FR3109457A1 - Presentation attack detection method for fingerprints - Google Patents

Abstract

The invention relates to a method and a system for attack detection by presenting fingerprints on a suitable sensor in which a texture vector and a business vector are concatenated to build a model making it possible to discriminate the real fingerprints from the dummy fingerprints. Figure 1

Description

Presentation attack detection method for fingerprints

The invention relates to a method for verifying whether a fingerprint presented during a control is a real fingerprint or a dummy fingerprint. It is used in particular to detect presentation attacks (fake fingers) on fingerprint sensors.

The expression "attack detection by fingerprint presentation" in the present invention means that we will detect whether a fingerprint presented on a suitable sensor is a real fingerprint or a dummy fingerprint, and thus avoid fraudulent use of the identity of a person.

The fingerprint is one of the most widely used biometric methods to secure access and the issuance of sovereign titles. This massive use of fingerprints has led to the appearance and multiplication of attacks on biometric systems. For example, an individual previously expelled from a country can re-enter the territory by replacing the fingerprints of his right hand with those of his left hand, at an access control gate or by using a finger dummy.

We speak of a "level 1" attack, by presentation on the fingerprint sensor, where the imposters deposit fake fingers (silicone, latex, wood glue) or dead fingers, on the fingerprint sensors in order to impersonate or change identity. These attacks take place at the level of the biometric sensor and the impostor will usurp the biometric data of another individual, or create new biometric data, in order to access confidential information over which he has no rights.

It is therefore very important to set up a presentation attack detection system, also called "anti-spoofing" in the prior art. The purpose of this system will be to generate an alert in the presence of a false finger, in order to avoid the issuance or use of sovereign titles to impostors.

Several solutions are proposed in the state of the art for presentation attack detection. These solutions are of two types: the hardware approach and the software approach.

The hardware approach requires the integration of specific electronic components for fingerprint sensors. It is therefore very expensive, dependent on the types of false fingers when learning the system and not very industrializable.

The software approach is the most explored approach in the prior art. It comes in a dynamic approach and a static approach.

The dynamic software approach consists of capturing multiple images of the fingerprint over the duration of a finger movement on the sensor, a rotation and a long press of the fingerprint lasting from zero to five seconds. These methods analyze the variations over several successive images. They have the disadvantage of being less precise and above all they require more time during the acquisition of the imprint, which can challenge an impostor.

The static software approach consists of using a single image of the fingerprint to determine whether it is a real finger or a dummy finger. This is the most common state-of-the-art approach. Only one image is needed and the acquisition time is thus reduced.

The solutions known in the prior art consider a fingerprint image as any image on which we will apply methods for extracting the classic image texture descriptors before making a decision thanks to a "classifier or classifier" previously trained. The texture descriptors measure the local variations of intensity on each of the pixels of the image. The measurement of these variations, globally, gives the texture of the analyzed image. The texture descriptors are calculated for each pixel of the image and correspond to a redefinition of a pixel with respect to its local neighborhood. One of the best-known descriptors is the "local binary pattern" known by the abbreviation LBP, anglo-saxon acronym for "Local Binary Pattern". The descriptors are then inserted into a support vector machine type classifier better known by the Anglo-Saxon abbreviation SVM (Support Vector Machine) or neural networks (NNET) which are machine learning models (or Machine Learning) to learn the discriminating factors on the descriptors. These models use notions of probability calculations to find the set of descriptors allowing the best possible separation between real fingerprints and dummy fingerprints. Once the model is learned on the basis of a chosen descriptor, when a new image arrives at the input of the system, the same type of descriptor is extracted, then it is submitted to the model in order to make a classification decision. 'picture.

Other descriptors called “deep” or “deep” descriptors are also learned by deep learning, usual in image classification. It should be noted that the “deep” descriptors are more precise and provide better results. However, they take a very long time to set up because the learned model is complex.

In summary, the methods of the prior art have one or more of the following disadvantages:

They are expensive and not easy to use.

The fingerprint acquisition time to determine an attack is too long for usual control applications.

The different types of descriptors generally have at least one of the following disadvantages:
- They do not make it possible to obtain the desired precision,
- They are complex to implement and provide results that are difficult for humans to understand.

By way of illustration, the document by Xiaofei et al, entitled “Multi-scale local binary pattern with filters for spoof fingerprint detection”, Information Sciences 268 (2014) 91–102, and the document by Kumar Abhishek et al, entitled “ A Minutiae Count Based Method for Fake Fingerprint Detection”, Procedia Computer Science 58 (2015) 447 – 452, disclose fingerprint detection methods.

The idea of the present invention is to propose a new presentation attack detection method which will exploit business descriptors, resulting from the knowledge of fingerprints, combined with classic texture descriptors.

In the remainder of the description, the expression “business descriptors” designates descriptors which reflect the characteristics of a minutia, which encompasses the global descriptor and the local descriptor of a fingerprint, and under the expression “descriptors texture or LBP” the descriptors associated with the texture of the image of the imprint. The minutiae are specific points of the imprint which materialize a particular deformation of a ridge and valley.

The idea of the method according to the invention is based on the exploitation of business descriptors based on statistical estimators of characteristic elements of a fingerprint as well as on the quality of these indices. The method will use minutia extractors for fingerprints which provide several exploitable information and which will help in the construction of discriminating descriptors to discriminate between real fingerprints and dummy fingerprints.

The object of the invention relates to a method for detecting an attack by presentation of fingerprints comprising at least the following steps:
- Generate a presentation attack detection model by performing the following steps:
- Acquire one or more fingerprints labeled real or dummy by means of a sensor,
- Submitting an image of said fingerprint(s) to a texture extractor in order to generate a texture vector V ₁ (lbp ₁ ,…, lbp _m ),
- Submit said image to a minutia extractor and extract a number n of minutiae M _n , a minutia being characterized by at least its abscissa, x, its ordinate y, its type t, its orientation θ and its quality index q,
- Run the four statistical indicators:
The average ,
The standard deviation ,
The level of asymmetry of the values around the mean ,
Information on the flattening of the distribution of the variable w ,
for at least the following descriptors w : the type of the minutia, its orientation, its quality index, the distance separating said minutia from at least one neighboring minutia, the number of peaks separating said minutia M _j from said neighboring minutia considered, in order to calculate the components vm _k , of a business vector V ₂ =(vm ₁ ,..vm _k ),
- Determine the global quality Q _g of the imprint, and add this global quality value to the components vm _k ,
- Concatenate the texture vector V ₁ with the business vector V ₂ containing the properties of the imprint, to form a vector V _c containing the characteristics of the k variables described by all the minutiae acquired for a given imprint,
- Subjecting this vector V _c to a discrimination algorithm configured to generate an attack detection model by presenting fingerprints,
- Submitting a new fingerprint acquired by the fingerprint sensor to said fingerprint presentation attack detection model in order to verify whether said fingerprint is real or fictitious.

The method may further comprise the following steps:
- Calculate the values of the frequencies (Q ₁ ), f(Q ₂ ), f(Q ₃ ), f(Q ₄ ), f(Q ₅ ) of appearance for global quality values varying from 1 to 5 and adding said values to the business descriptor vector V ₂ .

According to one embodiment, the method further comprises a step of calculating the global variation of the directions of minutiae of the imprint and the addition of the global variation value of the directions to the business descriptor vector V ₂ .

The method can further comprise a step of determining the reading error frequency on the directions of the minutiae f( _erlect ) and the addition of this value to the business descriptor vector V ₂ .

The method can also comprise a step of calculating the number of empty zones, N( _zv ), on the image of the fingerprint and the addition of this value to the business descriptor vector V _2.

According to a variant embodiment, the statistical descriptors will be applied to the distance characteristic of a minutia, by considering the distance taken with respect to the three minutiae considered as the closest neighbors.

The invention also relates to a fingerprint presentation attack detection system comprising a fingerprint sensor connected to a texture extractor configured to generate a texture vector V ₁ , a concatenation device, a discrimination algorithm configured to generate an attack detection model, and a comparison device, characterized in that it further comprises the following elements:
- a minutiae extraction module configured to run the four statistical indicators:
The average ,
The standard deviation ,
The level of asymmetry of the values around the mean ,
Information on the flattening of the distribution of the variable w ,
on at least the following descriptors w : the type of the minutia, its orientation, its quality index, the distance separating said minutia from at least one neighboring minutia, the number of peaks separating said minutia M _j from said neighboring minutia considered, in order to calculate the components vm _k , of a business vector V ₂ =(vm ₁ ,..vm _k ), and
- To determine the global quality Q _g of the fingerprint, add this global quality value to the components vm _k in order to generate a business vector V ₂ containing the properties of the fingerprint,
- A module configured to concatenate the texture vector V ₁ with the vector V ₂ , to form a vector V _c containing the characteristics of the k variables described by all the minutiae acquired for a given fingerprint,
- Said discrimination algorithm being configured to generate an attack detection model by presenting fingerprints,
- Said comparator being configured to compare a fingerprint to be verified with the attack detection model and decide whether the fingerprint is a real fingerprint or a dummy fingerprint.

The minutiae extractor module is, for example, configured to determine at least one of the following values:
- The values of the frequencies f(Q ₁ ), f(Q ₂ ), f(Q ₃ ), f(Q ₄ ), f(Q ₅ ) of appearance for global quality values varying from 1 to 5 ,
- The global variation of the minutia directions of the imprint,
- The reading error frequency on the directions of the minutiae f( _erlect ) and the number of empty areas, N( _zv ), on the image of the fingerprint,
- The number of empty areas, N( _zv ), on the fingerprint image,
- The addition of one or more of these values to the business descriptor vector V _2.

The discrimination algorithm used is for example an automatic machine learning model of the SVM type or of the neural network type.

Other characteristics, details and advantages of the invention will become apparent on reading the description given with reference to the appended drawings given by way of non-limiting example and which represent, respectively:

illustrates an example of architecture allowing the implementation of the method according to the invention,

a representation table of a fingerprint by all the minutiae extracted with the characteristic variables of its local behavior,

an illustration of "holes" between the ridges and valleys of a footprint,

an example of a file containing the quality indices by zone of the image of a fingerprint,

a sequence of the steps of the method according to the invention using a combination of texture descriptors and business descriptive statistics descriptors in the same vector to differentiate dummy fingerprints from real fingerprints.

In order to clearly understand the method implemented by the invention, the following example is given to detect whether a fingerprint acquired by a fingerprint reader is a dummy fingerprint or a real fingerprint. The method is based in particular on the concatenation of conventional texture descriptors with business descriptors based on statistics of the biometric data. This will advantageously improve the precision of the classifier and therefore of the control of the “veracity” of a fingerprint, real fingerprint or dummy fingerprint. For the construction of the model, the process will use labeled fingerprints, that is to say fingerprints which we know if they are real or fake. The process will use for this construction, a sufficient number of fingerprints, in the sense usually used for the construction of databases. Once the G model has been built and learned by a system, it can be applied to a unique fingerprint to decide whether it is real or fake.

FIG. 1 illustrates an example of system architecture according to the invention comprising a fingerprint sensor 10 connected to a module 20 for processing data acquired by the fingerprint sensor. The processing module 20 comprises a first module 21, texture extractor, configured to determine image texture descriptors, a second module 22 configured to process the data of the acquired fingerprint, in order to define business descriptors complementary to the descriptors statistics as will be detailed below. This module 22 contains a minutiae extractor whose statistical indicators are used to produce business descriptors which will be combined by the person skilled in the art with the texture descriptors.

The fingerprint sensor will allow the taking of labeled fingerprints for the construction of the G model, during a first phase I of the process, then the capture of a fingerprint whose authenticity one wishes to verify, during a second phase. , phase II of the process.

The texture extractor 21 consists of descriptors with local binary patterns or LBP (Local Binary Pattern). The output of the texture extractor 21 can be written in the form of a texture vector LBP, V ₁ =(lbp ₁ ,…lbp _m ) with m=59 for example. The output of the second module 22 corresponding to business vector (local and global) will be written in the following form V ₂ = (vm ₁ , ….., vm _n ) with n=49, for example. The texture vector V ₁ and the business vector V ₂ will be transmitted to a concatenation module 23 in order to generate a vector V _c result of the concatenation V _c = (vm ₁ , ….,.vm _n , lbp ₁ , …, lbp _m ). The concatenated vector is subjected to a discriminating algorithm 24 in order to generate a model for verifying the veracity of a fingerprint 25, phase I of the method.

The generated G model will be used to decide if a fingerprint is a real fingerprint or a dummy fingerprint, phase II of the process.

For this, the system according to the invention comprises a comparator 26 taking as input data from a fingerprint acquired on the fingerprint sensor 20 and the data from the model 25 to detect whether the captured fingerprint is a real fingerprint or a fingerprint dummy. The result can be displayed on a screen of an enrollment station or the result of the comparison will generate an alarm signal at an access control gate in the event of identity theft.

The detection system according to the invention can be implemented in the enrollment stations available at the town hall to apply for a passport. These stations allow the capture of images of the applicant's ten fingerprints. The image of each fingerprint can thus be processed by the comparator 26. If the result of this comparator indicates that one of the fingerprints is dummy, then the town hall agent carrying out the enrollment will receive an alert in order to be able to interrupt the process. passport application.

The algorithm 24 is a classifier of the SVM (Support Vector Machine) type or of the neural network or NNET type which are automatic machine learning models (Machine Learning) for learning the discriminating factors on the descriptors. Any supervised learning technique algorithm intended to solve discrimination problems may be used. These models use notions of probability calculations to find the set of descriptors allowing the best possible separation between the false fingerprints and the real ones. These algorithms are known to those skilled in the art and will not be detailed. The method according to the invention “injects” at the input of these discriminating algorithms, the vector V _c resulting from the concatenation of the texture vector V ₁ and of the business vector V ₂ .

A footprint is like an alternating surface of a set of parallel ridges and valleys over most regions in the footprint. The deformations between the ridges and valleys form the minutiae which are the most stable representation used for comparison and identification of fingerprints. Minutiae represent local discontinuities and mark positions where a ridge ends or bifurcates. On a fingerprint, it is possible to detect between [1, 150] minutiae knowing that fourteen minutiae are generally sufficient to make a comparison.

A minutia m( x, y, t, θ , q, dst ₁ , nb_cr ₁ , dst ₂ , dst_cr ₂ , dst ₃ , bd_cr ₃ ) is characterized by its abscissa x , its ordinate y , its type t , its orientation θ , the quality index q associated with the thoroughness on the imprint. Two types of minutiae are used, ie, bifurcations and terminations. The orientation of a minutia is the angle formed by the deviation of the crest used to identify the minutia with respect to the horizontal. The variables dst ₁ , nb_cr ₁ , respectively represent the distance which separates the minutiae from its nearest neighbor and the number of peaks which separate them. The indices 2 and 3 in the notation dst _i , nb_cr _i , represent the same measurements for the second and third nearest neighboring minutiae. We will generalize by using the notations dst _i , nb_cr _i , i being the “rank” of the nearest neighboring minutia with respect to the minutia concerned.

Descriptors related to thoroughness are classic descriptors used in the field of fingerprint biometrics. They are known to those skilled in the art and will therefore not be detailed.

Thus, a fingerprint can be represented from the aforementioned variables characterizing a minutia, by considering three closest neighboring minutiae:

, with And .

Figure 2 is a local representation of a fingerprint with ten minutiae and the variables described above. In the example, four statistical indicators are used: the mean , the standard deviation , skewness S , kurtosis K . The mean represents the central tendency indicator of a distribution, the standard deviation indicates the fluctuation of the different values around the mean value. The skewness indicates the level of asymmetry of the values around the mean while the kurtosis gives information on the flattening of the distribution.

The following assumptions are made: each of the characteristics of a minutia m is considered as a variable, then all of these four statistical indicators are calculated for all the minutiae of a fingerprint. For example, taking the quality variable for a minutia for a given imprint, indicates the average level of local quality observed on all the minutiae detected on the fingerprint , Then determines the deviation of the individual quality indices of each minutia from the central tendency of the quality value. On dummy fingerprints, the value of the standard deviation is generally low, because the imprint is homogeneous and its variations are small, unlike a real imprint. Similar observations are made on the orientation variable θ which indicates the variation of the directions on the minutiae and shows the homogeneity on the false or dummy fingerprints unlike the real ones.

The method according to the invention will in particular use the following four statistical indicators, which it will apply to selected descriptors w :

, the mean for w ,

, the standard deviation for w ,

, the “skewness” for w .

, the “kurtosis” for w .

The process will consider all of the n minutiae M ₁ ,…, M _n of the captured fingerprint, then each variable of a minutia (with the exception of x and y), t, θ , q, dst ₁ , nb_cr ₁ , dst ₂ , dst_cr ₂ , dst ₃ , bd_cr ₃ . The method will calculate for each of these variables the value of the four aforementioned statistical estimators.

The calculation generates a set of values for all the n minutiae of the footprint and on each of the minutiae variables:

v _m1 = M ₁ [ ], ……, v _{m j} = M _j [ ], which form the components of a business vector V ₂ which will be concatenated with the texture vector V ₁ . The values V _m are linked to the j variables which constitute the minutiae of the imprint.

From the example representing the fingerprint with ten minutiae and eleven variables per minutia, we have 36 values v_m1, …, v_m36, forming this business vector V₂.

To generate the business vector V _2, the method uses in particular the business descriptors following:

The number of minutiae of the imprint,

The frequency of termination-type minutiae,

The frequency of appearance-type minutiae,

The average quality of the minutiae taken from the impression,

The standard deviation calculated on the distribution of minutiae of the fingerprint,

The "skewness" calculated on the distribution or on the quality parameter,

The average distance of the orientations on the minutiae,

The average distance with an nth nearest neighbour, for a minutia, or more generally with a minutia close to the minutia considered, the first neighboring minutia, the second neighboring minutia, etc.

To these values, the method will add information on the overall quality Q _g of the imprint.

Regardless of the effort made by an attacker to reproduce a fingerprint, a homogeneous character of the ridges induces failures in the normal quality known for a fingerprint. In general, the overall quality Q _g of a real impression will be higher than the overall quality obtained by a dummy impression.

The difficulty of positioning a finger evenly on the fingerprint sensor causes small empty areas to appear on the fingerprint image or “holes” in the image. A dummy fingerprint image generally has more “empty areas,” 30, shown in Figure 3, than a fingerprint image obtained with a real fingerprint.

To the previous parameters, the process can add additional business descriptors making it possible to better differentiate between a real fingerprint and a dummy fingerprint, to improve and make decision-making more reliable.

Thus, the process may use:

The frequency of quality equal to zero on the global quality (figure 4) – f(Q ₀ ),

The frequency of quality equal to one over the overall quality - f(Q ₁ ),

The frequency of quality equal to two over the overall quality - f(Q ₂ ),

The frequency of quality equal to three on the overall quality - f(Q ₃ ),

The frequency of quality equal to four on the overall quality - f(Q ₄ ),

The frequency of quality equal to five on the overall quality - f(Q ₅ ),

The overall total quality of the fingerprint Q _tg , which corresponds to the sum of the qualities obtained for all the frequencies f(Q ₀ ), f(Q ₁ ), f(Q ₂ ), f(Q ₃ ), f (Q ₄ ), f(Q ₅ ),

The overall variation of minutiae directions on the recorded fingerprint,

The frequency of reading error on minutia directions, f( _erlect ),

The number of empty areas or holes present on a fingerprint, N( _zv ),

The variation of directions of thoroughness which includes:

-∆θ gives the variation linked only to the minutiae of the imprint,

- The reading of a file (.dm) which contains the global directions of the imprint (ridge and valley directions) and the reading of this file gives a variable of the directions complementary to those read specifically on the minutiae with the variable ∆θ .

The four statistical indicators presented above will not be applied to these last descriptors.

The output of minutiae extractor 22 generates several files:

An F _qm file containing the quality values of the acquired impression image. The quality will be read in this file, an example of which is illustrated in FIG. 4. This file contains quality values varying from 0 to 5 per zone of the image, the value of 5 being given by way of illustration. Thus, the overall quality of a fingerprint image is extracted by summing all the quality values per area of the image. Then, for each of the values between 0 and 5, the number of occurrences is counted which represents the frequency of appearance for each quality value. The frequency of quality equal to 2 indicates the total number of occurrences of "2" found in the file F _qm ;

An F _lfm file which contains values “0” and “1”, is representative of a map of the empty areas of the footprint or Low Flow Map. This file contains values of 0 and 1 only where 1 indicates an area read as empty in the footprint. By counting the number of occurrences of 1, we obtain the total number of empty areas of the imprint;

An F _hcm file representative of a High Curvature Map. This file makes it possible to count other additional singular points of the imprints which are called the "delta", center and loop. Thus, the .hcm file contains 0s and 1s with the frequency of the 1s which indicates the presence of the singular points on the imprint;

An F _lcm file representative of a low contrast map or Low Contrast Map and also contains "0" and "1". The value 1 indicates areas of high contrast indicating the presence of a finger. In this file, the frequency of “1”s is counted to construct the contrast variable useful for identifying the total space of the impression;

The F _dm file contains information about the direction of the ridges and valleys of the footprint. The overall variation of the directions on the footprint, then the associated reading errors are read in the .dm file which stands for Direction Map. This file contains values from -1 to 15. The value "-1" indicates an inability to read the direction of the footprint. Thus, by counting all the occurrences of the values at "-1", we form the descriptor of the frequency of errors in the direction of minutiae. The rest of values from 0 to 15 indicate a read direction value. Thus, by reading the file F _dm , if at a step n the value of direction is equal to d and at step n+1 the value is d' different from d, then this variation corresponds to a change of direction and the method records this change of direction. Otherwise there is no change of direction. By traversing all the values of the file, one obtains the global change of direction on the footprint which forms the descriptor global variation of the directions of the footprint.

The method will record in the files F _lcm , F _dm , F _lfm , F _hcm , the numerical values supplied by the minutia extractor 22. For each reading of these files, line by line, a change in value corresponds to a variation . The frequency is obtained by counting the number of times a value appears. Direction errors correspond to values equal to "-1" in the F _dm file. The line-by-line traversal makes it possible to take into account the entire footprint, by counting the frequencies of an entire file as well as their occurrence.

Figure 5 illustrates a succession of steps implemented by the method according to the invention.

During phase I for the construction of a model from labeled fingerprints,

The first step 51 consists in acquiring several labeled fingerprints as defined previously,

During a second step 52, an algorithm will extract a first texture vector, according to a known method such as LBP for example.

During a third step 53 which is carried out in parallel with the second step 52, the second module will extract statistics descriptors from the local and global information provided by a minutia extractor in order to generate indicators allowing the construction of a business vector (local and global),

During a fourth step 54, the first texture vector and the business vector are concatenated to generate a combination result vector,

During a fifth step 55, the method executes an SVM-type algorithm for learning a decision model that will make it possible to discriminate between a real fingerprint and a dummy fingerprint. To do this, the model will use a database containing several real fingerprints and several dummy fingerprints. For each of these fingerprints, steps 51, 52, 53 and 54 are executed then at 55, all these extractions are learned automatically by the SVM which produces the separating model of a real fingerprint compared to a dummy fingerprint.

The sixth step 56 generates a model which will be used to carry out a comparison, during a seventh step 57, with acquired fingerprints in order to determine whether they are dummy fingerprints or real fingerprints.

During phase II, the process will capture a fingerprint to be verified, i.e., which we are trying to verify if it is a real or fake fingerprint. The fingerprint is submitted to the model generated by the steps explained above, using a comparison technique known to those skilled in the art.

Detailed description of step 53

Step 53 of construction of the business vector comprises at least the following steps.

The four statistical indicators, the average, the standard deviation, the "skewness", the "kurtosis" are executed, for all the minutiae extracted on each of the following variables:

The type of minutia, its orientation, its quality index, its distance separating a minutia from a neighboring minutia, the number of ridges separating a minutia M at a neighboring minutia considered, in order to generate a business vector V_m,

Determine an overall quality value of the fingerprint captured,

Define a vector containing the result of the application of the statistical indicators and the value of the overall quality of the fingerprint, for the n minutiae for each of the j variables,

In order to improve the detection method, step 53 for constructing the business vector can add to the business descriptor vector V ₂ one or more of the following values:

The global total quality of the fingerprint Q _tg , the frequency of quality equal to zero on the global quality f(Q ₀ ), or equal to one f(Q ₁ ), or equal to two f(Q ₂ ), or equal at three f(Q ₃ ), or equal to four f(Q ₄ ), or equal to five f(Q ₅ ), the overall variation of the directions of minutiae on the registered imprint, the frequency of reading error on the directions of the minutiae, f( _erlect ), the number of empty zones or holes present on a fingerprint, N( _zv ).

The descriptor vector thus formed will be concatenated with the texture descriptor vector before being transmitted to the learning algorithm to generate an attack detection model by presenting fingerprints.

As previously described, the process will generate a business vector which will be concatenated with the texture vector.

According to a first embodiment, to generate this business vector, the method considers the following parameters:

During a step of checking the "veracity" of a fingerprint, a person presents his fingerprint on the fingerprint reader, the fingerprint is analyzed according to the processing chain explained above. The result is submitted to the model in order to decide whether or not the fingerprint is true. If the fingerprint is considered a real fingerprint, then the identity verification process can continue.

The method according to the invention has in particular the following advantages:

The information from a minutia extractor and the use of descriptors based on estimators of descriptive statistics of mean, variance, skewness, kurtosis make it possible in particular to obtain more precision to validate or reject a fingerprint as being a real fingerprint or dummy fingerprint, a quick check that can be used in real-time verification systems.

Claims (9)

Method for detecting an attack by presentation of fingerprints comprising at least the following steps:
- Generate an attack detection model (25) per presentation by performing the following steps:
- Acquire (51) one or more fingerprints labeled real or dummy by means of a sensor (10),
- Submitting an image of said fingerprint to a texture extractor (21) in order to generate a texture vector V ₁ (lbp ₁ ,..lbp _m ), (52),
- Submitting said image to a minutia extractor (22) and extracting (53) a number n of minutiae M _n , a minutia being characterized by at least its abscissa, x, its ordinate y, its type t, its orientation θ and its quality index q,
- Run the four statistical indicators:
The average ,
The standard deviation ,
The level of asymmetry of the values around the mean ,
Information on the flattening of the distribution of the variable w
,
for at least the following descriptors w : the type of the minutia, its orientation, its quality index, the distance separating said minutia from at least one neighboring minutia, the number of peaks separating said minutia M _j from said neighboring minutia considered, in order to calculate the components vm _k , of a business vector V ₂ =(vm ₁ ,.., vm _k ),
- Determine the global quality Q _g of the imprint, and add this global quality value to the components vm _k ,
- Concatenate (54) the texture vector V ₁ with the business vector V ₂ containing the properties of the imprint, to form a vector V _c containing the characteristics of the k variables described by the set of minutiae acquired for a given imprint,
- Submitting (55) this vector V _c to a discrimination algorithm (24) configured to generate (56) an attack detection model M by presenting fingerprints,
- Submitting a new fingerprint acquired by the fingerprint sensor to said fingerprint presentation attack detection model in order to verify (57) whether said fingerprint is real or fake. Method according to Claim 1, characterized in that it also comprises the following steps: calculating the values of the frequencies (Q ₁ ), f(Q ₂ ), f(Q ₃ ), - f(Q ₄ ), - f( Q ₅ ) of appearance for global quality values varying from 1 to 5 and adding said values to the business descriptor vector V ₂ . Method according to one of Claims 1 or 2, characterized in that it also comprises a step of calculating the global variation of the minutiae directions of the imprint and the addition of the global variation value of the directions to the descriptor vector profession V ₂ . Method according to one of Claims 1 to 3, characterized in that it also comprises a step of determining the reading error frequency on the directions of the minutiae f( _erlect ) and the addition of this value to the descriptor vector profession V ₂ . Method according to one of Claims 1 to 4, characterized in that it also comprises a step of calculating the number of empty zones, N( _zv ), on the image of the imprint and the addition of this value to the business descriptor vector V ₂ . Method according to one of the preceding claims, characterized in that the statistical descriptors are applied to the three distances of a minutia corresponding to its three closest neighbours. Fingerprint presentation attack detection system comprising a fingerprint sensor (10) connected to a texture extractor (21) configured to generate a texture vector V ₁ , a concatenation device (23), a discrimination algorithm (24) configured to generate an attack detection model M, and a comparison device (26), characterized in that it further comprises the following elements:
- a minutiae extraction module (22) configured to execute the four statistical indicators:
The average ,
The standard deviation ,
The level of asymmetry of the values around the mean ,
Information on the flattening of the distribution of the variable w
,
on at least the following descriptors w : the type of minutia, its orientation, its quality index, the distance separating said minutia from at least one neighboring minutia, the number of peaks separating said minutia M _j from said neighboring minutia considered, in order to calculate the components vm _k , of a business vector V ₂ =(vm ₁ ,..vm _k ), and to determine the overall quality Q _g of the imprint, and add this overall quality value to the components vm _k ,
- Determine the global quality Q _g of the fingerprint and add this global quality value to the components vm _k in order to generate a business vector V ₂ containing the properties of the fingerprint,
- A module configured to concatenate (54) the texture vector V ₁ with the business vector V ₂ , to form a vector V _c containing the characteristics of the k variables described by all the minutiae acquired for a given fingerprint,
- Said discrimination algorithm (24) being configured to generate (56) an attack detection model M by presenting fingerprints,
- Said comparator (26) being configured to compare a fingerprint to be verified with the attack detection model and decide whether the fingerprint is a real fingerprint or a dummy fingerprint. Detection system according to Claim 7, characterized in that the minutiae extractor module (22) is configured to determine at least one of the following values:
- The values of the frequencies f(Q ₁ ), f(Q ₂ ), f(Q ₃ ), - f(Q ₄ ), - f(Q ₅ ) of appearance for global quality values varying from 1 to at 5,
- The global variation of the minutia directions of the imprint,
- The reading error frequency on the directions of the minutiae f( _erlect ) and the number of empty areas, N( _zv ), on the image of the fingerprint,
- The number of empty areas, N( _zv ), on the fingerprint image,
- The addition of one or more of these values to the business descriptor vector V _2. Detection system according to one of Claims 7 or 8, characterized in that the discrimination algorithm (24) is an automatic machine learning model of the SVM type or neural networks.