) if Ss≥SmAnd Ls≥Lm, by the original signal sample of acquisition with Ls/LmThe interval of (rounding) carries out at equal intervals
It is down-sampled, go out L from original signal samplings/Lm(rounding) a standardization sample, by curtailment LmSample abandon;
(3) if Ss< SmAnd Ls≥Lm, by the original signal sample of acquisition with LmLength be segmented, by original signal
Sample decomposition is Ls/Lm(rounding) a standardization sample, by curtailment LmSample abandon;
(4) if Ls< Lm, then the curtailment of sample, re-starts acquisition.
Will be carried out the standardization of sample of signal by above-mentioned down-sampled handle, in order to subsequent progresss deep learning with
Batch training.Wherein, the length unit of sample of signal is the number of sampled point, and the precision unit of sample of signal is sample number/second.
Preferably, step S12) and step S22) in about the red information cepstrum feature of electromagnetism extracting method it is as follows:
Fourier transformation is carried out to the electromagnetic leakage signal after down-sampled:
S (k)=FFT (S (n))
Wherein, S (k) is the frequency spectrum of electromagnetic leakage signal, and FFT is Fast Fourier Transform (FFT), S (n) be it is down-sampled after electricity
The time-domain signal of magnetic dispersion signal;
Logarithm operation is done to the amplitude of the frequency spectrum of electromagnetic leakage signal:
L (k)=log | S (k) |
Wherein, L (k) is the frequency domain amplitude logarithmic spectrum of electromagnetic leakage signal, | S (k) | it is the width of electromagnetic leakage signal frequency spectrum
Value and by S (k) carry out modulo operation obtain;
Inverse Fourier transform is carried out to the frequency domain amplitude logarithmic spectrum of electromagnetic leakage signal, the cepstrum for extracting the red information of electromagnetism is special
Sign:
C (n)=IFFT (L (k))
Wherein, C (n) is that the cepstrum feature of electromagnetic leakage signal indicates, IFFT is the inverse operation of Fast Fourier Transform (FFT).
Preferably, the convolutional neural networks include two convolutional layers, a pond layer and a full articulamentum, convolutional layer
Calculation formula are as follows:
Output(Ho×Lo×T0)=W (Hc×Lc×Ti×T0)*Input(Hi×Li×Ti)+b
Wherein, Output is the output tensor of convolutional calculation, and Input is the input tensor of convolutional calculation, and W is convolutional Neural
Network parameter simultaneously carries out convolution algorithm with Input, and b is Bayes's parameter vector and participates in machine learning, and Ho is output tensor square
The line number of battle array, LoTo export tensor matrix column number, ToFor the port number of output, HiFor the line number for inputting tensor matrix, LiIt is defeated
Enter tensor matrix column number, TiFor the port number of input.
Preferably, Relu excitation function is used in the learning process of convolutional neural networks, calculation formula is as follows:
Relu (x)=max (0, x)
Preferably, Dropout function is added in convolutional layer, calculation formula is as follows:
R~Bernoulli (p)
Wherein, r is independent Bernoulli random variable, with Probability p by r value be 1 (due to r value can for 0 or 1, and
As r=1, which participates in network query function, and as r=0, which is not involved in network query function, therefore p can be considered
Node discard probability), x is the calculating input of Dropout function,It is the calculating output of Dropout function.
In the case where limited sample size, the detection model that training obtains may be excessively quasi- with the distribution of training sample
It closes, is effectively identified so as to cause model can not be detected there are the sample to be tested of larger difference with training sample, gained model lacks
Weary generalization ability.It is desirable that machine can not only effectively be identified the sample similar with training sample, can also be pushed away by deep learning
It is wide to identify unknown sample to effective.And the introducing of above-mentioned two function is exactly excessive quasi- during deep learning in order to prevent
Conjunction problem, wherein Relu excitation function in multilayer neural network indicate upper layer node output and lower level node it is defeated
Functional relation between entering, and Dropout function is that the neuron in network is allowed to stop working with certain probability.
Preferably, Probability p=0.5 in Dropout function.That is the probability that r is randomly set to 0 or 1 is 50%,
Purpose is to abandon neuron at random to realize, it is not allowed to participate in network query function.
Preferably, after completing to the identification decision of the red information of electromagnetism, if the electromagnetic leakage signal to be measured band of input itself
There is priori label, then can assess the Detection accuracy of the red information of electromagnetism by comparing label and testing result.
Technical solution provided by the invention at least has the following beneficial effects:
1, the present invention by utilizing cepstral analysis and convolutional neural networks knot in electromagnetic information leakage detection field for the first time
The establishment process of the method for conjunction, extraction process and detection model to the red leakage of information feature of electromagnetism optimizes processing, real
The red information of electromagnetic leakage can be effectively detected in the environment of present low signal-to-noise ratio;Specifically, cepstral analysis is weighted using logarithm, is expanded
The big dynamic range of frequency spectrum, improves the precision converted again, to enhance the cyclophysis in signal, and cepstral analysis
With deconvolution effect, convenient for the red information of electromagnetism is separated and extracted from complex electromagnetic environment, in addition, convolutional neural networks can be with
Using the red information characteristics hidden in the self study analysis identification electromagnetic signal of machine, avoid the artificially defined and red information of searching special
Sign, therefore, the method for the present invention volume high sensitivity and Detection accuracy are higher than conventional method.
2, the present invention carries out machine learning to finite sample using artificial intelligence deep learning method, in the red information sample of electromagnetism
In the cumulative process of this big data, pass through the continuous study to increment sample, sustainable improvement intelligent measurement ability.
It is 3, of the invention by using Relu excitation function and Dropout function in the learning process of convolutional neural networks,
Help to inhibit deep learning existing overfitting problem in the process,
Detailed description of the invention
In order to illustrate more clearly of technical solution of the present invention, below the embodiment of the present invention will be described in required use
Attached drawing is briefly described, it should be apparent that, following drawings be used only for helping understanding section Example in the present invention rather than
The whole of technical solution, in which:
Fig. 1 is the flow chart of the red information detecting method of electromagnetism provided by the present invention;
Fig. 2 is the structure chart of convolutional neural networks in the red information detecting method of electromagnetism provided by the present invention;
Fig. 3 is the sample of signal of electromagnetic leakage in the embodiment of the present invention 1 in down-sampled preceding time domain samples;
Fig. 4 is that the sample of signal of electromagnetic leakage in the embodiment of the present invention 1 is down-sampled, cepstral analysis extraction leakage of information is special
Cepstrum sample after sign.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution of the present invention is clearly and completely described,
Obviously, described embodiment is only section Example of the invention, instead of all the embodiments.Based on the reality in the present invention
Example is applied, all other embodiment obtained by those of ordinary skill in the art without making creative efforts all belongs to
In the scope of protection of the invention.
The present invention provides a kind of red information detecting methods of the electromagnetism based on cepstrum and convolutional neural networks, first acquisition electricity
The sample of magnetic dispersion signal, and the red information characteristics hidden in electromagnetic leakage signal are extracted by cepstral analysis, then pass through
Convolutional neural networks carry out classification based training to the red information characteristics extracted, obtain connecing about the detection model of the red information of electromagnetism
Input electromagnetic leakage signal to be measured, and form the red information characteristics of electromagnetism of the signal based on cepstrum and indicate, finally utilize instruction
The red infomation detection model of the electromagnetism perfected carries out identification decision to red information characteristics.
As shown in Figure 1, specific step is as follows for the detection method:
One, detection model is obtained
Step S10) the multiple electromagnetic leakage signal samples of acquisition;
Step S11) it is to solve the problems, such as that different acquisition environment down-sampling precision is inconsistent, down-sampled place is done to all signals
Reason;
Step S12) cepstral analysis is made to each signal after down-sampled and extracts leakage of information feature therein, form base
It is indicated in the red information characteristics of the electromagnetism of cepstrum;
Step S13) the red information characteristics of above-mentioned electromagnetism are indicated to carry out classification based training in input convolutional neural networks;
Step S14) by a large amount of sample training, obtain and save the detection model of the red information of electromagnetism;
Two, the red infomation detection of electromagnetism
Step S20) by electromagnetic leakage signal sample to be measured input detection model;
Step S21) it is down-sampled to electromagnetic leakage signal sample to be measured progress according to acquisition environment down-sampling precision;
Step S22) cepstral analysis is made to the measured signal sample after down-sampled and extracts leakage of information feature therein, shape
It is indicated at the red information characteristics of electromagnetism based on cepstrum;
Step S23) utilize the red infomation detection model of the trained electromagnetism based on convolutional neural networks to red information characteristics
Carry out identification decision;
Step S24) input electromagnetic leakage signal to be measured per se with priori label, pass through and compare label and testing result
To assess the Detection accuracy of the red information of electromagnetism.
In detection method provided by the invention, step S11) and step S21) in down-sampled process it is as follows:
(1) the length L of training/test sample is setmAnd the smallest precision S needed for training/test samplem, and acquire
Original signal sample length be LsAnd the precision of acquisition is Ss;
(2) if Ss≥SmAnd Ls≥Lm, by the original signal sample of acquisition with Ls/LmThe interval of (rounding) carries out at equal intervals
It is down-sampled, go out L from original signal samplings/Lm(rounding) a standardization sample, by curtailment LmSample abandon;
(3) if Ss< SmAnd Ls≥Lm, by the original signal sample of acquisition with LmLength be segmented, by original signal
Sample decomposition is Ls/Lm(rounding) a standardization sample, by curtailment LmSample abandon;
(4) if Ls< Lm, then the curtailment of sample, re-starts acquisition.
In detection method provided by the invention, step S12) and step S22) in about the red information cepstrum feature of electromagnetism
Extracting method is as follows:
Fourier transformation is carried out to the electromagnetic leakage signal after down-sampled:
S (k)=FFT (S (n))
Wherein, S (k) is the frequency spectrum of electromagnetic leakage signal, and FFT is Fast Fourier Transform (FFT), S (n) be it is down-sampled after electricity
The time-domain signal of magnetic dispersion signal;
Logarithm operation is done to the amplitude of the frequency spectrum of electromagnetic leakage signal:
L (k)=log | S (k) |
Wherein, L (k) is the frequency domain amplitude logarithmic spectrum of electromagnetic leakage signal, | S (k) | it is the width of electromagnetic leakage signal frequency spectrum
Value and by S (k) carry out modulo operation obtain;
Inverse Fourier transform is carried out to the frequency domain amplitude logarithmic spectrum of electromagnetic leakage signal, the cepstrum for extracting the red information of electromagnetism is special
Sign:
C (n)=IFFT (L (k))
Wherein, C (n) is that the cepstrum feature of electromagnetic leakage signal indicates, IFFT is the inverse operation of Fast Fourier Transform (FFT).
As shown in Fig. 2, the convolutional neural networks include two convolutional layers, one in detection method provided by the invention
A pond layer and a full articulamentum, parameter setting are as shown in table 1:
1 convolutional neural networks structure of table
Convolutional neural networks layering | Calculate core | Material calculation |
Convolutional layer 1 | 20×8 | 1×1 |
Pond layer | 2×2 | 2×2 |
Convolutional layer 2 | 10×4 | 1×1 |
Full articulamentum | 1×1 | 1×1 |
The calculation formula of convolutional layer are as follows:
Output(Ho×Lo×T0)=W (Hc×Lc×Ti×T0)*Input(Hi×Li×Ti)+b
Wherein, Output is the output tensor of convolutional calculation, and Input is the input tensor of convolutional calculation, and W is convolutional Neural
Network parameter simultaneously carries out convolution algorithm with Input, and b is Bayes's parameter vector and participates in machine learning, and Ho is output tensor square
The line number of battle array, LoTo export tensor matrix column number, ToFor the port number of output, HiFor the line number for inputting tensor matrix, LiIt is defeated
Enter tensor matrix column number, TiFor the port number of input.
In detection method provided by the invention, the overfitting problem in deep learning in order to prevent, in convolutional Neural
Relu excitation function is used in the learning process of network, calculation formula is as follows:
Relu (x)=max (0, x)
Meanwhile Dropout function being added also in convolutional layer, calculation formula is as follows:
R~Bernoulli (p)
Wherein, r is independent Bernoulli random variable, with Probability p by r value be 1, x be Dropout function calculating it is defeated
Enter,It is the calculating output of Dropout function.
Embodiment 1 (uses the red information detecting method of the above-mentioned electromagnetism based on cepstrum and convolutional neural networks)
1, m electromagnetic leakage signal sample is collected using signal receiver:
Si(t), i=1,2,3....m
Its time domain samples is as shown in Figure 3.
2, down-sampled processing is done to electromagnetic leakage signal sample, L is setmFor 16000 sampled points, SmFor 2MS/s, obtain
Standardized electromagnetic leakage signal time series:
Si(n), i=1,2,3....m
Wherein, the length of each sample sequence is 16000, i.e. 0 < n < 16000.
3, cepstral analysis is made to the electromagnetic leakage signal sample after down-sampled, it is special to form the red information of electromagnetism based on cepstrum
Sign indicates that process is as follows:
(1) to Si(n) Fourier transformation is carried out, the frequency spectrum of electromagnetic leakage signal is obtained:
Si(k)=FFT (Si(n))
Wherein, there is 0 < k < 16000;
(2) to Si(k) logarithm operation is carried out, the frequency domain amplitude logarithmic spectrum of electromagnetic leakage signal is obtained:
Li(k)=log | Si (k) |
(3) to Li(k) inverse Fourier transform is carried out, the cepstrum of electromagnetic leakage signal is obtained:
Ci(n)=IFFT (Li(k))
Its cepstrum sample as shown in figure 4, and due to spectral line bilateral symmetry, so only need it is half of compose, the length of sample
Degree can be reduced to 8000.
4, the sample (as follows) after feature extraction is divided into two sample sets, i.e. training set and test set, wherein
Ci(n), i=1,2,3....m
Training set sample is inputted in convolutional neural networks and is trained, the detection model of the red information of electromagnetic leakage is obtained,
Process is as follows:
(1) training sample sequence is converted into two-dimentional input
It is more preferable to extract signal context feature in order to adapt to the neural network of convolutional layer, by Ci(n) 8000 point sequences drop
Dimension is mapped to two dimension input Input1i, Input1 is the tensor of one (98*40*1), indicates the size of the two-dimensional matrix of input
For 98 × 40 and input channel number is 1;
(2) calculating of convolutional layer 1
Convolutional layer 1 is slided using convolution kernel and carries out convolution algorithm, is carried out feature extraction to the signal input in 1 channel, is obtained
Obtain the feature output in 64 channels, calculation formula are as follows:
Output1=W1*Input1+b1
According to the convolutional neural networks structure in table 1, Input1 is the input tensor of (98*40*1), and Output1 is (98*
Output tensor 40*64), W1For the parameter tensor of { (20*8) * 1*64 } convolutional neural networks, convolution kernel size is (20*8),
The material calculation of convolution kernel is (1*1), and input channel number is 1, and output channel number is 64, b1For Bayes's parameter and size is
64;
(3) calculating of Relu excitation function
Calculation formula are as follows:
Relu (Output1)=max (0, Output1)
Wherein, the input tensor size of Relu excitation function is (98*40*64), and the output tensor of Relu excitation function is big
Small is (98*40*64);
(4) calculating of Dropout function
Calculation formula are as follows:
R~Bernoulli (p)
Wherein, node discard probability p=0.5 is set, is at random abandoned some node zero setting, to prevent training result excessively quasi-
It closes, the input tensor size of Dropout function is (98*40*64), and the output tensor size of Dropout function is (98*40*
64);
(5) calculating of pond layer
Pond layer carries out down-sampling calculating, and according to the design of neural network structure in table one, the ratio of down-sampling is 1/ (2*
2), the step-length of sampling is (2*2), and the input tensor size of pond layer is (98*40*64), and the output tensor size of pond layer is
(49*20*64);
(6) calculating of convolutional layer 2
The calculation formula of convolutional layer 2 is
Output2=W2*Input2+b2
Equally, according to the convolutional neural networks structure in table 1, Input2 is the input tensor of (49*20*64), Output2
It is the output tensor of (49*20*64), W2For the parameter tensor of { (10*4) * 64*64 } convolutional neural networks, convolution kernel size is
(10*4), the material calculation of convolution kernel are (1*1), and input channel number is 64, and output channel number is 64, b2For Bayes's parameter and
Size is 64;
(7) calculating of full articulamentum
(49*20*64) tensor that convolutional layer 2 exports is mapped to the sequence that length is 62720 by matrix operation first,
Then fully-connected network output category result is utilized, can independently set the classification quantity of output as needed in the present embodiment;
(8) feedback regulation of convolutional neural networks parameter
Using gradient descent method, according to the calibration of the sample of training set to the parameter W of convolutional neural networks1、W2、b1、b2Into certainly
Action feedback and adjusting, so that prediction result approaches sample calibration;
(9) repetitive exercise
Repeat the above steps (1)~(8), and the repetitive exercise of convolutional neural networks is carried out according to the frequency of training preset,
Until training detection model.
5, test set sample is inputted in detection model, sample to be tested is calculated according to above-mentioned steps (1)~(7),
To complete the recognition detection to the red information of electromagnetism.
Under the premise of using the same convolutional neural networks structure, by the method for the present invention and traditional time domain, frequency domain,
MFCC (mel cepstrum coefficients), db2 wavelet filteration method are compared.Related data is as follows:
Experimental design: the electromagnetic leakage signal generated when showing different content to display detects.
Using data: training set includes 74052 samples, sample rate 1.25GS/s, 208MS/s and 2MS/s;Test set
Including 5208 samples, sample rate 417MS/s.
The results show, the method for the present invention is by combining cepstral analysis and convolutional neural networks to ensure that the red information of electromagnetism is examined
Survey accuracy rate be excellent in up to 95% or more, and in terms of accuracy rate, accurate rate, recall rate, compared to conventional method
Or the combination of other methods, effect of optimization is obvious and has higher comprehensive evaluation index.
The above description is only a preferred embodiment of the present invention, is not intended to limit scope of patent protection of the invention, for
For those skilled in the art, the invention may be variously modified and varied.Within the spirit and principles in the present invention, all
Using any improvement or equivalent replacement made by description of the invention and accompanying drawing content, directly or indirectly it is used in other relevant
Technical field should all be included within the scope of the present invention.
Claims (8)
1. a kind of red information detecting method of electromagnetism based on cepstrum and convolutional neural networks, which is characterized in that acquisition electromagnetism first
The sample of leakage signal, and the red information characteristics hidden in electromagnetic leakage signal are extracted by cepstral analysis, then pass through volume
Product neural network carries out classification based training to the red information characteristics extracted, obtains the detection model about the red information of electromagnetism, then
Electromagnetic leakage signal to be measured is inputted, and form the red information characteristics of electromagnetism of the signal based on cepstrum to indicate, finally utilizes training
The good red infomation detection model of electromagnetism carries out identification decision to red information characteristics.
2. the red information detecting method of electromagnetism according to claim 1, which is characterized in that specifically comprise the following steps:
Obtain detection model
Step S10) the multiple electromagnetic leakage signal samples of acquisition;
Step S11) it is to solve the problems, such as that different acquisition environment down-sampling precision is inconsistent, down-sampled processing is done to all signals;
Step S12) cepstral analysis is made to each signal after down-sampled and extracts leakage of information feature therein, it is formed based on falling
The red information characteristics of the electromagnetism of spectrum indicate;
Step S13) the red information characteristics of above-mentioned electromagnetism are indicated to carry out classification based training in input convolutional neural networks;
Step S14) by a large amount of sample training, obtain and save the detection model of the red information of electromagnetism;
The red infomation detection of electromagnetism
Step S20) by electromagnetic leakage signal sample to be measured input detection model;
Step S21) it is down-sampled to electromagnetic leakage signal sample to be measured progress according to acquisition environment down-sampling precision;
Step S22) cepstral analysis is made to the measured signal sample after down-sampled and extracts leakage of information feature therein, form base
It is indicated in the red information characteristics of the electromagnetism of cepstrum;
Step S23) red information characteristics are carried out using the red infomation detection model of the trained electromagnetism based on convolutional neural networks
Identification decision.
3. the red information detecting method of electromagnetism according to claim 2, which is characterized in that step S11) and step S21) in
Down-sampled process is as follows:
(1) the length L of training/test sample is setmAnd the smallest precision S needed for training/test samplem, and the original acquired
Beginning sample of signal length is LsAnd the precision of acquisition is Ss;
(2) if Ss≥SmAnd Ls≥Lm, by the original signal sample of acquisition with Ls/LmThe interval of (rounding) carries out drop at equal intervals and adopts
Sample goes out L from original signal samplings/Lm(rounding) a standardization sample, by curtailment LmSample abandon;
(3) if Ss< SmAnd Ls≥Lm, by the original signal sample of acquisition with LmLength be segmented, by original signal sample
It is divided into Ls/Lm(rounding) a standardization sample, by curtailment LmSample abandon;
(4) if Ls< Lm, then the curtailment of sample, re-starts acquisition.
4. the red information detecting method of electromagnetism according to claim 3, which is characterized in that step S12) and step S22) in close
It is as follows in the extracting method of the red information cepstrum feature of electromagnetism:
Fourier transformation is carried out to the electromagnetic leakage signal after down-sampled:
S (k)=FFT (S (n))
Wherein, S (k) is the frequency spectrum of electromagnetic leakage signal, and FFT is Fast Fourier Transform (FFT), S (n) be it is down-sampled after electromagnetism let out
The time-domain signal of leakage signal;
Logarithm operation is done to the amplitude of the frequency spectrum of electromagnetic leakage signal:
L (k)=log | S (k) |
Wherein, L (k) is the frequency domain amplitude logarithmic spectrum of electromagnetic leakage signal, | S (k) | be the amplitude of electromagnetic leakage signal frequency spectrum and
It is obtained by carrying out modulo operation to S (k);
Inverse Fourier transform is carried out to the frequency domain amplitude logarithmic spectrum of electromagnetic leakage signal, extracts the cepstrum feature of the red information of electromagnetism:
C (n)=IFFT (L (k))
Wherein, C (n) is that the cepstrum feature of electromagnetic leakage signal indicates, IFFT is the inverse operation of Fast Fourier Transform (FFT).
5. the red information detecting method of electromagnetism according to claim 4, which is characterized in that the convolutional neural networks include two
A convolutional layer, a pond layer and a full articulamentum, the calculation formula of convolutional layer are as follows:
Output(Ho×Lo×T0)=W (Hc×Lc×Ti×T0)*Input(Hi×Li×Ti)+b
Wherein, Output is the output tensor of convolutional calculation, and Input is the input tensor of convolutional calculation, and W is convolutional neural networks
Parameter simultaneously carries out convolution algorithm with Input, and b is Bayes's parameter vector and participates in machine learning, HoFor output tensor matrix
Line number, LoTo export tensor matrix column number, ToFor the port number of output, HiFor the line number for inputting tensor matrix, LiFor input
The columns of moment matrix, TiFor the port number of input.
6. the red information detecting method of electromagnetism according to claim 5, which is characterized in that in the study of convolutional neural networks
Relu excitation function is used in journey, calculation formula is as follows:
Relu (x)=max (0, x).
7. the red information detecting method of electromagnetism according to claim 6, which is characterized in that Dropout is added in convolutional layer
Function, calculation formula are as follows:
R~Bernoulli (p)
Wherein, r is independent Bernoulli random variable, is inputted r value for the calculating that 1, x is Dropout function with Probability p,
It is the calculating output of Dropout function.
8. the red information detecting method of electromagnetism described according to claim 1~any one of 7, which is characterized in that in completion pair
After the identification decision of the red information of electromagnetism, if the electromagnetic leakage signal to be measured of input can pass through comparison per se with priori label
Label and testing result assess the Detection accuracy of the red information of electromagnetism.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811625057.5A CN109633289A (en) | 2018-12-28 | 2018-12-28 | A kind of red information detecting method of electromagnetism based on cepstrum and convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811625057.5A CN109633289A (en) | 2018-12-28 | 2018-12-28 | A kind of red information detecting method of electromagnetism based on cepstrum and convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109633289A true CN109633289A (en) | 2019-04-16 |
Family
ID=66078988
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811625057.5A Pending CN109633289A (en) | 2018-12-28 | 2018-12-28 | A kind of red information detecting method of electromagnetism based on cepstrum and convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109633289A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108680796A (en) * | 2018-05-17 | 2018-10-19 | 集美大学 | Electromagnetic information leakage detecting system and method for computer display |
CN110163099A (en) * | 2019-04-17 | 2019-08-23 | 中国电子科技网络信息安全有限公司 | A kind of abnormal behaviour identification device and method based on electromagnetic leakage signal |
CN110490154A (en) * | 2019-08-23 | 2019-11-22 | 集美大学 | A kind of multidimensional leakage information detection method, terminal device and storage medium |
CN110888623A (en) * | 2019-11-25 | 2020-03-17 | 集美大学 | Data conversion method, multiplier, adder, terminal device and storage medium |
TWI717221B (en) * | 2019-08-09 | 2021-01-21 | 開曼群島商創新先進技術有限公司 | Model parameter determination method, device and electronic equipment |
KR102225817B1 (en) * | 2019-11-14 | 2021-03-11 | 고려대학교 산학협력단 | Apparatus and method of testing semiconductor device by using frequency output feature of semiconductor device and machine learning |
CN113075462A (en) * | 2021-02-22 | 2021-07-06 | 中国人民解放军国防科技大学 | Electromagnetic field distribution positioning method based on neural network |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080019538A1 (en) * | 2006-07-24 | 2008-01-24 | Motorola, Inc. | Method and apparatus for removing periodic noise pulses in an audio signal |
CN106952649A (en) * | 2017-05-14 | 2017-07-14 | 北京工业大学 | Method for distinguishing speek person based on convolutional neural networks and spectrogram |
CN108663576A (en) * | 2018-05-08 | 2018-10-16 | 集美大学 | Weak electromagnetic red signal detection method under a kind of complex environment |
CN108680796A (en) * | 2018-05-17 | 2018-10-19 | 集美大学 | Electromagnetic information leakage detecting system and method for computer display |
CN108830135A (en) * | 2018-04-25 | 2018-11-16 | 中国电子科技集团公司电子科学研究院 | Computer and telephone intergration signal recognition method, device and storage medium |
CN108986787A (en) * | 2017-05-31 | 2018-12-11 | 英特尔公司 | Use the feature extraction of neural network accelerator |
-
2018
- 2018-12-28 CN CN201811625057.5A patent/CN109633289A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080019538A1 (en) * | 2006-07-24 | 2008-01-24 | Motorola, Inc. | Method and apparatus for removing periodic noise pulses in an audio signal |
CN106952649A (en) * | 2017-05-14 | 2017-07-14 | 北京工业大学 | Method for distinguishing speek person based on convolutional neural networks and spectrogram |
CN108986787A (en) * | 2017-05-31 | 2018-12-11 | 英特尔公司 | Use the feature extraction of neural network accelerator |
CN108830135A (en) * | 2018-04-25 | 2018-11-16 | 中国电子科技集团公司电子科学研究院 | Computer and telephone intergration signal recognition method, device and storage medium |
CN108663576A (en) * | 2018-05-08 | 2018-10-16 | 集美大学 | Weak electromagnetic red signal detection method under a kind of complex environment |
CN108680796A (en) * | 2018-05-17 | 2018-10-19 | 集美大学 | Electromagnetic information leakage detecting system and method for computer display |
Non-Patent Citations (1)
Title |
---|
周源: ""基于USRP平台的新型计算机视频信息电磁泄漏截获系统研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108680796A (en) * | 2018-05-17 | 2018-10-19 | 集美大学 | Electromagnetic information leakage detecting system and method for computer display |
CN108680796B (en) * | 2018-05-17 | 2023-09-19 | 集美大学 | Electromagnetic information leakage detection system and method for computer display |
CN110163099A (en) * | 2019-04-17 | 2019-08-23 | 中国电子科技网络信息安全有限公司 | A kind of abnormal behaviour identification device and method based on electromagnetic leakage signal |
CN110163099B (en) * | 2019-04-17 | 2022-04-01 | 中国电子科技网络信息安全有限公司 | Device and method for identifying abnormal behaviors based on electromagnetic leakage signals |
TWI717221B (en) * | 2019-08-09 | 2021-01-21 | 開曼群島商創新先進技術有限公司 | Model parameter determination method, device and electronic equipment |
CN110490154A (en) * | 2019-08-23 | 2019-11-22 | 集美大学 | A kind of multidimensional leakage information detection method, terminal device and storage medium |
KR102225817B1 (en) * | 2019-11-14 | 2021-03-11 | 고려대학교 산학협력단 | Apparatus and method of testing semiconductor device by using frequency output feature of semiconductor device and machine learning |
CN110888623A (en) * | 2019-11-25 | 2020-03-17 | 集美大学 | Data conversion method, multiplier, adder, terminal device and storage medium |
CN110888623B (en) * | 2019-11-25 | 2021-11-23 | 集美大学 | Data conversion method, multiplier, adder, terminal device and storage medium |
CN113075462A (en) * | 2021-02-22 | 2021-07-06 | 中国人民解放军国防科技大学 | Electromagnetic field distribution positioning method based on neural network |
CN113075462B (en) * | 2021-02-22 | 2022-05-17 | 中国人民解放军国防科技大学 | Electromagnetic field distribution positioning method based on neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109633289A (en) | A kind of red information detecting method of electromagnetism based on cepstrum and convolutional neural networks | |
Bao et al. | Threat of adversarial attacks on DL-based IoT device identification | |
CN112802484B (en) | Panda sound event detection method and system under mixed audio frequency | |
Lin et al. | Detection of gravitational waves using Bayesian neural networks | |
CN108573225A (en) | A kind of local discharge signal mode identification method and system | |
CN105488466B (en) | A kind of deep-neural-network and Acoustic Object vocal print feature extracting method | |
CN114564982B (en) | Automatic identification method for radar signal modulation type | |
Firuzi et al. | A novel method for differentiating and clustering multiple partial discharge sources using S transform and bag of words feature | |
CN112036239B (en) | Radar signal working mode identification method and system based on deep learning network | |
CN108680796A (en) | Electromagnetic information leakage detecting system and method for computer display | |
CN110211604A (en) | A kind of depth residual error network structure for voice deformation detection | |
CN109116312A (en) | Radar signal classification method based on QMFB and convolutional neural networks | |
CN110427878A (en) | A kind of sudden and violent signal recognition method of Rapid Radio and system | |
CN107064177B (en) | Ultra wide band soil signal water content recognition methods based on adaptive nuero-fuzzy inference system | |
CN105426832A (en) | Communication radar radiation source identification method in presence of unsteady SNR (Signal Noise Ratio) | |
CN115240651A (en) | Land-air communication speaker role identification method and device based on feature fusion | |
CN110458071A (en) | A kind of fiber-optic vibration signal characteristic abstraction and classification method based on DWT-DFPA-GBDT | |
Ma et al. | Artificial intelligence model for gravitational wave search based on the waveform envelope | |
Xu et al. | Magnetic anomaly detection using multifeature fusion-based neural network | |
CN113935367A (en) | Terahertz time-domain spectrum hidden dangerous goods classification method based on ResNet and LSTM fusion | |
Orlic et al. | Earthquake—explosion discrimination using genetic algorithm-based boosting approach | |
CN116466408B (en) | Artificial neural network superbedrock identification method based on aeromagnetic data | |
Harid et al. | Automated large‐scale extraction of whistlers using mask‐scoring regional convolutional neural network | |
CN112581940A (en) | Discharging sound detection method based on edge calculation and neural network | |
CN111983569A (en) | Radar interference suppression method based on neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |