Based on the audio authentication method to ENF phase spectrums and instantaneous frequency spectrum analysis
Technical field
The invention belongs to digital audio and video signals processing technology fields, more particularly to one kind is based on to ENF phase spectrums and instantaneously
The audio authentication method of frequency spectrum analysis.
Background technology
Currently, the prior art commonly used in the trade is such:
With the development of computer and internet the relevant technologies, people are more dependent on using digital multimedia data.Number
The advantages of word multi-medium data is easy to preservation, editor and propagates brings many facilities and enjoyment to people's daily life.Such as people
Do not need any professional knowledge quickly and easily digital audio file can be spliced using audio edited software, plus
Enter the operations such as noise and transformation, this is Internet era entertainment way prevailing.But the development of technology is a double-edged sword, together
When allow some criminals to have an opportunity to take advantage of.Criminal can maliciously distort digital audio and be carried out wide-scale distribution,
And it only is difficult to discover out with sense organ.It is propagated in court's recording proof, Deceptive news if applying such digital audio file
Etc. under occasions, may result in serious consequence, damage the just and social trust order of law.Thus it is guaranteed that digital audio
Authenticity and integrity, it is highly important to carry out tampering detection to digital audio.Digital audio tampering detection is digital audio
One important branch of evidence obtaining, in the fields such as judicial evidence collection, Justice of Journalism and scientific discovery extensive application.
In current digital audio altering detecting method, most efficient method is the detection based on mains frequency consistency
Method in the past decade almost becomes the common standard of digital audio identification, obtain in world wide academic research personnel and
The concern of law enforcement agency.Its principle is, if sound pick-up outfit downward recording audio the case where accessing power grid, audio letter
Mains frequency (Electirc Network Frequency, ENF) information will necessarily be carried in number.This not only enables ENF become one
Kind is naturally embedded into the watermark signal in audio signal, and can be used as timestamp.Embedded ENF in audio file
Ingredient (ENF component, ENFC), can extract by bandpass filtering.Using ENFC stability and uniqueness into
Generally there are two Research Thinkings for row digital audio tampering detection, and first is the mains frequency for the ENFC and power supply department that will be extracted
Data in database are compared, and determine whether the audio recording time is consistent with what is declared, establishes and preserves and is large-scale
ENF Signals Data Bases difficulty is high, cost is high, and there is presently no the relatively high ENF databases of practical value.Grigoras most early in
Romanian part establishes ENF reference databases.Liu Yuming etc. analyzes north American power grid detecting system, proposes to establish mark
The method of quasi- mains frequency;Second is the certain features extracted in ENF signals, carries out consistency or Regularity Analysis.
Grigoras proposes the audio forgery detection algorithm based on ENF earliest, mainly passes through the fluctuation and reference of ENF in audio to be detected
The data in time are compared, to judge whether audio is tampered with.Then Grigoras verifications add in short-term audio signal
Window is analyzed, and can carry out more careful, accurate comparison with database.The Research foundation in Grigoras such as Rodr í guez
On, the consistency of ENF phase changes is detected audio by the method for proposing that ENF standard databases need not be used as feature
It distorts, chooses boundary value and categorised decision is carried out to this feature.Hu Yongjian etc. is ideal by one on the basis of Rodr í guez
Sinusoidal signal, which is used as, refers to signal, constructs new characteristic quantity to detect the discontinuity of ENF phases.Hu Yongjian etc. is then to above-mentioned
Method is improved, and proposes not needing additional reference signal and the method that directly calculates ENF maximum offsets, furthermore with mostly special
Sign, which is combined, is accurately positioned tampered region.Esquef etc. can cause to distort the mutation of point ENF instantaneous frequencys according to operation is distorted, and propose
TPSW (Two-Pass Split-Window) method estimates ENF background change levels, is more than background by practical Instantaneous frequency variations
The peak point of change level is known as distorting a little.
In conclusion problem of the existing technology is:
There are problems that currently based on the ENF researchs for carrying out the passive tampering detection of digital audio:
1) do not have authoritative ENF comparison databases.Using being carried out in ENF ingredients and the ENF databases in measured signal pair
Than distorting no reliable result to judge whether voice signal passes through;
2) most of method does not extract characteristic crucial in voice signal, can directly be to voice signal
It is no to be tampered carry out decision;
3) correlation between override feature collection, it is not further to the initial characteristic data extracted to be handled;
4) existing most methods the degree of automation is not high, ineffective, and to the adaptivity of disparate databases signal
Difference.
Solve the difficulty and meaning of above-mentioned technical problem:
The ENF comparison databases for establishing authority, cost dearly and difficult management, practical operation have little significance;Extract language
Whether key feature data are come directly to being tampered that make decisions be researcher's asking of wanting to capture all the time in sound signal
Topic.
The phase spectrum sensitive to signal cutout and instantaneous frequency spectrum are used as feature in the ENF ingredients of present invention selection signal,
Carry out tampering detection;The present invention is tested using the voice signal of three databases, and using deep learning method depth with
Machine forest carries out Model Construction, ensure that the adaptivity of the program and the degree of automation can be applied to actual conditions.
Invention content
In view of the problems of the existing technology, the present invention provides one kind based on to ENF phase spectrums and instantaneous frequency spectrum point
The audio authentication method of analysis.The present invention analyzes the phase spectrum and frequency spectrum of ENFC by extracting the ENFC in voice signal, carries
Take phase and frequency feature.Fusion Features are carried out to phase spectrum signature and frequency spectrum signature using DCA methods, it is random using depth
Forest carries out Model Construction to fusion feature, and obtained model can distort carry out decision to whether arbitrary measured signal passes through,
Realize voice signal insertion, the automatic detection of delete operation.This method is by merging phase representative in ENF ingredients
With instantaneous frequency feature, and use deep learning method training pattern, obtain that automatic detection model can be carried out, improve detection
Efficiency realizes the automation of digital audio tampering detection.
The invention is realized in this way a kind of digital audio true and false based on to ENF phase spectrums and instantaneous frequency spectrum analysis
Identification method, including:It is pre-processed, including down-sampling and narrow-band filtering, is obtained with mains frequency to measured signal first
Narrow band signal centered on (Electirc Network Frequency, ENF) standard frequency;Then ENF signals are carried out special
Sign extraction analyzes the phase spectrum and instantaneous frequency spectrum of ENF signals, extracts the phase spectrum fluctuation characteristic of ENF signals, phase spectrum and frequency
Rate composes fitting parameter feature;By differentiating correlation analysis (discriminant correlation analysis, DCA) method
Fusion Features are carried out, maximize the correlation between different feature sets, while eliminating correlation between class, and restricted interior phase
Guan Xing;Depth random forest is finally applied to carry out Model Construction to the feature after fusion, trained model carries out transfer learning,
After i.e. model preserves, carry out decision whether can be tampered to arbitrary measured signal.The present invention is based on the ENF marks in measured signal
Remember that signal carries out tampering detection, extraction ENF signals affected phase and frequency feature because distorting, and this method is to carrying
The feature set taken carries out DCA Fusion Features, is trained classification to the feature after fusion using depth random forest method, obtains
To disaggregated model, which can be obtained good detection result for the insertion of signal and deletion situation, it is multiple to reduce calculating
Miscellaneous degree, substantially increases classification accuracy, can realize automatic classifier system.
Specifically include following steps:
Step 1:It is pre-processed to measured signal;
Step 2:The feature extraction of phase spectrum and frequency spectrum is carried out to the ENF ingredients in signal;
Step 3:Fusion Features are carried out to multiple feature sets of extraction using DCA methods;
Step 4:Model Construction is carried out to the feature after fusion using depth random forest, can be determined to measured signal
Plan.
Further, step 1, following steps are specifically included:
Step 1.1:X [n] is pre-processed to measured signal, and pretreatment includes down-sampling, goes DC component, obtains xd
[n];
Step 1.2:The signal x of down-sampling will be passed through in step 1.1d[n], by centre frequency at ENF standard frequencies
Bandpass filter, obtain the ENF ingredients x in signalENFC[n]。
Further, step 2, following steps are specifically included:
Step A1:To xENFC[n] carries out being based on DFT1Phase Power estimation, extraction phase spectrum fluctuation characteristic F;
Step A2:To xENFC[n] carries out the instantaneous frequency Power estimation based on Hilbert;
Step A3:It carries out curve fitting respectively to phase spectrum and frequency spectrum, extracts phase spectrum fit characteristicWith instantaneous frequency
Rate composes fit characteristic
Further, in step A1, to xENFC[n] carries out being based on DFT1Phase Power estimation, first to xENFC[n] signal into
Leaf transformation DFT in the conventional N point discrete Fouriers of row, to be based on DFT0Phase estimation, obtain estimation phaseBased on DFT1Phase
Position estimation is in DFT0On the basis of phase estimation, calculate xENFCThe approximate first derivative of [n] at point n:
x′ENFC[n]=fd(xENFC[n]-xENFC[n-1])
In conjunction with approximate first derivative andThe phase estimation of higher order is carried out, and linear interpolation is carried out to estimated result,
Obtain phase spectrum estimated result, extraction phase spectrum fluctuation characteristic F;
In step A2, to xENFC[n] carries out the instantaneous Frequency Estimation converted based on Hilbert, obtains x firstENFC[n's]
Analytical function:
x(a) ENFC[x]=xENFC[x]+i*Η{xENFC[x] },
WhereinΗ represents Hilbert transformation;Instantaneous frequency is Η { xENFC[n] } phase angle change rate, estimation
The instantaneous frequency f [n] of ENF signals removes oscillation and boundary effect to f [n], builds xENFC[n] instantaneous frequency is composed;
In step A3, according to xENFCThe characteristics of phase spectrum and frequency spectrum of [n], respectively use Sum of Sines and
Gaussian comes fit phase spectrum and frequency spectral curve;
Sum of Sines expression formula forms:
Gaussian expression formula forms:
Wherein expression argument is fit characteristic,
Further, step 3, it specifically includes:
The target of Fusion Features be by the relevant information in two or more feature vectors be combined into one it is more single than any
Input feature value has more the information of discrimination, or in the case where intrinsic dimensionality is excessive, and spy is reduced by Fusion Features
Sign dimension can still reach and the approximate accuracy of high dimensional feature.Using the phase for differentiating that correlation analysis DCA will be obtained in step 2
Position feature set and frequecy characteristic collection carry out Fusion Features, and DCA is carried out by the correlation two-by-two between maximizing two feature sets
Fusion Features, and restricted interior correlation.The conversion of feature set is calculated by maximizing the covariance matrix between feature set
Matrix, while ensureing the diagonalization of scatter matrix in class.
Further, step 4, it specifically includes:
Step 4.1:Model Construction is carried out to the feature after fusion using depth random forest;
Depth random forest is a kind of deep neural network model, can be used for classifying.By fusion feature part for training
Depth random forest, the training process of depth random forest is different with traditional random forest, the variation and layer that it can be according to precision
Number limits and automatically determines the model parameters such as the number of plies, will not stop after training precision is not promoted or the number of plies reaches maximum value
Training, using classification results at this time as final classification precision.
Step 4.2:After preservation model decision is carried out to whether arbitrary measured signal is tampered.
The number of plies and structural parameters of obtained depth random forest after the completion of the training process of depth random forest are constituted
The fusion feature disaggregated model of gained of the invention, can carry out arbitrary measured signal fusion feature classification and decision.
Another object of the present invention is to provide described in a kind of realize based on to ENF phase spectrums and instantaneous frequency spectrum analysis
The computer program of digital audio authenticity identification method.
Another object of the present invention is to provide described in a kind of realize based on to ENF phase spectrums and instantaneous frequency spectrum analysis
The letter digital audio signal processing system of digital audio authenticity identification method.
Another object of the present invention is to provide a kind of computer readable storage mediums, including instruction, when it is in computer
When upper operation so that computer execution is described to be reflected based on the digital audio true and false to ENF phase spectrums and instantaneous frequency spectrum analysis
Determine method.
In conclusion advantages of the present invention and good effect are:
The present invention analyzes phase spectrum sensitive to signal cutout in ENF signals and instantaneous frequency spectrum, and extraction is effective respectively
Feature set, and the feature set extracted is handled;
The feature-based fusion technology that the present invention uses carries out characteristic processing, improves and knows while reducing intrinsic dimensionality
Other gap carries out model training using deep learning method, substantially increases the accuracy rate of the passive tampering detection of digital audio;
The present invention is high for complex environment recording and noisy speech stability, has very strong robustness.
The present invention is that the accuracy of the passive tampering detection of digital audio and automation propose a kind of algorithm of popularity.
The experimental data that the present invention uses come from three different databases totally 500 voices (including original language
Sound and distort voice), import these voice signals using MATLAB, it is special to extract the fluctuation of ENF content consistencies by inventive step 1
Sign.According to step 2, phase fluctuation and instantaneous frequency fluctuation are fitted using 5 sin cores and 5 Gaussian kernels;According to step
3, using phase fluctuation feature and frequency fluctuation feature as a feature set, DCA Fusion Features are carried out, two-dimentional fusion is obtained
Feature is characterized addition label, uses ten folding cross validations to fusion feature using depth random forest, is finally obtaining classification just
True rate reaches 99.8%.
Description of the drawings
Fig. 1 is the digital audio true and false provided in an embodiment of the present invention based on to ENF phase spectrums and instantaneous frequency spectrum analysis
Identification method flow chart.
Fig. 2 is provided in an embodiment of the present invention based on DFT1Phase spectrum feature extraction flow chart;
Fig. 3 is the embodiment of the present invention based on Hilbert transformation instantaneous frequency spectrum signature extraction flow charts.
Specific implementation mode
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to embodiments, to the present invention
It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to
Limit the present invention.
Referring to Fig.1, a kind of digital audio true and false based on to ENF phase spectrums and instantaneous frequency spectrum analysis provided by the invention
Identification method includes the following steps:
Step 1:It is pre-processed to measured signal;
Specific implementation includes following sub-step:
Step 1.1:X [n] is pre-processed to measured signal, including down-sampling, goes DC component, obtains xd[n];
In view of (over-sampling can carry for frequency alias effect, signal message loss and the signal-to-noise ratio of signal in the present embodiment
The signal-to-noise ratio of high RST) balance, by the resampling frequency f of signaldIt is set to 1000HZ or 1200HZ (by the ENF frequencies of standard
Rate is placed on ω0=π/10rad/sample).
Step 1.2:The signal x of down-sampling will be passed through in step 1.1d[n], by centre frequency at ENF standard frequencies
Bandpass filter, obtain the ENF ingredients x in signalENFC[n]。
The present embodiment carries out narrow-band filtering using the linear zero phase FIR filter of 10000 ranks prevents phase delay.Center
Frequency is at ENF standard frequencies, bandwidth 0.6HZ, passband ripple 0.5dB, stopband attenuation 100dB.Use high-grade filting
Device is ideal narrow band signal in order to obtain.Zero padding (zero padding) refer to the end of time-domain signal plus zero with
The way for increasing signal length, frequency resolution can be improved before DFT using zero padding, and frequency is more accurately found in help
Peak point in spectrum.
Step 2:The feature extraction of phase spectrum and frequency spectrum is carried out to the ENF ingredients in signal;
Specific implementation includes following sub-step:
Step A1:To xENFC[n] carries out being based on DFT1Phase Power estimation, extraction phase spectrum fluctuation characteristic F;
Such as Fig. 2, first to xENFC[n] signal carries out leaf transformation (DFT) in conventional N point discrete Fouriers, obtains X (k), enables
kpeakAs every frame | X (k) | the integer index of maximum value is referred to as based on DFT0Phase estimation:
Calculate ENF signals xENFCThe approximate first derivative of [n] at point n:
x′ENFC[n]=fd(xENFC[n]-xENFC[n-1]) (2)
To x 'ENFC[n] carries out DFT0Phase estimation obtains | X ' (k) |, will | X ' (k) | it is multiplied by a scale coefficient F (k).
DFT can be obtained in this way0[k]=X (k) and DFT1[k]=F (k) | X ' (k) |.xENFC[n estimates that frequency values are
ENFC, which is a narrow band signal, to be write as:xENFC[n]=acos (ω0n+
φ0), wherein ω0=2 π fENFC/fd, φ0Represent xENFCInitial phase, and fENFCBe ENF it is actual frequency.It is pushed away according to mathematics
Calculation can obtain:
Whereinθ represents x 'ENFCEstimation phase, to X ' (k) carry out linear interpolation to obtain more
Accurate value.Based on DFT1The estimation phase spectrum of method is:
Characteristic quantity F is calculated the phase fluctuation feature of ENFC is described.It enablesIt is corresponding n-thbThe estimation phase of frame
Position,Wherein 2≤nb≤NBlock,It indicatesFrom nb=2 arrive NBlockAverage value.
Step A2:To xENFC[n] carries out the instantaneous frequency Power estimation based on Hilbert;
Such as Fig. 3, to signal xENFC[n] carries out discrete Hilbert transform.X is obtained firstENFCThe analytical function of [n]:x(a) ENFC[x]=xENFC[x]+i*Η{xENFC[x] }, whereinΗ represents Hilbert transformation.Instantaneous amplitude is Η { xENFC
[n] } amplitude, instantaneous frequency is Η { xENFC[n] } phase angle change rate.Estimate the instantaneous frequency f [n] of ENF signals.It is using
Due to there is numerical radius during Hilbert transformation, so obtained f [n] there are certain unwanted oscillation, needs further
Low-pass filtering, removal oscillation are carried out to f [n].Due to the boundary effect of Frequency Estimation, remove f [n] each 2000 sampled points end to end,
Last gained f [n] is the instantaneous frequency Power estimation of ENFC.
Step A3:It carries out curve fitting respectively to phase spectrum and frequency spectrum, extracts phase spectrum fit characteristicWith instantaneous frequency
Rate composes fit characteristic
The characteristics of the present embodiment is distributed for ENF phase distributions and instantaneous frequency, uses different analytical expressions respectively
Discrete data point group is fitted.For phase or frequency curve selection selection analytical expression standard be:The expression formula
Original signal curve and editor's signal curve can be not only fitted respectively, and the difference of the two can be embodied in parameter
On.Based on this standard, the present embodiment has selected Sum of two fitting expressions of Sines and Gaussian to be intended respectively
Phase curve and frequency curve are closed, wherein expression argument is fitting parameter feature.
Analytical expression Sum of Sines are adapted to fit phase spectrum, and form is:
Wherein a is amplitude, and b is frequency, and c is the phase constant of each sine wave item, and n refers to the quantity of this sequence, is taken
Value range is 1≤n≤9.It enablesFor phase spectrum fit characteristic, i.e.,:
Analytical expression Gaussian is adapted to fitting peak value, and form is:
Wherein a is the amplitude of peak value, and b is the position where peak value, and c is related with the secondary lobe at peak, and n refers to being fitted how many
Peak value, value range are 1≤n≤8.It enablesFor frequency spectrum fit characteristic, i.e.,:
Step 3:Fusion Features are carried out to multiple feature sets of extraction using DCA methods;
The phase property collection obtained in step 2 and frequecy characteristic collection feature is carried out using differentiation correlation analysis (DCA) to melt
It closes.DCA carries out effective Fusion Features by the correlation two-by-two between maximizing two feature sets, while eliminating correlation between class
Property, and restricted interior correlation.Intrinsic dimensionality can also be reduced simultaneously, reduce the gap on recognition result.DCA is to apply to ask
With the feature-based fusion of method, having reduces intrinsic dimensionality, while the advantages of reduce the gap on recognition result.
Assuming that X ∈ Rp×nWith Y ∈ Rq×nIndicate that two matrixes, each matrix include the n training from different mode
Feature vector.If the sample in data matrix is collected from c independent classes.N row in this way in data matrix can be with
It is divided into c independent group, wherein niDependent of dead military hero is in ithClassEnable xij∈ X are indicated and ithJth in classthSample
This corresponding feature vector.WithX is indicated respectivelyijIn ithMean value in class and in entire feature set, i.e.,Scatter matrix is defined as between class
Wherein
If characteristic is more than classification number (p>>C), calculate covariance matrixIt will be than calculating
It is more prone to.By rightMapping can effectively obtainUpper significant feature vector.Therefore it only needs
Find the covariance matrix of c × c dimensionsFeature vector.As can be distinguished well between fruit, thenIt will
It is a diagonal matrix, becauseIt is symmetric positive semidefinite matrix, the present invention can be by becoming its diagonalization of changing commanders:
P is orthogonal eigenvectors matrix,It is diagonal matrix of the nonnegative real number characteristic value by sequence sequence of successively decreasing.Q(c×r)For
The matrix of the r feature vector compositions from matrix P corresponds to first r maximum nonzero eigenvalue.Have:
Mapping in this way can obtain SbxMiddle r important feature vectors:Q→ΦbxQ
(ΦbxQ)TSbx(ΦbxQ)=Λ(r×r), (13)
Wbx=ΦbxQΛ-1/2S can be unified by being one kindbxThe transformation for reducing data matrix dimension X simultaneously is tieed up from p dimensions to r.
I.e.:
X ' is the projections of X in space, and scatter matrix is I between class, and class is separable.Pay attention to being up to c-1 here
A generalized eigenvalue, therefore the upper limit of r is c-1, other upper limits of r are made of the order of data matrix, i.e. r≤min (c-1, rank
(X),rank(Y))。
Similar above-mentioned method handles second feature collection Y, and finds transformation matrix Wby, unify between the class of second mode to spread
Matrix SbyThe dimension for reducing data matrix Y simultaneously is tieed up from q dimensions to r.
Φ′bxWith Φ 'byMore new capital be r × c non-square quadrature matrix.In spite of Sb′x=Sb′y=I, matrixWithAll it is stringent diagonal matrixElement wherein on diagonal line is non-right close to 1
Element on linea angulata is close to 0.It is minimum related that this so that the center of class has before, therefore can well be divided class
From.Next needing to enable the character pair that the feature in same feature set is only concentrated with another feature has non-zero correlation.In order to
Realize that this target, the present invention need the scatter matrix between the class of transformation matrix to carry out diagonalization, i.e. S 'xy=X ' Y 'T.Using strange
Different value decomposes (SVD) diagonalization Sx′y。
X ' and Y ' orders therein are all r, S 'xy(r×r)It is non-reduced.It is the member on a diagonal matrix and leading diagonal
Element is all nonzero value.Enable Wcx=U Σ-1/2, Wcy=V Σ-1/2, have:
(UΣ-1/2)TS′xy(VΣ-1/2)=I, (19)
It is connected to the covariance matrix S ' between feature setxy.Next feature set is converted:
WhereinIt is the final transformed matrix of X and Y respectively.It can easily be proven that after transformation
Scatter matrix is still diagonal between the class of feature set, therefore, can be separated between class.Class between scatter matrix be:
It is known in formula (14)And U is an orthogonal matrix, is had:
Here it can equally proveIt is diagonal matrix.Obtain converting characteristic collectionRepresent the association between feature
Variance is a leading diagonal Striking symmetry matrix, shows that the correlation between single feature concentrates different characteristic is minimum.Transformation
Feature setCovariance between representative sample is block diagonal matrix, shows that sample has more with the sample in same class
High correlation.
Step 4:Model Construction is carried out to the feature after fusion using depth random forest, can be determined to measured signal
Plan.
Step 4.1:Model Construction is carried out to the feature after fusion using depth random forest;
The present invention needs to carry out data the data volume of the scanning enlarged sample of more granularities first, is carried out by sliding window
Sampling.Window size is 100, step-length 1, then the sample that 301 groups of characteristics are 100 can be obtained after sampling, but these samples are complete
An original sample of portion source, so the quantity to sample is expanded.It is then complete using a random forest and one
Full random forest is trained.The generation of decision tree in completely random forest is need not to calculate gini index or entropy increasing
Benefit randomly selects an attribute as attribute is divided gradually to generate completion.Assuming that the present invention needs to do three classification, then pass through
As soon as generating the characteristic information that 301 groups of dimensions are three after a random forest and completely random forest respectively, generated after combination
1806 dimension datas.In the generation and test process of the two random forests and completely random forest, cross validation is rolled over using k
Mode predict, first using k-1 groups again this be also equivalent to 300 groups of data and train random forest, with other one
It organizes and is tested in data distribution area k-1 numbers, then test set is done to be averaged also just having obtained the output of random forest, every group
Data do a test, and recycling k times also just can still obtain the output of k groups.Certainly feature extraction is being carried out using sliding window
When can also set different serial ports sizes and different step-lengths, then by after random forest and completely random forest again
It combines again together.
In cascading forest, by output (the 3*4=12 dimensions of two completely random forests and two common random forests
According to) and initial data (referring to 3618 dimension datas exported after the scanning of more granularities) series connection after as next layer of input (12+
3618=3630 dimension datas) because being all that the output series connection of last layer has been come in each layer of input has each time
Therefore 3630 dimension datas are also equivalent to and are corrected to the parameter of random forest, so, the number of plies of depth random forest is not
The present invention oneself setting, it can be depending on the variation of precision and number of plies limitation, when training precision does not have promotion or the number of plies
Reach after maximum value will deconditioning, using classification results at this time as final classification precision.
Step 4.2:Whether can be tampered to arbitrary measured signal after preservation model and carry out decision.
The number of plies and structural parameters of obtained depth random forest after the completion of the training process of depth random forest are constituted
The fusion feature disaggregated model of gained of the invention, can carry out arbitrary measured signal fusion feature classification and decision.
With reference to specific embodiment/experiment/emulation credit analysis, the invention will be further described.
The experimental data that the present invention uses come from three different databases totally 500 voices (including original language
Sound and distort voice), import these voice signals using MATLAB, it is special to extract the fluctuation of ENF content consistencies by inventive step 1
Sign.According to step 2, phase fluctuation and instantaneous frequency fluctuation are fitted using 5 sin cores and 5 Gaussian kernels;According to step
3, using phase fluctuation feature and frequency fluctuation feature as a feature set, DCA Fusion Features are carried out, two-dimentional fusion is obtained
Feature is characterized addition label, uses ten folding cross validations to fusion feature using depth random forest, is finally obtaining classification just
True rate reaches 99.8%.
In the above-described embodiments, can come wholly or partly by software, hardware, firmware or its arbitrary combination real
It is existing.When using entirely or partly realizing in the form of a computer program product, the computer program product include one or
Multiple computer instructions.When loading on computers or executing the computer program instructions, entirely or partly generate according to
Flow described in the embodiment of the present invention or function.The computer can be all-purpose computer, special purpose computer, computer network
Network or other programmable devices.The computer instruction can store in a computer-readable storage medium, or from one
Computer readable storage medium is transmitted to another computer readable storage medium, for example, the computer instruction can be from one
A web-site, computer, server or data center pass through wired (such as coaxial cable, optical fiber, Digital Subscriber Line (DSL)
Or wireless (such as infrared, wireless, microwave etc.) mode is carried out to another web-site, computer, server or data center
Transmission).The computer read/write memory medium can be that any usable medium that computer can access either includes one
The data storage devices such as a or multiple usable mediums integrated server, data center.The usable medium can be magnetic Jie
Matter, (for example, floppy disk, hard disk, tape), optical medium (for example, DVD) or semiconductor medium (such as solid state disk Solid
State Disk (SSD)) etc..
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all essences in the present invention
All any modification, equivalent and improvement etc., should all be included in the protection scope of the present invention made by within refreshing and principle.