A method and device for generating a verification code based on handwriting signature

Abstract

The application discloses a method for generating a verification code based on handwriting signatures. The method relates to computer information processing technology, and comprises the following steps: obtaining a real handwriting signature of a person to be verified; adding disturbance to the real handwriting signature of the user to generate multiple signatures similar to the real handwriting signature, and forming candidate images together with the real signature; generating an adversarial image by combining the real handwriting signature image and the multiple disturbed signature images; encapsulating the adversarial image to generate a verification code; displaying the verification code as a candidate answer on a terminal screen; and receiving data returned by a terminal selection operation to verify whether the operation is a real user operation or a robot attack. The method for generating a verification code based on handwriting signatures can prevent the verification code from being attacked and improve the overall security of the verification code. The method can be widely applied to system login verification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer information processing, in particular to a handwriting signature generation verification code technology. BACKGROUND

[0002] As a visual and convenient technical means to distinguish between real users and malicious robots, the verification code has been widely used in the fields of website security, data security, operation security, transaction security and other Internet fields. The common verification codes at present mainly include sliding type, point selection type behavior verification code, character recognition click verification code, image target detection type verification code, semantic understanding type verification code and intelligence test verification code.

[0003] The forms of graphic verification codes are various, and the common ones at present include character recognition (Chinese and English) type, sliding block puzzle type, word selection type, scene recognition type and the like, but the character type verification code is still one of the most widely used verification code forms. From the initial simple character deformation to gradually increasing the distortion degree, adding various interference lines and noise methods, character superposition and adhesion and the like, the difficulty of machine recognition is increased. Some initial machine learning methods (SVM, OCR optical character recognition and the like) cannot effectively recognize in the present highly complex verification code scene. With the improvement of OCR technology and artificial intelligence technology in recent years, the ordinary verification code can be easily recognized by automatic machine, and both the character verification code and the puzzle verification code are easy to be cracked by the OCR technology, resulting in the reduction of the security of the verification code. If the picture verification code is designed too complex, the difficulty of human eye recognition is increased, and the user friendliness is not good. Therefore, how to balance the convenience of user use and reduce the probability of cracking the picture verification code by the verification code recognition software.

[0004] A method and device for generating picture verification code are disclosed in CN104065666B, entitled "A method and device for generating picture verification code". A predetermined number of picture verification codes with a certain interference intensity level are generated according to the input interference intensity level. The verification code recognition accuracy is obtained by verifying the string and the corresponding original verification string, which is easy to be cracked by OCR technology. CN200710301626A, entitled "Picture verification code generation method and picture verification code generation system", discloses a method for calculating the color distribution and brightness distribution of a natural background image, determining the text color and brightness for foreground text according to the calculated color distribution and brightness distribution, and then embedding the background text into the natural background image according to the determined text color and text brightness. CN110009057B, entitled "A graphical verification code recognition method based on deep learning", generates automatic annotations for verification codes through a generative adversarial network, and then obtains a data set for training a verification code recognition network. The trained image verification code recognition network is verified and tested through real verification code data sets and simulated verification code data sets, and the graphical verification code recognition is realized. The verification code needs to be annotated and the verification code recognition network needs to be verified and tested. It mainly aims at the recognition of verification code rather than the generation of verification code, and a large amount of data set for training needs to be obtained in advance. It has certain difficulty and complexity to realize. With the development of image recognition, character recognition, text understanding and natural language processing technology, some verification code methods have the risk of being broken and have security risks. Some verification codes are related to reasoning and logic, which reduces the user experience. SUMMARY

[0005] The present application aims at the problems of easy cracking and attack of verification code and poor user experience in the prior art, and proposes a method and device for generating verification code based on handwriting signature. The user handwriting signature is integrated into the verification code to solve the problem that the verification code in the prior art is easy to be attacked by technical means. The personal handwriting signature has characteristics. Once the writing habit is formed, the external presentation form of the handwriting is fixed. Therefore, the user handwriting signature is integrated into the verification code for verification, which can effectively prevent malicious attacks of robots.

[0006] The technical scheme for solving the above technical problems of the present application is to provide a system for generating a verification code based on handwriting signatures, comprising a handwriting signature disturbance subsystem, an attack defense subsystem, and a verification code judgment subsystem. The handwriting signature disturbance subsystem obtains a real handwriting signature of a person to be verified, generates a reconstructed handwriting signature information using a generative adversarial network to obtain a disturbed handwriting signature, and / or adjusts a sub-speed parameter at any time of a real handwriting signature stroke sequence to reversely restore a disturbed handwriting signature similar to the real handwriting signature. The disturbed handwriting signature and the real signature form a candidate picture. The attack defense subsystem obtains the candidate picture to generate an adversarial picture, and encapsulates the adversarial picture to generate a verification code. The verification code judgment subsystem outputs the verification code to a terminal screen for display, receives a trigger signal returned by a terminal selection operation, judges that the user operation passes the verification when the selection operation is a selection of the real handwriting signature trigger signal, and judges that it is a robot attack otherwise.

[0007] Further preferably, the handwriting signature disturbance subsystem generates the disturbed signature using a generative adversarial network, decomposes the real handwriting signature into a superposition of multiple log-Gaussian signals, trains the generative adversarial network by fitting, reconstructs handwriting signature information, and obtains multiple disturbed handwriting signatures with different degrees of similarity to the real handwriting signature.

[0008] Further preferably, the training process of the generative adversarial network specifically includes: sampling random noise from an arbitrary uniform distribution; inputting the generative adversarial network generator to fit the real handwriting signature data distribution and generate an imitated handwriting signature; inputting the discriminator to distinguish whether it is an imitated handwriting signature or a real handwriting signature, and updating the generator and the discriminator according to the real handwriting signature data P data , calling the formula: calculating the loss function, repeatedly updating the generator and the discriminator through the loss function, and completing the training of the generative adversarial network until Nash equilibrium is reached, wherein P z represents the noise distribution.

[0009] Further preferably, the one or more sub-speed parameters are adjusted to different degrees to change the speed value and / or angle of the sub-speed sequence of the signature stroke to obtain the disturbed handwriting signature, wherein the sub-speed parameters include: the amplitude value, occurrence time, log time delay, log response time, starting angle, and ending angle of the sub-speed of the signature stroke sequence at a certain time. According to the log-Gaussian distribution of the speed value sequence of the stroke sequence at a certain time, the amplitude value, occurrence time, log time delay, and log response time of a certain sub-speed of the stroke sequence are obtained, the signature stroke trajectory information in the horizontal and vertical directions is obtained, and the starting angle and ending angle of a certain sub-speed of the signature stroke sequence at any time are calculated.

[0010] Further preferably, the amplitude value D i , occurrence time , and log time delay μ iand the logarithmic response time sigma i , the formula is called:

[0011] The speed value |v i (t) of the i-th sub-speed sequence v i (t) of the signature stroke sequence at time t is calculated. The starting angle and the ending angle of the i-th sub-speed at time t are calculated. i (t) of the signature stroke sequence.

[0012] Further preferably, the attack defense subsystem generates the verification code according to the adversarial image package, which specifically comprises: sampling random noise from an arbitrary uniform distribution; inputting the random noise into the attack defense subsystem generator to fit the data distribution of the real handwriting signature image and the perturbed signature image, generating a corresponding imitated signature image, and inputting the real handwriting signature image, the perturbed signature image and the imitated signature image into the discriminator; the generator and the discriminator are repeatedly updated and iterated until the loss function reaches Nash equilibrium, and the corresponding adversarial encrypted image is obtained; the perturbed signature image and the corresponding adversarial encrypted image are input into the similarity discrimination network for discrimination, and the adversarial encrypted image obtained when the similarity threshold is less than the predetermined threshold is taken as the verification code.

[0013] The application also provides a method for generating a verification code based on handwriting signature, which obtains a real handwriting signature of a person to be verified, adds noise to a generative adversarial network, decomposes the real handwriting signature into superposition of multiple lognormal Gaussian signals through mutual game of the generator and the discriminator, trains the generative adversarial network through fitting, reconstructs handwriting signature information, and obtains multiple perturbed handwriting signature images with different similarity degrees from the real handwriting signature; adjusts the sub-speed parameter of the real handwriting signature stroke sequence at any moment, generates multiple perturbed handwriting signatures similar to the real handwriting signature; the perturbed handwriting signature and the real signature form a candidate image, and the candidate image generates an adversarial image package to generate a verification code; the verification code is output to a terminal screen for display, and a signal returned by a terminal triggered selection operation is received; when the triggered operation is a real handwriting signature signal, it is judged that the user operation passes the verification, otherwise it is a robot attack.

[0014] Further preferably, the training process of the generative adversarial network specifically comprises: sampling random noise from an arbitrary uniform distribution; inputting the generative adversarial network generator to fit the data distribution of the real handwriting signature, generating an imitated handwriting signature; inputting the discriminator to distinguish whether it is an imitated handwriting signature or a real handwriting signature, and according to the real handwriting signature data P data , the formula is called: Calculate the loss function, and iterate through the generator and discriminator using the loss function until Nash equilibrium is reached, completing the training of the generative adversarial network. Here, P... z This indicates the noise distribution.

[0015] Further optimization involves adjusting one or more sub-velocity parameters to varying degrees to alter the velocity values ​​and / or angles of the signature stroke sub-velocity sequence, thereby obtaining a perturbed handwriting signature. The sub-velocity parameters include: the amplitude value, occurrence time, logarithmic time delay, logarithmic response time, start angle, and end angle of the sub-velocity at a given moment in the signature stroke sequence. Based on the log-Gaussian distribution of the stroke sequence velocity values ​​at a given moment, the amplitude value, occurrence time, logarithmic time delay, and logarithmic response time of a specific sub-velocity in the stroke sequence are obtained. The signature stroke trajectory information in the horizontal and vertical directions is acquired. The start and end angles of a specific sub-velocity in the signature stroke sequence at any given moment are calculated. The amplitude value D of the i-th sub-velocity in the signature stroke sequence is obtained. i Time of occurrence Logarithmic time delay μ i and logarithmic response time σ i Call the formula:

[0016] Calculate the i-th sub-velocity sequence v of the signature stroke sequence at time t. i The velocity value of (t)|v i (t)|;Based on the starting angle of the i-th sub-velocity at time t and ending angle Call the formula: Calculate the angle φ of the i-th sub-velocity in the signature stroke sequence. i (t).

[0017] Further optimization involves generating a verification code based on an adversarial image encapsulation, specifically including: sampling random noise from an arbitrary uniform distribution; inputting the random noise into the generator of the attack defense subsystem to fit the data distribution of the real handwriting signature image and the perturbed signature image, generating a corresponding imitation signature image; inputting the real handwriting signature image, the perturbed signature image, and the imitation signature image into the discriminator; repeatedly updating and iterating the generator and the discriminator until the loss function reaches Nash equilibrium, obtaining the corresponding adversarial encryption image; inputting the perturbed signature image and the corresponding adversarial encryption image in pairs into a similarity discrimination network for discrimination; if the similarity threshold is less than a predetermined threshold, the obtained adversarial encryption image is used as the verification code.

[0018] This invention applies user handwriting signatures to CAPTCHAs. A handwriting signature perturbation subsystem generates multiple candidate answers. This involves simulating multiple handwriting signature images with varying degrees of similarity to the real signature through a game between the generator and discriminator. Based on writing kinematics theory, model parameters are fitted to reconstruct the original handwriting signature information, resulting in multiple perturbed handwriting signatures with different levels of similarity to the original. Subsequently, an attack defense subsystem encrypts both the real and perturbed handwriting signatures, effectively preventing CAPTCHA breaches even if an attacker obtains the user's real signature, thus improving overall CAPTCHA security. Furthermore, this invention avoids lengthy text readings and brain teasers, resulting in a better user experience. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a system for generating verification codes based on handwriting signatures;

[0020] Figure 2 This is a diagram illustrating handwriting signature verification. Detailed Implementation

[0021] To make the technical means, objectives and effects of the present invention easy to understand, the implementation methods and means of the present invention will be specifically described below in conjunction with the accompanying drawings and specific embodiments.

[0022] like Figure 1 The diagram shows a system schematic for generating CAPTCHAs based on handwriting signatures according to the present invention. It mainly includes three subsystems: a handwriting signature perturbation subsystem, an attack defense subsystem, and a CAPTCHA judgment subsystem. The handwriting signature perturbation subsystem obtains the user's authentic handwriting signature, adds perturbation to the signature to generate candidate images (e.g., multiple signatures similar to the authentic signature can be generated and combined with the authentic signature to form candidate images); the attack defense subsystem obtains the candidate images and generates corresponding adversarial encryption images to prevent malicious attacks; the CAPTCHA judgment subsystem stores candidate answers and judges human-machine behavior based on the user's returned answer.

[0023] The handwriting signature perturbation subsystem obtains the genuine handwriting signature and generates multiple (e.g., n) perturbation signatures similar to the genuine signature. The number of perturbation signatures generated is determined based on the required verification difficulty. To increase the verification difficulty, n is increased, thus increasing the number of candidate options and decreasing the probability of the robot selecting the correct answer. Generating multiple similar perturbation signatures through the handwriting signature perturbation subsystem primarily serves to increase the number of verification options. Besides the methods described below, other feasible methods can also be used to generate perturbation signatures.

[0024] The handwriting signature information can be decomposed into superposition of multiple log Gaussian signals by using a generative adversarial network or fitting model parameters to reconstruct the original handwriting signature information, to obtain multiple perturbed handwriting signatures with different degrees of similarity to the original handwriting signature, and to realize perturbation of the real handwriting signature.

[0025] Embodiment 1: generating a perturbed signature by using a generative adversarial network. The generative adversarial network (GAN) is a deep learning model applied to an unsupervised learning network for image generation, including a generator G and a discriminator D. Noise (a set of random vectors can be used) is input into the generator, and the generator generates imitated handwriting signatures. Then, the imitated handwriting signatures and real handwriting signatures are input into the discriminator at the same time, and the generator and the discriminator are mutually game-played, so that thousands of imitated handwriting signatures with different degrees of similarity to the real handwriting signature can be simulated. The training process of the generative adversarial network is as follows:

[0026] 1. Random noise is sampled from an arbitrary uniform distribution;

[0027] 2. The random noise is input into the generator as an input, and the generator tries to fit the distribution of the real handwriting signature data as much as possible to generate imitated handwriting signatures;

[0028] 3. The generated imitated handwriting signatures and the real handwriting signatures are input into the discriminator, and the discriminator is trained to distinguish the imitated handwriting signatures and the real handwriting signatures as much as possible;

[0029] 4. The generator and the discriminator are repeatedly updated and iterated through a loss function until a Nash equilibrium is reached, at which time the generator has been able to fit the distribution of the real handwriting signature data and generate data close to the real handwriting signature. According to the real handwriting signature data P data , the formula is called:

[0030]

[0031] The loss function is calculated, where represents the loss function of the target function (the generator G and the discriminator D), P z represents a noise distribution subject to an arbitrary distribution, which can be subject to a Gaussian distribution, x ~ P data represents that x is selected within the range of P data , and z ~ P z represents that z is selected within the range of P z . E represents mathematical expectation, G(z) represents the output of the data z after passing through the generation network, i.e. the generated fake signature picture, and D(x) represents the probability of outputting a real signature picture after passing through the discrimination network.

[0032] The generative adversarial network does not need to be trained well so that the generated imitated handwriting signature is obviously different from the real handwriting signature, and the user can distinguish them in a short time. The loss function reaches Nash equilibrium, and the obtained generator is used to generate the imitated handwriting signature.

[0033] In Embodiment 2, the sub-speed parameter of the signature stroke sequence at any moment is disturbed, the amplitude value, occurrence time, logarithmic time delay, and logarithmic response time of the sub-speed of the signature stroke sequence at the moment are obtained according to the logarithmic Gaussian distribution of the speed value sequence of the stroke sequence at the moment, the signature stroke trajectory information in the horizontal and vertical directions is obtained, the starting angle and ending angle of the sub-speed of the signature stroke sequence at any moment are calculated, one or more of the above parameters (amplitude value, occurrence time, logarithmic time delay, logarithmic response time, starting angle, and ending angle of the sub-speed of the signature stroke sequence at the moment) are adjusted to different degrees, and the disturbed handwriting signature is obtained.

[0034] Based on the writing kinematics theory, the sigma lognormal model is adopted, and the formula is called.

[0035] The speed sequence v(t) of the signature is calculated,

[0036] where v i (t) is the i th decomposed sub-speed sequence in the speed sequence, that is, v(t) is the sum of N sub-sequences v i (t). The speed value sequence |v i (t)| can be expressed by a logarithmic Gaussian distribution, the amplitude value D i , occurrence time logarithmic time delay μ i , and logarithmic response time σ i of the i th sub-speed of the signature stroke sequence are obtained, and the formula is called.

[0037] The speed value |v i (t)| of the i th sub-speed sequence v i (t) of the signature stroke sequence at time t is calculated. Wherein, exp( ) is an exponential function.

[0038] The signature stroke trajectory information x(t) and y(t) in the horizontal and vertical directions are obtained according to the formula:

[0039]

[0040]

[0041] The angle φ i(t), where ε x (t) and ε y (t) are two correction terms.

[0042] The formula is called as follows: The initial angle and the end angle of the i-th sub-velocity of the signature stroke sequence at the time t are calculated and

[0043] The amplitude value D i , the occurrence time , the logarithmic time delay μ i , and the logarithmic response time σ i , the initial angle , and the end angle of the i-th sub-velocity of the signature stroke sequence obtained by fitting are disturbed to obtain a new signature sequence. The related parameters can be disturbed in the following manner: the related parameters are disturbed according to a uniformly distributed random variable, and the following formula is called to obtain the corresponding new model parameters after disturbance

[0044]

[0045]

[0046]

[0047]

[0048]

[0049]

[0050] where R D , R μ , R σ , and are uniformly distributed random variables, the value range is determined by the visual Turing test, and after disturbance, they are taken as the updated values of the corresponding D i , μ i , σ i , , respectively, and substituted into formula (3) to obtain |v i (t)|, substituted into formula (4) to obtain Finally, substituted into formula (5) and formula (6), a new signature sequence can be obtained. That is, by adjusting D i , μ i , σi , The size of one or several of these parameters reverses to obtain a simulated handwriting signature similar to the real handwriting signature, that is, a disturbed handwriting signature with different degrees of similarity to the original handwriting signature.

[0051] In embodiment 3, a simulated signature can be generated by adding single-point random noise to the handwriting signature, accumulating single-point random noise, or directly taking the corresponding Chinese characters from the font library.

[0052] One disturbance method can be used alone to generate a simulated signature, or several disturbance methods can be used simultaneously to generate a simulated signature. A total of n signatures that meet the identification difficulty requirements can be generated.

[0053] Generally, the attacker cannot obtain the user's real handwriting signature, so the handwriting signature disturbance subsystem can better prevent attacks. However, in special cases, if the attacker has obtained the user's real handwriting signature through some means, it is possible to break the verification code. Therefore, the attack defense subsystem is added as a second encryption defense line to further effectively prevent attackers from breaking the verification code system.

[0054] The attack defense subsystem generates an adversarial image based on the real handwriting signature and the n disturbed signatures (i.e., n+1 images) generated by the handwriting signature disturbance subsystem, encapsulates the adversarial image to generate a verification code, and greatly increases the difficulty of identifying the correct answer for the attacker.

[0055] Embodiment 1:

[0056] 1. Random noise is sampled from an arbitrary uniform distribution;

[0057] 2. The random noise is input into the attack defense subsystem generator, which generates as many false images as possible to fit the data distribution of the real handwriting signature image and the n disturbed signature images;

[0058] 3. The real handwriting signature image, the n disturbed signature images, and the false n+1 images are input into the discriminator to train the discriminator to distinguish between the real handwriting signature image, the disturbed signature image, and the false image. The specific training method can be: the generator and the discriminator are repeatedly updated and iterated until the loss function reaches Nash equilibrium. At this time, the generator has been able to fit the distribution of the real handwriting signature image and the n disturbed signature images, and generate data very close to the n+1 images;

[0059] 4. After the model training is completed, the attack defense subsystem obtains n+1 adversarial encrypted images corresponding to the real handwriting signature image and the n disturbed signature images;

[0060] 5. The n perturbed signatures generated by the handwriting signature perturbation subsystem are input into the similarity discrimination network in pairs with the corresponding adversarial encrypted images generated by the handwriting signature perturbation subsystem. If the similarity threshold value obtained by the similarity discrimination network is less than a predetermined threshold value (such as 0.5), the result is taken as the final result generated by the attack defense subsystem.

[0061] The image generated by the attack defense subsystem does not affect the visual effect of the user, but the pixel values of the image have been changed, which can greatly confuse the way of cracking the verification code by a machine. Regardless of whether the attacker uses OCR technology or deep learning technology, it is difficult to find the correct answer, making the verification code difficult to break, and effectively improving the security of the verification code.

[0062] Embodiment 2:

[0063] The real handwriting signature and the n+1 images generated by the handwriting signature perturbation subsystem are respectively added with Gaussian noise, blurring, etc. to generate corresponding n+1 signature images that can achieve attack prevention.

[0064] As Figure 2 is a handwriting signature verification display schematic diagram. The candidate answers are composed of the real handwriting signature of the user and the n handwriting signatures generated by the system, and the candidate answer arrangement order is random. When the user selects the real signature image, the verification can be passed. The real handwriting signature image of the user and the n imitated handwriting signature images generated are output to the verification page display, the verification code decision subsystem completes the data interaction between the front end and the back end, makes a decision to judge the human-computer behavior. The verification code decision subsystem obtains n+1 candidate image answers generated by the attack defense subsystem, displays the n+1 candidate image answers as the verification code on the screen to the user, receives the data returned after the operation, makes a decision in combination with the real handwriting signature, and triggers the operation through the verification code to determine whether it is a real user operation or a robot attack.

[0065] The verification code decision subsystem obtains the selected signature image and compares it with the stored real handwriting signature to identify whether the selected signature image is the real handwriting signature of the user. If the selected signature image is the real handwriting signature of the user, the decision is passed, and the user is allowed to access the system.

[0066] To further increase the fault tolerance and facilitate the user to accurately select the target, the n+1 encrypted candidate images generated by the attack defense subsystem are randomly arranged in sequence and can be displayed in multiple independent different shapes, such as circles, rectangles, triangles, etc. The candidate answer display part can be set with a background image, skin, etc. If the user needs to change the options, the change prompt can be “change a question”, “change a change”, “refresh”, etc.

[0067] The above-mentioned method examples are only specific method examples of the present application, and are not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A system for generating verification codes based on handwriting signatures, characterized in that, include: The system includes a handwriting signature perturbation subsystem, an attack defense subsystem, and a verification code judgment subsystem. The handwriting signature perturbation subsystem obtains the real handwriting signature of the person to be verified, generates an adversarial network to reconstruct the handwriting signature information to obtain a perturbed handwriting signature, and / or adjusts the sub-velocity parameter of the real handwriting signature stroke sequence at any time to reversely restore a perturbed handwriting signature similar to the real handwriting signature. The perturbed handwriting signature and the real signature form a candidate image. The attack defense subsystem obtains candidate images, generates adversarial images, and encapsulates the adversarial images to generate CAPTCHAs. The verification code judgment subsystem outputs the verification code to the terminal screen and receives the trigger signal returned by the terminal's selection operation. If the selection operation is to select a real handwriting signature trigger signal, it is determined that the user operation has passed the verification; otherwise, it is considered a robot attack. By adjusting one or more sub-velocity parameters to varying degrees, the velocity values ​​and / or angles of the signature stroke sub-velocity sequence are changed to obtain a perturbed handwriting signature. The sub-velocity parameters include: the amplitude value, occurrence time, logarithmic time delay, logarithmic response time, start angle, and end angle of the sub-velocity at a certain moment in the signature stroke sequence. Based on the logarithmic Gaussian distribution of the velocity value sequence of the stroke sequence at a certain moment, the amplitude value, occurrence time, logarithmic time delay, and logarithmic response time of a certain sub-velocity of the stroke sequence are obtained. The signature stroke trajectory information in the horizontal and vertical directions is obtained, and the start angle and end angle of a certain sub-velocity of the signature stroke sequence at any moment are calculated. Obtain the amplitude value D of the i-th sub-velocity in the signature stroke sequence. i Time of occurrence Logarithmic time delay μ i and logarithmic response time σ i Call the formula: Calculate the i-th sub-velocity sequence v of the signature stroke sequence at time t. i The velocity value of (t)|v i (t)|;Based on the starting angle of the i-th sub-velocity at time t and ending angle Call the formula: Calculate the angle of the i-th sub-velocity in the signature stroke sequence.

2. The system according to claim 1, characterized in that, The handwriting signature perturbation subsystem uses a generative adversarial network (GAN) to generate perturbation signatures. It decomposes the real handwriting signature into a superposition of multiple log-Gaussian signals, trains the GAN through fitting, reconstructs the handwriting signature information, and obtains multiple perturbation handwriting signatures with different degrees of similarity to the real handwriting signature.

3. The system according to claim 2, characterized in that, The training process of a generative adversarial network (GAN) specifically includes: sampling random noise from an arbitrary uniform distribution; inputting the GAN generator to fit the distribution of real handwriting signature data to generate a simulated handwriting signature; inputting the discriminator to distinguish whether it is a simulated handwriting signature or a real handwriting signature, based on the real handwriting signature data P. data Call the formula: Calculate the loss function, and iterate through the generator and discriminator using the loss function until Nash equilibrium is reached, completing the training of the generative adversarial network. Here, P... z This indicates the noise distribution.

4. The system according to any one of claims 1-3, characterized in that, The attack defense subsystem generates a verification code based on the adversarial image encapsulation, specifically by: sampling random noise from an arbitrary uniform distribution; inputting the random noise into the generator of the attack defense subsystem to fit the data distribution of the real handwriting signature image and the perturbed signature image, generating a corresponding imitation signature image; inputting the real handwriting signature image, the perturbed signature image, and the imitation signature image into the discriminator; repeatedly updating and iterating the generator and the discriminator until the loss function reaches Nash equilibrium, obtaining the corresponding adversarial encryption image; inputting the perturbed signature image and the corresponding adversarial encryption image in pairs into a similarity discrimination network for discrimination; if the similarity threshold is less than a predetermined threshold, the obtained adversarial encryption image is used as the verification code.

5. A method for generating verification codes based on handwriting signatures, characterized in that, The process involves obtaining the real handwriting signature of the person to be verified, adding noise to a generative adversarial network (GAN), and decomposing the real handwriting signature into a superposition of multiple log-Gaussian signals through a game between the generator and the discriminator. The GAN is then trained by fitting the signal to reconstruct the handwriting signature information, resulting in multiple perturbed handwriting signature images with varying degrees of similarity to the real signature. The sub-velocity parameters of the stroke sequence of the real handwriting signature are adjusted at any given time to generate multiple perturbed handwriting signatures similar to the real signature. These perturbed handwriting signatures and the real signature are combined to form candidate images, which are then used to generate adversarial images and encapsulate them to generate a CAPTCHA. The verification code is displayed on the terminal screen. The terminal receives the signal returned by the selected operation. If the selected operation is a real handwriting signature signal, it is determined that the user operation has passed the verification. Otherwise, it is a robot attack. By adjusting one or more sub-velocity parameters to varying degrees, the velocity values ​​and / or angles of the signature stroke sub-velocity sequence are changed to obtain a perturbed handwriting signature. The sub-velocity parameters include: the amplitude value, occurrence time, logarithmic time delay, logarithmic response time, start angle, and end angle of the sub-velocity at a certain moment in the signature stroke sequence. Based on the log-Gaussian distribution of the stroke sequence velocity values ​​at a certain moment, the amplitude value, occurrence time, logarithmic time delay, and logarithmic response time of a certain sub-velocity in the stroke sequence are obtained. The signature stroke trajectory information in the horizontal and vertical directions is obtained. The start angle and end angle of a certain sub-velocity in the signature stroke sequence at any moment are calculated. The amplitude value D of the i-th sub-velocity in the signature stroke sequence is obtained. i Time of occurrence Logarithmic time delay μ i and logarithmic response time σ i Call the formula: Calculate the i-th sub-velocity sequence v of the signature stroke sequence at time t. i The velocity value of (t)|v i (t)|;Based on the starting angle of the i-th sub-velocity at time t and ending angle Call the formula: Calculate the angle of the i-th sub-velocity in the signature stroke sequence.

6. The method according to claim 5, characterized in that, The training process of a generative adversarial network (GAN) specifically includes: sampling random noise from an arbitrary uniform distribution; inputting the GAN generator to fit the distribution of real handwriting signature data to generate a simulated handwriting signature; inputting the discriminator to distinguish whether it is a simulated handwriting signature or a real handwriting signature, based on the real handwriting signature data P. data Call the formula: Calculate the loss function, and iterate through the generator and discriminator using the loss function until Nash equilibrium is reached, completing the training of the generative adversarial network. Here, P... z This indicates the noise distribution.

7. The method according to any one of claims 5-6, characterized in that, The specific steps for generating a CAPTCHA based on adversarial image encapsulation are as follows: 1) Sample random noise from an arbitrary uniform distribution; 2) Input the random noise into the generator of the attack defense subsystem to fit the data distribution of the real handwriting signature image and the perturbed signature image, generating a corresponding imitation signature image; 3) Input the real handwriting signature image, the perturbed signature image, and the imitation signature image into the discriminator; 4) Repeatedly update and iterate the generator and the discriminator until the loss function reaches Nash equilibrium, obtaining the corresponding adversarial encryption image; 5) Input the perturbed signature image and the corresponding adversarial encryption image in pairs into a similarity discrimination network for discrimination; 6) If the similarity threshold is less than a predetermined threshold, the obtained adversarial encryption image is used as the CAPTCHA.

Images