KR20190036796A - Apparatus and method of signature authentication using deep learning - Google Patents

Abstract

The present invention relates to a signature authentication apparatus using deep learning and a method thereof. The signature authentication apparatus comprises: an input unit receiving a signature; an acceleration sensor measuring acceleration of the signature authentication apparatus; and a control unit extracting feature data from the data inputted through the input unit and acceleration data measured through an acceleration sensor and performing the signature authentication by inputting the extracted feature data to a learning model, wherein the learning model is generated by learning data for signature registration.

Description

[0001] APPARATUS AND METHOD OF SIGNATURE AUTHENTICATION USING DEEP LEARNING [0002]

The present invention relates to a signature authentication apparatus and method using deep learning, and more particularly, to a signature authentication apparatus and method for authenticating a dynamic signature using deep learning.

Recently, as the spread of mobile devices such as smart phones and smart pads has increased, researches on biometric authentication in these devices have been developed.

The fingerprint is the most widely used authentication method. However, in authentication using static bio information such as fingerprints, voices, and faces, forgery can be performed using fingerprints, recorded voices, and photographs made with a 3D printer, and the data can not be changed once exposed.

As a result, studies on authentication based on dynamic behaviors that are difficult to imitate easily, such as recognition of gait recognition and gesture recognition, have been actively carried out. Among these dynamic behaviors, dynamic signature in a smartphone can be easily performed by a user, There is an advantage that it can be freely changed even when exposed, and no additional device is required.

However, it is an important issue to identify skilled forgery such as shoulder surfing in the dynamic signature, Smudge Attack in the sign of the signature, and so on.

In order to distinguish such falsification signatures, a lot of characteristic information about a dynamic action in which a signature is performed is required. In the past, studies related to signature authentication using a stylus pen have been mainly made, so that the signature shape, pen pressure, A method to utilize it as feature information has been proposed. However, there is a problem that this method can not be applied to a finger signature using a hand.

Meanwhile, the background art of the present invention is disclosed in Korean Patent Publication No. 10-2012-0047973 (2012.05.14).

An object of the present invention is to provide a signature authentication apparatus and method using deep learning that enables more precise signature classification.

A signature authentication apparatus using deep learning according to the present invention includes an input unit for receiving a signature; An acceleration sensor for measuring the acceleration of the signature authentication device; And a controller for extracting feature data from data input through the input unit and acceleration data measured through the acceleration sensor and inputting the extracted feature data to a learning model to perform signature authentication, And is characterized by being generated by learning data for signature registration.

In the present invention, the controller extracts the feature data by inputting the two-axis coordinate values of each signature point sampled according to the set period and the three-axis acceleration values at the corresponding points into the CNN (convolutional neural network).

In the present invention, the learning model is a model of an autoencoder system.

In the present invention, the control unit calculates an unillustrated figure based on a value output from the learning model and a value input to the learning model, and compares the calculated unillustrated figure with a threshold value to determine whether or not the authentication is successful. .

In the present invention, the controller may calculate a mean squared error between a value output from the learning model and a value input to the learning model using the non-inference diagram.

In the present invention, the control unit may normalize data to be input to the CNN (convolutional neural network).

In the present invention, the controller extracts a feature vector according to a predetermined reference from data input through the input unit and acceleration data measured through the acceleration sensor, compares the extracted feature vector with previously registered signature data, Signature data is extracted through the feature data extraction and learning model, or authentication failure processing is performed.

A signature authentication method using deep learning according to the present invention includes: a step in which a control unit receives a signature through an input unit; Measuring the acceleration of the device through which the signature is authenticated through the acceleration sensor while the controller receives the signature; Extracting feature data from the data input through the input unit and the acceleration data measured through the acceleration sensor; Inputting the extracted feature data to a learning model generated by learning data for sign registration, and calculating an output value; Calculating a non-derivation based on the output value and the value input to the learning model; And comparing the calculated non-guide figure with a threshold value to determine whether the authentication is successful.

In the step of extracting the characteristic data of the present invention, the control unit inputs the two-axis coordinate value of each signature point sampled according to the set period and the three-axis acceleration value at the time point to the CNN (convolutional neural network) And extracts the feature data.

The signature authentication method using deep learning according to the present invention includes the steps of: receiving, by the control unit, a signature for signature registration through the input unit; Measuring acceleration of the apparatus through the acceleration sensor while the control unit receives a signature for sign registration; Extracting a feature vector according to a preset reference in signature data for signature registration; Analyzing a difference between the inputted signatures based on the extracted feature vectors; And a step of requesting re-signature or building the learning model according to an analysis result of the step of analyzing the difference between the inputted signatures.

The signature authentication apparatus and method using deep learning according to the present invention can improve the accuracy of signature authentication by enhancing the recognition rate of signatures and lowering the recognition rate of falsification signatures using acceleration information as well as shape information of signatures It is effective.

Further, the apparatus and method for signature authentication using deep learning according to the present invention uses an effective deep learning technique for distinguishing falsification signatures and extracts feature data to be input for deep learning through machine learning, thereby improving the speed and accuracy of signature authentication It is possible to make it possible to make it possible.

1 is a block diagram illustrating a configuration of a signature authentication apparatus according to an embodiment of the present invention.
FIG. 2 is an exemplary diagram illustrating a relationship between a two-axis coordinate value and a three-axis acceleration value of a signature performed in the signature authentication apparatus according to an exemplary embodiment of the present invention.
3 and 4 are exemplary diagrams for illustrating normalization of data input through an input unit in a signature authentication apparatus according to an embodiment of the present invention.
5 is an exemplary diagram illustrating a learning model of a signature authentication apparatus according to an embodiment of the present invention.
FIG. 6 is an exemplary diagram illustrating a CNN (convolutional neural network) used in a signature authentication apparatus according to an embodiment of the present invention. Referring to FIG.
7 is a flowchart illustrating a signature registration process performed in the signature authentication method according to an embodiment of the present invention.
8 is a flowchart illustrating a signature authentication process performed in the signature authentication method according to an embodiment of the present invention.

Hereinafter, an embodiment of a signature authentication apparatus and method using deep learning according to the present invention will be described with reference to the accompanying drawings. In this process, the thicknesses of the lines and the sizes of the components shown in the drawings may be exaggerated for clarity and convenience of explanation. In addition, the terms described below are defined in consideration of the functions of the present invention, which may vary depending on the intention or custom of the user, the operator. Therefore, definitions of these terms should be made based on the contents throughout this specification.

FIG. 1 is a block diagram illustrating a configuration of a signature authentication apparatus according to an exemplary embodiment of the present invention. FIG. 2 is a diagram illustrating a signature verification apparatus according to an exemplary embodiment of the present invention, FIGS. 3 and 4 are diagrams for explaining the normalization of data input through the input unit in the signature authentication apparatus according to an embodiment of the present invention, and FIG. 5 is a view for explaining the relationship between values FIG. 6 is an exemplary diagram for explaining a CNN (convolutional neural network) used in a signature authentication apparatus according to an embodiment of the present invention; FIG. The signature authentication apparatus according to the present embodiment will be described with reference to the following.

1, a signature authentication apparatus according to an exemplary embodiment of the present invention includes a control unit 100, an input unit 110, an acceleration sensor 120, and a storage unit 130.

The input unit 110 may receive a signature from a user. For example, the input unit 110 may be configured in the form of a touch screen to receive a signature input by the user. That is, when the user inputs a signature to the input unit 110 through a finger or the like, the coordinate value of the touched position can be grasped through the input unit 110. [

The control unit 100 can perform registration and authentication of signatures using such coordinate data. 2) of each signature point (indicated by a dotted line in FIG. 2) sampled according to a setting period (for example, 32 ms) by a signature inputted through the input unit 110, ) Coordinate values can be used to authenticate and register the signature. That is, this coordinate data is data related to the shape of the signature. Using the data periodically sampled in this way, even if the time information is not used separately, the same effect as using the signature speed information can be obtained.

The acceleration sensor 120 can measure the acceleration of the signature authentication device. For example, when the signature authentication apparatus according to the present embodiment is a smart phone, it is possible to measure the acceleration during signing using the sensor built in the smart phone. The acceleration information includes three axes (x axis, y Axis, z-axis).

That is, in the present embodiment, the accuracy of the signature authentication can be improved by using the acceleration information of the signature authentication device in which the signature is performed as the feature information about the dynamic behavior.

Specifically, as shown in FIG. 2, when the user inputs a signature, if the direction is changed, a rapid change occurs in the acceleration of the signature authentication device. Depending on the signature shape, speed, signer's habit, Value is determined. Therefore, the acceleration value can be used as feature information for distinguishing the forged signature.

3 and 4, the control unit 100 processes the input data or the measured acceleration data through the input unit 110 to analyze the signature data and the input signature data It can be optimized. Specifically, the initial value may be subtracted from the coordinate value or the acceleration value to match the starting point, and data normalization may be performed to match the size of the signature.

The storage unit 130 may store authentication data, that is, data related to a signature input for signature registration.

For example, the control unit 100 learns data for signature registration to generate a learning model, and stores the signature in the form of a signature (130). ≪ / RTI >

That is, the control unit can perform signature authentication using the deep learning technique. Specifically, the control unit 100 can learn a signature for registration and construct a model of an autoencoder (AE) method.

The AE has a structure as shown in FIG. 4 and is a type of FNN (Feedforward Neural Network), which is a neural network that learns the inherent characteristics of data. Specifically, AE is a learning model that is learned to generate similar output values to the input values, and generates output values with high similarity for input values similar to the learned data, while output values for input values that do not have low similarity .

는 활성화 함수이고, L은 손실 함수(loss function)이다. 이러한 관계식에서의 AE의 학습 과정은 손실함수를 최소화하는 W와 W'를 찾는 과정이며, 유사한 데이터의 입력이 반복되면 해당 데이터의 특성에 대응하는 W와 W'를 산출할 수 있다.The learning method of the AE can be described by Equation (1), where h is the result of encoding from signature data x through AE, z is the result of decoding h, ,

Is an activation function, and L is a loss function. The learning process of AE in this relation is a process of finding W and W 'that minimizes the loss function, and if input of similar data is repeated, W and W' corresponding to the characteristic of the data can be calculated.

That is, when learning is performed by inputting a plurality of similar data, such an AE outputs data having a similarity degree between the input value and the output value with respect to the input value similar to the learned data. And outputs data having a low degree of similarity to the output value, that is, having a high degree of similarity.

Accordingly, the control unit 100 can distinguish the normal signature from the fake signature by using the AE.

However, the control unit 100 extracts the feature data from the corresponding data instead of inputting the two-axis coordinate values of the sampled signature points and the three-axis acceleration values at the time point to the AE, And may be configured to input data to the AE.

That is, as the amount of data input to the AE increases, the accuracy increases but the processing speed decreases. Therefore, the control unit 100 increases the processing speed by extracting the characteristic data and inputting the data to the AE, However, rather than extracting arbitrary data or lowering the sampling rate to reduce the amount of data, it is possible to increase the accuracy of signature authentication by extracting the feature data.

For example, the control unit 100 may extract feature data from signature data using a CNN (convolutional neural network) that is mainly used in the field of image recognition.

Specifically, as shown in FIG. 6, by repeating the convolutional layer and the pooling layer at least once, characteristic data can be output from the input data and the amount of data can be reduced have.

However, since the CNN (convolutional neural network) used in the present invention is one-dimensional data arranged in a time-series manner, unlike that used in general image analysis, convolution for a 1D signal can be used as shown in Equation 2 below.

That is, the calculation result F of the convolution layer can be derived using Equation 2, where f is the input vector (input data), g is the kernel (filter), u and n are the input vector The length of the kernel. Here, the filter can be arbitrarily generated at the beginning of the learning.

The output value passed through the convolution layer is made a characteristic map through an activation function having nonlinearity, which is defined as Equation (3) below. The ReLU (Rectified Linear Unit) in Equation (3) means an activation function.

If a feature has been extracted to some extent through the convolution layer, it is not necessary to make a judgment with all of the features. Therefore, it is possible to reduce the length (amount) of data and use a pooling technique.

For example, the Max-Pulling technique as shown in Equation (4) below can be used, and the Max-Pulling can be performed by setting the stride to 2 in the characteristic map F. [

The extracted data can be arranged as a series of data that can be input to the neural network in the Fully Connected layer, and the control unit 100 can construct the learning model by inputting the feature data of the signature data thus extracted to the AE .

Meanwhile, in the sign authentication process, the control unit 100 inputs feature data extracted through the CNN in the signature input in the authentication process, calculates the output value, and outputs the value output from the learning model and the learning model By calculating the non-derivation based on the input value, a fake signature can be determined.

For example, the controller 100 may calculate a mean squared error between a value output from the learning model and a value input to the learning model as the non-inference. If the non-inference is smaller than the threshold It can be determined that the signature authentication is successful.

This learning model used by the control unit 100 is not a two-class method for dividing input data into a subject and a non-object, and is a one-class method. Therefore, Can be more effective. That is, in the two-class method, the fake signature is classified closer to the subject class, so the one class method is more suitable for authentication for security.

However, since the learning model used in the present invention is not limited to the model of the AE system, data having high similarity between the input value and the output value is output for an input value similar to the learned data. However, And a method of outputting data with a low degree of similarity between output values can be employed.

In addition, the control unit 100 may further perform signature authentication by inputting the distance between each signature point to the CNN or the learning model. The more the feature information is used, the more accurate the authentication can be.

Meanwhile, the signature authentication device may be a mobile device such as a mobile phone, a smart phone, a smart pad PMP (Portable Media Player), a PDA (Personal Digital Assistant), a portable game machine, a digital camera, an e-book reader, The present invention is not limited to this, and various devices can be employed as signature authentication devices.

FIG. 7 is a flowchart illustrating a signature registration process performed in the signature authentication method according to an embodiment of the present invention. FIG. 8 illustrates a signature authentication process performed in the signature authentication method according to an embodiment of the present invention. A signature authentication method according to the present embodiment will be described with reference to FIG.

As shown in FIG. 7, the control unit 100 may perform the signature registration process first.

Specifically, the control unit 100 receives a signature for signature registration through the input unit 110 (S200). At this time, the controller 100 can measure the acceleration of the signature authentication device through the acceleration sensor 120, and can use the input or measured data as signature registration data. For example, the control unit 100 determines whether the three-axis acceleration value sampled according to the setting period, which is the data indicating the two-axis coordinate value of each signature point sampled according to the setting period, Value can be used as signature data, and the distance between each signature point can be further utilized.

Then, the control unit 100 extracts the feature vector from the signature data for signature registration according to a preset reference (S210). That is, the control unit 100 sets the feature vector according to the set reference (for example, the inflection point of the acceleration, the maximum point, the minimum point, the maximum or minimum value of the x coordinate, the maximum or minimum value of the y coordinate, etc.) Can be extracted.

In addition, the control unit 100 calculates a median value of the signature data, and analyzes the deviation of the feature vector on the basis of the calculated median value (S220). For example, the control unit 100 may calculate a deviation (e.g., standard deviation) of a point at which the value of the x-coordinate becomes the maximum based on the calculated median value. At this time, the median value can be calculated for each item (x coordinate, y coordinate, x axis acceleration, y axis acceleration, z axis acceleration) of the signature data, and when the feature relating to the x coordinate is used, It can be used as a standard. Also, in the present embodiment, it has been described that the deviation is analyzed on the basis of the median, but other representative values other than the median can be used to analyze the deviation.

Then, the control unit 100 determines whether the difference between the input signatures is equal to or greater than a reference value (S230). That is, in the signature registration process according to the present embodiment, at least two times of signatures can be inputted. If the difference between the deviations calculated for each signature thus inputted is larger than the reference value, sign registration (learning) The server 100 requests a re-signature (S240).

On the other hand, if the input signature deviation is not large, the controller 100 optimizes the inputted signature data (S250), inputs the optimized signature data to the CNN, and extracts the feature data (S260).

Thereafter, the control unit 100 constructs a learning model by learning the extracted data (S270). That is, the controller learns signatures so as to perform signature authentication using a deep learning technique, and can construct a model of an autoencoder (AE) method, for example.

Meanwhile, as shown in FIG. 8, in order to authenticate a signature, the control unit 100 first receives a signature through the input unit 110 (S300). Like the above-described step S200, the controller 100 can measure the acceleration of the signature authentication device through the acceleration sensor 120 and utilize it as signature data.

Then, the control unit 100 extracts a feature vector from the input signature data (S310), and compares the difference between the extracted feature vector and the previously registered signature data (S320). That is, the control unit 100 performs the same pre-processing as that used in the signature registration process described above, and can filter a signature having a large difference from the registered signature before performing the signature authentication through the CNN-AE.

Then, the control unit 100 optimizes the inputted signature data (S330), inputs the optimized signature data to the CNN to extract the feature data (S340), inputs the extracted feature data to the learning model, calculates the output value S350), the learning model is calculated based on the input data and the output value in the learning model (S360). That is, as described above, the AE outputs data having similarity between the input value and the output value with respect to the input value similar to the learned data. However, with respect to the input value other than the input value, the similarity between the input value and the output value is low The control unit 100 can determine the fake signature by calculating the non-derivation based on the value output from the learning model and the value input to the learning model.

If the calculated non-default value is less than the threshold value, the control unit 100 determines that the authentication is successful. Otherwise, if the calculated non-default value is greater than or equal to the threshold value, the control unit 100 determines that the authentication fails (S370 to S390). For example, the controller 100 may calculate a mean squared error between a value output from the learning model and a value input to the learning model as a non-dioptric value. In this case, if the non- The control unit 100 can determine that the signature is correct, and in the opposite case, the control unit 100 determines that the signature is a forged signature or the like, and can perform the authentication failure process.

As described above, the signature authentication apparatus and method using deep learning according to the embodiment of the present invention can increase the recognition rate of signatures and lower the recognition rate of falsification signatures by utilizing acceleration information as well as shape information of signatures, Learning method, and extracts feature data to be input to the deep learning through machine learning, thereby improving the speed and accuracy of signature authentication.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it will be understood by those of ordinary skill in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. I will understand. Accordingly, the technical scope of the present invention should be defined by the following claims.

100:
110: input unit
120: Accelerometer
130:

Claims (10)

An input unit for inputting a signature;
An acceleration sensor for measuring the acceleration of the signature authentication device; And
A controller for extracting feature data from data input through the input unit and acceleration data measured through the acceleration sensor and inputting the extracted feature data into a learning model to perform signature authentication,
Wherein the learning model is generated by learning data for signature registration.
The method according to claim 1,
Wherein the controller extracts the feature data by inputting the two-axis coordinate values of each signature point sampled according to the set period and the three-axis acceleration value at the corresponding point into the CNN (convolutional neural network) Signature authentication device.
The method according to claim 1,
Wherein the learning model is a model of an autoencoder system.
The method of claim 3,
Wherein the control unit calculates an unillustrated figure based on a value output from the learning model and a value input to the learning model and compares the calculated unillustrated figure with a threshold value to determine whether or not the authentication is successful. Signature authentication device using learning.
5. The method of claim 4,
Wherein the control unit calculates a mean squared error between a value output from the learning model and a value input to the learning model using the non-inference diagram.
3. The method of claim 2,
Wherein the control unit normalizes data to be input to the CNN (convolutional neural network).
The method according to claim 1,
The control unit extracts a feature vector according to a predetermined reference from data input through the input unit and acceleration data measured through the acceleration sensor, compares the extracted feature vector with previously registered signature data, And a sign authentication through a learning model or an authentication failure process.
The control unit receiving the signature through the input unit;
Measuring the acceleration of the device through which the signature is authenticated through the acceleration sensor while the controller receives the signature;
Extracting feature data from the data input through the input unit and the acceleration data measured through the acceleration sensor;
Inputting the extracted feature data to a learning model generated by learning data for sign registration, and calculating an output value;
Calculating a non-derivation based on the output value and the value input to the learning model; And
And comparing the calculated non-default value with a threshold value to determine whether authentication is successful.
9. The method of claim 8,
In the step of extracting the feature data, the control unit inputs the two-axis coordinate values of the sampled signature points and the three-axis acceleration values at the corresponding points in the CNN (convolutional neural network) according to the set period, And extracting the signature from the signature.
9. The method of claim 8,
Before receiving the signature,
Receiving, by the control unit, a signature for signature registration through the input unit;
Measuring acceleration of the apparatus through the acceleration sensor while the control unit receives a signature for sign registration;
Extracting a feature vector according to a preset reference in signature data for signature registration;
Analyzing a difference between the inputted signatures based on the extracted feature vectors; And
Further comprising the step of requesting re-signature or building the learning model according to an analysis result of the step of analyzing the difference between the inputted signatures.

Images