CN105787497A - Account-stealing case analysis method and device - Google Patents
Account-stealing case analysis method and device Download PDFInfo
- Publication number
- CN105787497A CN105787497A CN201410815613.0A CN201410815613A CN105787497A CN 105787497 A CN105787497 A CN 105787497A CN 201410815613 A CN201410815613 A CN 201410815613A CN 105787497 A CN105787497 A CN 105787497A
- Authority
- CN
- China
- Prior art keywords
- neural network
- network model
- deep neural
- layer
- characteristic information
- 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
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The embodiments of the invention relates to an account-stealing case analysis method and a device. The method comprises the steps of acquiring the first feature data of a to-be-analyzed account-stealing case; according to the first feature data, constructing an original feature input vector; adopting the original feature input vector as an input vector of an original feature input layer of a depth neural network model, and acquiring the analysis result of the to-be-analyzed account-stealing case through the output layer of the depth neural network model. The depth neural network model comprises at least three layers of recessive feature layers. The output vectors of the recessive feature layers and the input vectors of an original feature input layer are the same in element number. Therefore, the account-stealing case can be analyzed through the depth neural network, so that the working intensity of the case trial personnel is lightened. Meanwhile, the objectivity of judgment results is improved.
Description
Technical field
The application relates to technical field of network security, particularly relates to a kind of robber's account case analysis method and apparatus.
Background technology
Along with development and the evolution of network payment, steal account case and also increasingly trend towards complicating and intelligent.Especially, the proximal segment time starts to emerge acquaintance and operates account falseness and report a case to the security authorities stealing in account case.This acquaintance operates falseness and reports a case to the security authorities the intentional or unintentional fraudulent claim character of ubiquities, and network payment company needs strong examination also to give the response do not compensated decidedly.
Robber's account case analysis method of the prior art generally includes following steps, i.e. data acquisition, data prediction, feature extraction and selection and last pattern classification.Although present most of research and be applied in last category of model, but good feature extraction and selection, often the final accuracy of algorithm is played vital effect.Existing feature extraction and selection are typically all and are accomplished manually, namely final model performance is served conclusive effect by the understanding of business by Modeling Research personnel and engineer, and, the most work time shared by modelling application is all consumed among the calculating and test process of feature extraction and selection.
Therefore, more robber's account case analysis method of the prior art depends on and manually completes, and not only takes time and effort, and court verdict subjectivity is strong.
Summary of the invention
The embodiment of the present application provides a kind of robber's account case analysis method and apparatus, it is possible to alleviates the working strength of using case trial electric personnel, improves the objectivity of court verdict simultaneously.
First aspect, it is provided that a kind of robber's account case analysis method, the method includes:
Obtain the fisrt feature data of robber's account case to be analyzed；
According to described fisrt feature data construct primitive character input vector；
Using the described primitive character input vector input vector as the primitive character input layer of deep neural network model, the analysis result of described robber's account case to be analyzed is obtained by the output layer of described deep neural network model, wherein, described deep neural network model includes the input vector of the recessive character layer of at least three layers, the output vector of described recessive character layer and described primitive character input layer and has identical element number.
Preferably, described according to described fisrt feature data construct primitive character input vector, specifically include:
Described fisrt feature data are normalized and obtain second feature data；
According to described second feature data construct primitive character input vector.
Preferably, before the fisrt feature data of described acquisition robber's account case to be analyzed, described method also includes:
Obtain sample and steal the sample characteristics data of account case and the sample analysis result of described sample robber's account case；
According to described sample characteristics data construct sample primitive character input vector；
Using the described sample primitive character input vector input vector as the primitive character input layer of deep neural network model, using the described sample analysis result output result as the output layer of described deep neural network model, carry out deep neural network model learning.
Preferably, described in carry out deep neural network model learning, specifically include:
Based on minimizing reconstructed error criterion, when meeting performance threshold, successively carry out characteristic information refinement for described deep neural network model；
By the probability that the classification results of statistical sorter modeling acquisition is identical with described sample analysis result, described deep neural network model is carried out Performance Evaluation；
When described Performance Evaluation assessment result meet require time, confirmed described deep neural network model learning；
When the assessment result of described Performance Evaluation is unsatisfactory for requiring, adjusts described performance threshold and re-start characteristic information refinement.
Preferably, described based on minimizing reconstructed error criterion, when meeting performance threshold, successively carry out characteristic information refinement for described deep neural network model, specifically include:
After each characteristic information refines, counter pushing away verifies described characteristic information refines whether meet performance threshold；
When the result be described characteristic information refine meet described performance threshold time, has confirmed this layer characteristic information refinement；
When the result be described characteristic information refine be unsatisfactory for described performance threshold time, adjust the weight matrix of this layer and re-start the characteristic information of this layer and refine.
Second aspect, it is provided that a kind of robber's account case analysis device, this device includes: the first data capture unit, primary vector construction unit and analytic unit；
Described first data capture unit, for obtaining the fisrt feature data of robber's account case to be analyzed；
Described primary vector construction unit, for the fisrt feature data construct primitive character input vector obtained according to described first data capture unit；
Described analytic unit, for input vector as the primitive character input layer of deep neural network model of primitive character input vector that described primary vector construction unit is built, the analysis result of described robber's account case to be analyzed is obtained by the output layer of described deep neural network model, wherein, described deep neural network model includes the input vector of the recessive character layer of at least three layers, the output vector of described recessive character layer and described primitive character input layer and has identical element number.
Preferably, described primary vector construction unit specifically includes:
Normalized subelement, obtains second feature data for described fisrt feature data are normalized；
Vector builds subelement, for the second feature data construct primitive character input vector obtained according to described normalized subelement.
Preferably, described device also includes:
Second data capture unit, for, before the fisrt feature data of described first data capture unit acquisition robber's account case to be analyzed, obtaining sample and steal the sample characteristics data of account case and the sample analysis result of described sample robber's account case；
Secondary vector construction unit, for the sample characteristics data construct sample primitive character input vector obtained according to described second data capture unit；
Unit, for input vector as the primitive character input layer of deep neural network model of sample primitive character input vector that described secondary vector construction unit is built, the sample analysis result obtained by described second data capture unit, as the output result of the output layer of described deep neural network model, carries out deep neural network model learning.
Preferably, described unit specifically includes:
Feature refines subelement, for based on minimizing reconstructed error criterion, when meeting performance threshold, successively carrying out characteristic information refinement for described deep neural network model；
Classification subelement, the feature for refining subelement refinement according to described feature obtains classification results by grader modeling；
Performance Evaluation subelement, for the probability that the classification results by adding up the acquisition of described classification subelement is identical with the sample analysis result that described second data capture unit obtains, carries out Performance Evaluation to described deep neural network model；
Confirm subelement, for when described Performance Evaluation subelement carry out Performance Evaluation assessment result meet require time, confirmed described deep neural network model learning；
Adjust subelement, for when the assessment result that described Performance Evaluation subelement carries out Performance Evaluation is unsatisfactory for requiring, adjusting described performance threshold and re-started characteristic information refinement by described feature refinement subelement.
Preferably, described feature refinement subelement specifically includes:
Authentication module, for after each characteristic information refines, counter pushing away verifies described characteristic information refines whether meet performance threshold；
Confirm module, for when the result of described authentication module be described characteristic information refine meet described performance threshold time, has confirmed this layer characteristic information refinement；
Adjusting module, for when the result of described authentication module be described characteristic information refine be unsatisfactory for described performance threshold time, adjust the weight matrix of this layer and re-start the characteristic information of this layer and refine.
In robber's account case analysis method that the application provides, first obtain the fisrt feature data of robber's account case to be analyzed, then according to fisrt feature data construct primitive character input vector, using the primitive character input vector input vector as the primitive character input layer of deep neural network model, the analysis result of robber's account case to be analyzed is obtained by the output layer of deep neural network model, wherein, deep neural network model includes the recessive character layer of at least three layers, the input vector of the output vector of recessive character layer and primitive character input layer has identical element number.Therefore, the application is undertaken stealing account case analysis by deep neural network, it is possible to alleviates the working strength of using case trial electric personnel, improves the objectivity of court verdict simultaneously.Further, deep neural network is more and more abstract from bottom to high-level characteristic, is increasingly able to express the intention of initial characteristics, and this feature extraction as far as possible not breaking one's promise breath is optimal strategy, and the comprehensive explanation of feature can be reached gratifying precision.
Accompanying drawing explanation
Robber's account case analysis method flow diagram that Fig. 1 provides for the embodiment of the present application one；
Fig. 2 is the deep neural network model structure schematic diagram in the embodiment of the present application one；
Robber's account case analysis method flow diagram that Fig. 3 provides for the embodiment of the present application two；
Fig. 4 is that the feature in the embodiment of the present application two refines schematic diagram of mechanism；
Fig. 5 is the model performance assessment schematic diagram in the embodiment of the present application two；
Robber's account case analysis apparatus structure schematic diagram that Fig. 6 provides for the embodiment of the present application three.
Detailed description of the invention
For making the purpose of the embodiment of the present application, technical scheme and advantage clearly, below in conjunction with the accompanying drawing in the embodiment of the present application, technical scheme in the embodiment of the present application is clearly and completely described, obviously, described embodiment is some embodiments of the present application, rather than whole embodiments.Based on the embodiment in the application, the every other embodiment that those of ordinary skill in the art obtain under not making creative work premise, broadly fall into the scope of the application protection.
For ease of the understanding to the embodiment of the present application, being further explained explanation below in conjunction with accompanying drawing with specific embodiment, embodiment is not intended that the restriction to the embodiment of the present application.
Robber's account case analysis method flow diagram that Fig. 1 provides for the embodiment of the present application one, the executive agent of described method can be the special equipment performing described method, it is also possible to for existing terminal unit, for instance computer, as it is shown in figure 1, the method specifically includes:
Step 101, obtains the fisrt feature data of robber's account case to be analyzed.
In the embodiment of the present application, whether belong to acquaintance and operate account falseness in order to accurate analysis goes out to steal account case and report a case to the security authorities, as much as possible can obtain the characteristic information stealing account case, such as, environmental characteristic information, performance characteristic information, transaction feature information, temporal characteristics information and characteristics of objects information, using features described above information as fisrt feature data, in order to follow-up can be analyzed stealing account case according to fisrt feature data.
Step 102, according to fisrt feature data construct primitive character input vector.
Wherein, owing to fisrt feature packet is containing polytype data, such as, logical data (true or false) and numeric type data (100), in order to balanced polytype data are for the impact of case analysis result, first fisrt feature data can be normalized and obtain second feature data, then according to second feature data construct primitive character input vector.
Above-mentioned normalized is specifically as follows second feature data fisrt feature data being converted in the numerical intervals being limited in 0 to 1 after treatment, for instance, logical truth is converted into data 1, transaction count 3 is converted into data 0.3.Data below can be made to process by normalized convenient, and ensure that when program is run, convergence is accelerated.
Step 103, using the primitive character input vector input vector as the primitive character input layer of deep neural network model, obtains the analysis result of robber's account case to be analyzed by the output layer of deep neural network model.
Wherein, above-mentioned deep neural network model includes the recessive character layer of at least three layers, and the input vector of the output vector of recessive character layer and primitive character input layer has identical element number.
Fig. 2 is the deep neural network model structure schematic diagram in the embodiment of the present application one, with reference to Fig. 2, deep neural network model in the embodiment of the present application can have the recessive character layer of at least three layers, such as, there is five layers even the recessive character layer of ten layers, Fig. 2 only has three layers recessive character layer exemplarily with deep neural network model, compared to conventional Feature Selection, the application is by changing features successively, between layers by effective Feature Space Transformation, thus obtaining the feature representation of more accurately refine, it is easy to follow-up obtain case analysis result accurately.
In Fig. 2, prototype network structure includes following variable:
Primitive character input vector: x=(x1, x2 ..., xN1)；
Recessive character layer L1 output variable: h1=(h11, h12 ..., h1n)；
Recessive character layer L2 output variable: h2=(h21, h22 ..., h2n)；
Recessive character layer L3 output variable: h3=(h31, h32 ..., h3n)；
Output layer output vector: z；
The connection weights of primitive character input layer and recessive character layer L1: wih1；
The connection weights of recessive character layer L1 and recessive character layer L2: wih2；
The connection weights of recessive character layer L2 and recessive character layer L3: wih3；
The connection weights of recessive character layer L3 and output layer: whj.
Wherein, primitive character input layer contains the Criminal characteristic information of all retrievable robber's account cases, these characteristic informations include the magnanimity attribute field based on several storehouses, such as identity information, region, place, amount of money involved and trading volume etc., and a large amount of characteristic variables accumulated along with the continuous experience accumulation of wind control business, for instance transaction count in the past period and dealing money, different environment cumulative amounts etc..And recessive character layer L1 on this basis is simplifying and refining primitive character input layer, it is achieved increasingly complex and structurized feature representation, say, that higher level character representation.The characteristic variable that recessive character layer comprises often does not possess the physical meaning of dense thickness, but possesses the feature representation of more accurately refine.
By that analogy, these levels are that level goes forward one by one, it is that the organic assembling expressed by low-level image feature forms that high-level characteristic is expressed, and it is achieved in that by connection weights between layers, use the sigmod function everywhere can led to map as the input-output function between each level in prototype network structure.
Input function input is: f (x)=x1w1+x2w2+...+xnwn；
Output function output is:
Owing in advance characteristic having been carried out normalized, the primitive character input vector then built, therefore can improve susceptiveness and the training speed of deep neural network model the saturation region of fair avoidance sigmod.
Feature self study process under above-mentioned deep neural network framework, it is to avoid manually choose and take time and effort and shortcoming that precision is inaccurate, be a kind of brand-new robber's account case feature extraction thinking.Robber's account case Feature Selection of the present invention is a kind of degree of depth learning method, and compared to the intelligent method such as neutral net of conventional only 1 layer of hidden node, this model usually haves three layers, 5 layers of even 10 layers of hidden node, analyzes result more accurate.
In the embodiment of the present application; before stealing account case analysis utilizing deep neural network model to carry out; generally also include the building process of deep neural network model; deep neural network model specifically can be trained by above-mentioned building process by the characteristic of robber's account case of case result known in Sample Storehouse, so that it is determined that the connection weights of each layer construct deep neural network model in deep neural network model.
Above-mentioned deep neural network model is trained, specifically may include that obtaining sample steals the sample characteristics data of account case and the sample analysis result of account case stolen by described sample；According to described sample characteristics data construct sample primitive character input vector；Using the described sample primitive character input vector input vector as the primitive character input layer of deep neural network model, using the described sample analysis result output result as the output layer of described deep neural network model, carry out deep neural network model learning.
Wherein, described in carry out deep neural network model learning, specifically may include that based on minimizing reconstructed error criterion, when meeting performance threshold, successively carry out characteristic information refinement for described deep neural network model；By the probability that the classification results of statistical sorter modeling acquisition is identical with described sample analysis result, described deep neural network model is carried out Performance Evaluation；When described Performance Evaluation assessment result meet require time, confirmed described deep neural network model learning；When the assessment result of described Performance Evaluation is unsatisfactory for requiring, adjusts described performance threshold and re-start characteristic information refinement.
Wherein, described based on minimizing reconstructed error criterion, when meeting performance threshold, successively carry out characteristic information refinement for described deep neural network model, specifically may include that after each characteristic information refines, counter pushing away verifies described characteristic information refines whether meet performance threshold；When the result be described characteristic information refine meet described performance threshold time, has confirmed this layer characteristic information refinement；When the result be described characteristic information refine be unsatisfactory for described performance threshold time, adjust the weight matrix of this layer and re-start the characteristic information of this layer and refine.
By above-mentioned processing procedure it can be seen that the application is undertaken stealing account case analysis by deep neural network, it is possible to alleviate the working strength of using case trial electric personnel, improve the objectivity of court verdict simultaneously.Further, deep neural network is more and more abstract from bottom to high-level characteristic, is increasingly able to express the intention of initial characteristics, and this feature extraction as far as possible not breaking one's promise breath is optimal strategy, and the comprehensive explanation of feature can be reached gratifying precision.
Robber's account case analysis method flow diagram that Fig. 3 provides for the embodiment of the present application two, the method carry out in the process of deep neural network model learning, after each feature is refined, it is necessary to counter pushing away verifies whether this refinement reaches certain performance threshold, as it is shown on figure 3, the method specifically includes:
Step 301, obtains sample and steals the sample characteristics data of account case and the sample analysis result of described sample robber's account case.
Step 302, according to described sample characteristics data construct sample primitive character input vector.
Step 303, using the sample primitive character input vector input vector as the primitive character input layer of deep neural network model, using the sample analysis result output result as the output layer of deep neural network model, carries out deep neural network model learning.
Wherein, carry out the learning process that deep neural network model learning can include stealing the refinement of account case feature and model performance two stages of assessment, specifically: based on minimizing reconstructed error criterion, when meeting performance threshold, successively carry out characteristic information refinement for described deep neural network model；By the probability that the classification results of statistical sorter modeling acquisition is identical with described sample analysis result, described deep neural network model is carried out Performance Evaluation；When described Performance Evaluation assessment result meet require time, confirmed described deep neural network model learning；When the assessment result of described Performance Evaluation is unsatisfactory for requiring, adjusts described performance threshold and re-start characteristic information refinement.
Fig. 4 is that the feature in the embodiment of the present application two refines schematic diagram of mechanism, described based on minimizing reconstructed error criterion, when meeting performance threshold, characteristic information refinement is successively carried out for described deep neural network model, specifically may include that after each characteristic information refines, counter pushing away verifies described characteristic information refines whether meet performance threshold；When the result be described characteristic information refine meet described performance threshold time, has confirmed this layer characteristic information refinement；When the result be described characteristic information refine be unsatisfactory for described performance threshold time, adjust the weight matrix of this layer and re-start the characteristic information of this layer and refine.
Fig. 5 is the model performance assessment schematic diagram in the embodiment of the present application two, with reference to Fig. 5, after feature has been refined automatically, only a small amount of tape label sample can be used to determine to the final mask realizing deep neural network, and classifier methods can use logistic regression algorithm or the algorithm of support vector machine of routine.
Step 304, obtains the fisrt feature data of robber's account case to be analyzed.
Described fisrt feature data are normalized and obtain second feature data by step 305.
Step 306, according to described second feature data construct primitive character input vector.
Step 307, using the described primitive character input vector input vector as the primitive character input layer of deep neural network model, obtains the analysis result of described robber's account case to be analyzed by the output layer of described deep neural network model.
Wherein, described deep neural network model includes the recessive character layer of at least three layers, and the input vector of the output vector of described recessive character layer and described primitive character input layer has identical element number.
Robber's account case analysis apparatus structure schematic diagram that Fig. 6 provides for the embodiment of the present application three, this device includes: the first data capture unit 601, primary vector construction unit 602 and analytic unit 603；
Described first data capture unit 601, for obtaining the fisrt feature data of robber's account case to be analyzed；
Described primary vector construction unit 602, for the fisrt feature data construct primitive character input vector obtained according to described first data capture unit 601；
Described analytic unit 603, for input vector as the primitive character input layer of deep neural network model of primitive character input vector that described primary vector construction unit 602 is built, the analysis result of described robber's account case to be analyzed is obtained by the output layer of described deep neural network model, wherein, described deep neural network model includes the input vector of the recessive character layer of at least three layers, the output vector of described recessive character layer and described primitive character input layer and has identical element number.
Preferably, described primary vector construction unit 602 specifically includes:
Normalized subelement, obtains second feature data for described fisrt feature data are normalized；
Vector builds subelement, for the second feature data construct primitive character input vector obtained according to described normalized subelement.
Preferably, described device also includes:
Second data capture unit, for, before described first data capture unit 601 obtains the fisrt feature data of robber's account case to be analyzed, obtaining sample and steal the sample characteristics data of account case and the sample analysis result of described sample robber's account case；
Secondary vector construction unit, for the sample characteristics data construct sample primitive character input vector obtained according to described second data capture unit；
Unit, for input vector as the primitive character input layer of deep neural network model of sample primitive character input vector that described secondary vector construction unit is built, the sample analysis result obtained by described second data capture unit, as the output result of the output layer of described deep neural network model, carries out deep neural network model learning.
Preferably, described unit specifically includes:
Feature refines subelement, for based on minimizing reconstructed error criterion, when meeting performance threshold, successively carrying out characteristic information refinement for described deep neural network model；
Classification subelement, the feature for refining subelement refinement according to described feature obtains classification results by grader modeling；
Performance Evaluation subelement, for the probability that the classification results by adding up the acquisition of described classification subelement is identical with the sample analysis result that described second data capture unit obtains, carries out Performance Evaluation to described deep neural network model；
Confirm subelement, for when described Performance Evaluation subelement carry out Performance Evaluation assessment result meet require time, confirmed described deep neural network model learning；
Adjust subelement, for when the assessment result that described Performance Evaluation subelement carries out Performance Evaluation is unsatisfactory for requiring, adjusting described performance threshold and re-started characteristic information refinement by described feature refinement subelement.
Preferably, described feature refinement subelement specifically includes:
Authentication module, for after each characteristic information refines, counter pushing away verifies described characteristic information refines whether meet performance threshold；
Confirm module, for when the result of described authentication module be described characteristic information refine meet described performance threshold time, has confirmed this layer characteristic information refinement；
Adjusting module, for when the result of described authentication module be described characteristic information refine be unsatisfactory for described performance threshold time, adjust the weight matrix of this layer and re-start the characteristic information of this layer and refine.
Professional should further appreciate that, the object of each example described in conjunction with the embodiments described herein and algorithm steps, can with electronic hardware, computer software or the two be implemented in combination in, in order to clearly demonstrate the interchangeability of hardware and software, generally describe composition and the step of each example in the above description according to function.These functions perform with hardware or software mode actually, depend on application-specific and the design constraint of technical scheme.Professional and technical personnel specifically can should be used for using different methods to realize described function to each, but this realization is it is not considered that exceed scope of the present application.
The method described in conjunction with the embodiments described herein or the step of algorithm can use the software module that hardware, processor perform, or the combination of the two is implemented.Software module can be placed in any other form of storage medium known in random access memory (RAM), internal memory, read only memory (ROM), electrically programmable ROM, electrically erasable ROM, depositor, hard disk, moveable magnetic disc, CD-ROM or technical field.
Above-described detailed description of the invention; the purpose of the application, technical scheme and beneficial effect have been further described; it is it should be understood that; the foregoing is only the detailed description of the invention of the application; it is not used to limit the protection domain of the application; all within spirit herein and principle, any amendment of making, equivalent replacement, improvement etc., should be included within the protection domain of the application.
Claims (10)
1. steal account case analysis method for one kind, it is characterised in that described method includes:
Obtain the fisrt feature data of robber's account case to be analyzed；
According to described fisrt feature data construct primitive character input vector；
Using the described primitive character input vector input vector as the primitive character input layer of deep neural network model, the analysis result of described robber's account case to be analyzed is obtained by the output layer of described deep neural network model, wherein, described deep neural network model includes the input vector of the recessive character layer of at least three layers, the output vector of described recessive character layer and described primitive character input layer and has identical element number.
2. method according to claim 1, it is characterised in that described according to described fisrt feature data construct primitive character input vector, specifically includes:
Described fisrt feature data are normalized and obtain second feature data；
According to described second feature data construct primitive character input vector.
3. method according to claim 1, it is characterised in that before the fisrt feature data of described acquisition robber's account case to be analyzed, described method also includes:
Obtain sample and steal the sample characteristics data of account case and the sample analysis result of described sample robber's account case；
According to described sample characteristics data construct sample primitive character input vector；
Using the described sample primitive character input vector input vector as the primitive character input layer of deep neural network model, using the described sample analysis result output result as the output layer of described deep neural network model, carry out deep neural network model learning.
4. method according to claim 3, it is characterised in that described in carry out deep neural network model learning, specifically include:
Based on minimizing reconstructed error criterion, when meeting performance threshold, successively carry out characteristic information refinement for described deep neural network model；
By the probability that the classification results of statistical sorter modeling acquisition is identical with described sample analysis result, described deep neural network model is carried out Performance Evaluation；
When described Performance Evaluation assessment result meet require time, confirmed described deep neural network model learning；
When the assessment result of described Performance Evaluation is unsatisfactory for requiring, adjusts described performance threshold and re-start characteristic information refinement.
5. method according to claim 4, it is characterised in that described based on minimizing reconstructed error criterion, when meeting performance threshold, successively carries out characteristic information refinement for described deep neural network model, specifically includes:
After each characteristic information refines, counter pushing away verifies described characteristic information refines whether meet performance threshold；
When the result be described characteristic information refine meet described performance threshold time, has confirmed this layer characteristic information refinement；
When the result be described characteristic information refine be unsatisfactory for described performance threshold time, adjust the weight matrix of this layer and re-start the characteristic information of this layer and refine.
6. steal account case analysis device for one kind, it is characterised in that described device includes: the first data capture unit, primary vector construction unit and analytic unit；
Described first data capture unit, for obtaining the fisrt feature data of robber's account case to be analyzed；
Described primary vector construction unit, for the fisrt feature data construct primitive character input vector obtained according to described first data capture unit；
Described analytic unit, for input vector as the primitive character input layer of deep neural network model of primitive character input vector that described primary vector construction unit is built, the analysis result of described robber's account case to be analyzed is obtained by the output layer of described deep neural network model, wherein, described deep neural network model includes the input vector of the recessive character layer of at least three layers, the output vector of described recessive character layer and described primitive character input layer and has identical element number.
7. device according to claim 6, it is characterised in that described primary vector construction unit specifically includes:
Normalized subelement, obtains second feature data for described fisrt feature data are normalized；
Vector builds subelement, for the second feature data construct primitive character input vector obtained according to described normalized subelement.
8. device according to claim 6, it is characterised in that described device also includes:
Second data capture unit, for, before the fisrt feature data of described first data capture unit acquisition robber's account case to be analyzed, obtaining sample and steal the sample characteristics data of account case and the sample analysis result of described sample robber's account case；
Secondary vector construction unit, for the sample characteristics data construct sample primitive character input vector obtained according to described second data capture unit；
Unit, for input vector as the primitive character input layer of deep neural network model of sample primitive character input vector that described secondary vector construction unit is built, the sample analysis result obtained by described second data capture unit, as the output result of the output layer of described deep neural network model, carries out deep neural network model learning.
9. device according to claim 8, it is characterised in that described unit specifically includes:
Feature refines subelement, for based on minimizing reconstructed error criterion, when meeting performance threshold, successively carrying out characteristic information refinement for described deep neural network model；
Classification subelement, the feature for refining subelement refinement according to described feature obtains classification results by grader modeling；
Performance Evaluation subelement, for the probability that the classification results by adding up the acquisition of described classification subelement is identical with the sample analysis result that described second data capture unit obtains, carries out Performance Evaluation to described deep neural network model；
Confirm subelement, for when described Performance Evaluation subelement carry out Performance Evaluation assessment result meet require time, confirmed described deep neural network model learning；
Adjust subelement, for when the assessment result that described Performance Evaluation subelement carries out Performance Evaluation is unsatisfactory for requiring, adjusting described performance threshold and re-started characteristic information refinement by described feature refinement subelement.
10. device according to claim 9, it is characterised in that described feature is refined subelement and specifically included:
Authentication module, for after each characteristic information refines, counter pushing away verifies described characteristic information refines whether meet performance threshold；
Confirm module, for when the result of described authentication module be described characteristic information refine meet described performance threshold time, has confirmed this layer characteristic information refinement；
Adjusting module, for when the result of described authentication module be described characteristic information refine be unsatisfactory for described performance threshold time, adjust the weight matrix of this layer and re-start the characteristic information of this layer and refine.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410815613.0A CN105787497A (en) | 2014-12-23 | 2014-12-23 | Account-stealing case analysis method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410815613.0A CN105787497A (en) | 2014-12-23 | 2014-12-23 | Account-stealing case analysis method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN105787497A true CN105787497A (en) | 2016-07-20 |
Family
ID=56377442
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410815613.0A Pending CN105787497A (en) | 2014-12-23 | 2014-12-23 | Account-stealing case analysis method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105787497A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108133436A (en) * | 2017-11-23 | 2018-06-08 | 科大讯飞股份有限公司 | Automatic method and system of deciding a case |
| CN110858269A (en) * | 2018-08-09 | 2020-03-03 | 清华大学 | Criminal name prediction method and device |
| CN112053273A (en) * | 2020-09-16 | 2020-12-08 | 北京偶数科技有限公司 | Method and device for guiding case analysis and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101276417A (en) * | 2008-04-17 | 2008-10-01 | 上海交通大学 | Method for filtering internet cartoon medium rubbish information based on content |
| CN103400577A (en) * | 2013-08-01 | 2013-11-20 | 百度在线网络技术（北京）有限公司 | Acoustic model building method and device for multi-language voice identification |
| CN103945533A (en) * | 2014-05-15 | 2014-07-23 | 济南嘉科电子技术有限公司 | Big data based wireless real-time position positioning method |
| CN104134062A (en) * | 2014-08-18 | 2014-11-05 | 朱毅 | Vein recognition system based on depth neural network |
-
2014
- 2014-12-23 CN CN201410815613.0A patent/CN105787497A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101276417A (en) * | 2008-04-17 | 2008-10-01 | 上海交通大学 | Method for filtering internet cartoon medium rubbish information based on content |
| CN103400577A (en) * | 2013-08-01 | 2013-11-20 | 百度在线网络技术（北京）有限公司 | Acoustic model building method and device for multi-language voice identification |
| CN103945533A (en) * | 2014-05-15 | 2014-07-23 | 济南嘉科电子技术有限公司 | Big data based wireless real-time position positioning method |
| CN104134062A (en) * | 2014-08-18 | 2014-11-05 | 朱毅 | Vein recognition system based on depth neural network |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108133436A (en) * | 2017-11-23 | 2018-06-08 | 科大讯飞股份有限公司 | Automatic method and system of deciding a case |
| CN110858269A (en) * | 2018-08-09 | 2020-03-03 | 清华大学 | Criminal name prediction method and device |
| CN110858269B (en) * | 2018-08-09 | 2022-03-08 | 清华大学 | Fact description text prediction method and device |
| CN112053273A (en) * | 2020-09-16 | 2020-12-08 | 北京偶数科技有限公司 | Method and device for guiding case analysis and storage medium |
| CN112053273B (en) * | 2020-09-16 | 2021-12-03 | 北京偶数科技有限公司 | Method and device for guiding case analysis and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ferrer et al. | Bias and discrimination in AI: a cross-disciplinary perspective | |
| Coşer et al. | PREDICTIVE MODELS FOR LOAN DEFAULT RISK ASSESSMENT. | |
| CN109492945A (en) | Business risk identifies monitoring method, device, equipment and storage medium | |
| CN108648074A (en) | Loan valuation method, apparatus based on support vector machines and equipment | |
| Didimo et al. | Combining network visualization and data mining for tax risk assessment | |
| CN110276552A (en) | Risk analysis method, device, equipment and readable storage medium storing program for executing before borrowing | |
| CN107644279A (en) | The modeling method and device of evaluation model | |
| CN109376766B (en) | Portrait prediction classification method, device and equipment | |
| Ibiwoye et al. | Artificial neural network model for predicting insurance insolvency | |
| CN108492001A (en) | A method of being used for guaranteed loan network risk management | |
| CN112215696A (en) | Personal credit evaluation and interpretation method, device, equipment and storage medium based on time sequence attribution analysis | |
| CN106155327A (en) | Gesture identification method and system | |
| CN110348238A (en) | A kind of application oriented secret protection stage division and device | |
| CN112927072A (en) | Block chain-based anti-money laundering arbitration method, system and related device | |
| Kadoić et al. | Structuring e-learning multi-criteria decision making problems | |
| Alzou'bi et al. | Artificial intelligence in law enforcement, a review | |
| CN105787497A (en) | Account-stealing case analysis method and device | |
| Khodabakhshi et al. | Fraud detection in banking using knn (k-nearest neighbor) algorithm | |
| CN110825929B (en) | Service permission recommendation method and device | |
| Banakar et al. | Justification and classification of issues for the selection and implementation of advanced manufacturing technologies | |
| US20240152818A1 (en) | Methods for mitigation of algorithmic bias discrimination, proxy discrimination and disparate impact | |
| Muzammil et al. | Determinants for the Adoption of Regulatory Technology (RegTech) Services by the Companies in United Arab Emirates: An MCDM Approach. | |
| Rodríguez de las Heras Ballell | The Layers of Digital Financial Innovation: Charting a Regulatory Response | |
| Bharathy et al. | Applications of social systems modeling to political risk management | |
| Lamba | SR-MLC: Machine Learning Classifiers in Cyber Security-An Optimal Approach |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20160720 |