RU2663475C1 - Method of identifying the pc user “man or internet robot” - Google Patents

Abstract

FIELD: information technology.SUBSTANCE: invention relates to the security of computer networks, namely to the formation of images when a user passes a fully automated Turing test. Method for identifying a PC user "human or Internet robot", in which each image is formed on the basis of a group of 3-dimensional models, wherein the lighting parameters and the shooting angle are set by a pseudo-random process, each such 3-dimensional model can be synthesized from 3-dimensional elementary primitives by means of applying operations of union, intersection, addition, while the image shows more than three objects and necessarily some have a complex shape, and there necessarily are cases of obscuring of some 3-dimensional models by the other ones, in addition, the objects cast shadows onto each other, and the unique image is automatically created for each test assignment.EFFECT: increased likelihood of distinguishing a person from an Internet robot when accessing Internet resources.5 cl, 5 dwg

Description

Technical field

The present invention generally relates to the security of computer networks, including unauthorized access to Internet resources. The invention provides a way for owners and administrators of network services to find out who the user of the computer from which an attempt is made to access a network resource is: a person or a computer program (Internet robot). So, in some cases, it is important to accurately distinguish whether the client and the initiator of specific requests to the server are a person or an Internet robot. Such Internet robots are usually used by cybercriminals whose purpose is to carry out illegal actions to obtain or disseminate information.

All large and medium-sized Internet services and Web portals, mail servers, search engines, as well as banking systems are very attractive platforms for attackers. In most cases, at present, the actions of hackers are not aimed at gaining administrator rights to manage a web resource (protection against such direct hacks is quite reliable and comes down to simple accuracy in administering the portal). The most massive way to organize online fraud is to simulate user actions using a special program (Internet robot). For example: an Internet robot can create thousands of new accounts (i.e. register supposedly new users of the resource, for example, a free email service), and then mass-send advertisements, fake invoices for payment for services, leave spam comments on forums, add friends on the social network "Vkontakte", publish messages provoking the disclosure of confidential information or personal data. If there were no protection from programs (Internet robots) that imitate human actions, then, without exception, all social Internet resources, including the websites of the State Service system, would turn into information garbage and simply could not be used.

The test that a person accesses the site has a common name - captcha. CAPTCHA (from the abbreviation SARTSNA-English. Completely Automated Public Turing test to tell Computers and Humans Apart - a fully automated Turing public test to distinguish between computers and people) - a computer test used to determine who the system user is: a person or a computer . The term appeared in 2000. The main idea of captcha: to offer the user a task that is easily solved by a person, but is extremely difficult and time-consuming for a computer (ideally unsolvable). Since in this case the test is carried out not by man, but by a machine, such a test is often called the Turing inverse test. As of 2013, approximately 320 million captchas are injected every day by users across the globe (see Wikipedia article https://en.wikipedia.org/wiki/CAPTCHA). Some requirements for Internet resources that use the captcha test are presented by GOST R 52872-2012. National standard of the Russian Federation. Internet resources. Accessibility requirements for the visually impaired

State of the art

The main measure to counter Internet robots is an additional user check, which is carried out in the form of an offer to the user to perform a small test task. It is assumed that if a person tries to access the resource, then the test will be easily and successfully passed by him, and for an Internet robot, even a specially programmed one, the probability of completing the verification task will be small. For example, on the most popular site of the Russian Internet sector (Vkontakte social network), such a check is organized as the following test: a person is shown a very distorted set of four characters, over which a request is displayed - read the “code word”, and then enter the code set of characters with using the keyboard in a special window. To make it difficult for Internet robots to recognize, the code symbols in the picture are specially crossed out, tilted, stretched, have a discontinuous mark, etc. The application of increasingly complex “noise” and frequent modification of distortions has to be done due to the appearance of increasingly sophisticated character recognition programs. So, in the description of the invention to the patent RU 2 608 262 "Distortion of the SYMBOLS FOR the SYMBOL-GRAPHIC REVERSE TURING TEST" it is proposed to obtain new types of graphically distorted characters by morphing (obtaining intermediate styles) based on two or more well-known standard fonts.

The software code of the Internet robot is known, which performs character recognition for any modern character captcha with a probability of 85% (see http://www.opennet.ru/opennews/art.shtml?num=47477). Site owners are forced to modify, complicate the test, make even more distortion and noise in the character design. However, highly complicated text captchas are easily recognized by modern recognition methods based on neural networks (see https: .// habrahabr.ru/post/343222/).

From the analysis of successful methods of “hacking” text captcha, we can draw conclusions about the main disadvantages of the test based on a symbolic code word:

a) there is a limited set of characters that can be demonstrated to a person, for example, for the Russian language it is 33 letters and 10 digits, the limitations of such a set of images for automatic recognition greatly facilitate the task of passing an Internet robot test, made, for example, on the basis of a neural network;

b) too distorted characters are already difficult to recognize even by humans, and the threshold of permissible noise and distortion for this technology has already been reached.

In addition to captcha based on distorted characters on the Internet, other methods of user verification are common. So, quite often captcha appears only in case of unusual user behavior (i.e., if it is suspected that an attack is being carried out by an Internet robot). For example, if a user performs a series of suspiciously fast actions of the same type, Google search engine protection first offers to pass a simple test (captcha) (example at: https://patrickhlauke.github.io/recaptcha/); if the user's actions are even more threatening, then a more complicated task (for the robot) appears - you need to specify images with the given content among several. For example, it is proposed to indicate with a mouse those (three to four) pictures of 16 in which, for example, a traffic sign is present. Google, apparently, uses this method of checking the user (for “humanity”) due to the fact that the protection by the “code word” had to be often modified, sophisticated in distortion and noise. Google showcases selected 16 images from a large database where all images are viewed and manually classified.

The problem with this method is that attackers can quite automatically collect a large part of the images stored in the Google database when they automatically go to the Google website, then classify these images themselves, and then “train” the neural network to recognize them. To prevent this, you need to create a lot of test images, as well as at least hundreds of categories of test images. But creating a large database of classified images is a very time-consuming and expensive task even for Google. In addition, since Google’s captcha image database is an image from the Internet, so it’s not difficult for cybercriminals to train modern student image recognition systems on the Internet. And, in addition, it turns out that free neural network training libraries from Google itself are quite suitable for use as cracking software. Those. in fact, recognition algorithms from Google (https://arxiv.org/pdf/1312.6082v4.pdf) crack Google captcha pretty well.

The publications suggest several ways to overcome the above disadvantages for the most common types of captcha (recall, distorted characters are shown in one type of captcha, in the second form you need to specify images (of several) with the given content). One of the most promising ways to overcome the above disadvantages is to create captcha images based on virtual computer 3D models of objects. Using one 3D model makes it possible to synthesize thousands of its 2D images. By changing the shooting angle of the model (i.e., turning in three-dimensional space), changing the texture and coloring, setting different lighting conditions, finally just stretching and compressing the 3D model or setting parametric skeleton-modified 3D models, you can achieve an extraordinary variety of test images for captcha . For example, in patent US 8510814 “Method and apparatus for network authentication of human interaction and user identity” it is proposed for each selected 3D model of an animal, person, piece of furniture, etc. to randomly change one or more visualization parameters, so that each the corresponding object remains a recognizable person. For example, in one of the ways to distinguish between a person and an Internet robot, the user is asked to move the mouse cursor and click on a cat, while other animals are also present on the image, or point to a sitting cat. Those. the user must recognize in the image various classes of objects and / or their state.

The prototype of the proposed method is the method described in patent U 7725395 "System and method for devising a human interactive proof that determines whether a remote client is a human or a computer program". The patent proposes a way to distinguish a person from an Internet robot, based on the ability of a person to recognize various parts of the body of a person or animal. For example, an image of a human face is shown on a computer screen to a user and the task is given: using the manipulator, move the mouse cursor and click on the earlobe. The image is formed on the basis of a 3D model of the face. And such a formation occurs with randomly selected (in a certain range) parameters: the position of the face in the frame, its size, stretching along the axes, turns (full face, half face, in profile), the face can change expression (smile, speak, etc.), except Moreover, the location and characteristics of the lighting sources of the model are also set randomly. If the database contains quite a lot of models of different faces (for example, 1000), then the number of models in the database, multiplied by the set of significantly different parameters for forming facial images, results in a huge variability of the images displayed to the user. And all these images are completely unique - none of them can be found on the Internet, i.e. The main advantage of the method is that each image is almost guaranteed to be displayed once. And even the individual parts of the face that you want to find and indicate to the user (for example, the nose) look different, because they can be rotated, tilted, lit, etc. in different ways.

The disadvantage of methods based on the task for the user to select an object of a given class on the image, or to recognize and indicate with the cursor the required part of the object, is that this object or part of the body must have a name (“hand”, “eye”, etc.), so that the person understands the essence of the task. However, at present, both paid and free training image recognition software systems based on deep neural network technology are available to attackers. Powerful hardware accelerators from NVidia have been created for this technology. An annual computer face recognition competition is held. There are also competitions in computer recognition of animal species in images. Of course, neural networks are trained on completely different images, not those that will then be offered to the Internet robot when performing the captcha test. The database of images for training, as a rule, is formed on the basis of publicly available pictures from the Internet. And it’s quite easy to collect such a volume of the training sample, which is enough to teach the neural network to steadily recognize the position, for example, of the nose on any, even synthetic 3-dimensional faces that form in the prototype method under discussion. Those. the likelihood of an Internet robot passing captcha based on recognition of body parts is probably about 50%, which is completely unacceptable. Although the authors of the method propose to impose “noise” on the image to make it difficult for the robot to recognize the location of the parts of the body specified in the captcha, this entails (as in the case of character noise) an increase in the difficulty to give correct answers to the person.

A common drawback of captcha methods that contain, albeit a large, but still finite database of 3D models, is that there is a high probability by requesting images to obtain so many images of each model stored in the databases that it will be possible to restore it in the form of the original 3-dimensional model. Methods for implementing this approach when cracking captcha are also well known and widely available.

The problem solved by this method is to increase the reliability of recognizing the user whether he is a person or an Internet robot.

The technical result is to increase the probability of using a captcha test to distinguish a person from an Internet robot when accessing Internet resources.

Disclosure of the invention.

To achieve this result, a method for identifying a computer user "a person or an Internet robot", including the steps of:

(a) select one or more 3-dimensional models from the library of 3-dimensional models, each of the 3-dimensional models in the library contains geometric shape data in three spatial dimensions,

(b) from the models selected in step (a) in a virtual three-dimensional space, a group of models is formed formed by their deformation, location on the stage, turns, lighting and shooting options,

(c) generate at least one 2-dimensional image of the group that is formed in step (b), while the image description parameters are automatically generated and stored,

(d) inform the user about the sequence of necessary actions that the user must carry out to confirm that the user is not a robot; (e) when the user performs the sequence of response actions that he is informed about at step (d), describe the response of the user with a set of parameters and keep it

(e) perform an automatic comparison of the stored parameters of the response of the user with the parameters of the image description, which form at step (c), while

the group of 3-dimensional models that is formed in step (a) includes at least three models, at least one model that has a convex-concave shape and at least two models of the group, which are partially obscured by at least one group model. Besides.

- when performing step (e), the parameters of user actions are generated and stored so that at least one parameter is uniquely related to the number of models detected by the user in the image that is generated in step (c), and this "user" parameter in step ( f) compare with the corresponding image description parameter automatically generated in step (c), which characterizes the true number of models in the image,

- when performing step (e), the parameters of the user’s actions are formed so that at least one parameter is uniquely associated with the missing at least one model in at least one image, in at least one of pairs of images that are formed in step (c) based on one group of models that are formed in step (a),

- describe the required sequence of user actions in step (d), using to describe the relative position of at least two models of the group formed in step (b), words from the series: “on”, “c”, “under”, “ above ”,“ for ”,“ before ”,“ y ”,“ when ”,“ between ”; “Opposite”, “inside”, “to the left of”, “to the right of”, “to the left”, “to the right”, “to the side”, “near”, “next to”, “closer to”, “further from”, “Behind”, “behind”, “among”, “outside”,

- the models used in step (b) are synthesized by applying the operations of combining, intersecting and supplementing from a set of separate primitive forms selected in a pseudo-random manner.

The essence of the proposed method is that it is based on the ability of a person and the inability of a computer to correctly distinguish different unknown objects from a group of 3-dimensional objects in an image, perceive the shape of each selected object, and correctly recognize the number and spatial arrangement of different objects. Moreover, the most significant factor that impedes automated image analysis by a software robot is that it is quite complex in form, i.e. convex-concave and unfamiliar to the robot (objects have no names and unfamiliar in the sense - the robot was not trained on analogs) 3-dimensional objects partially obscure each other and form a group containing more than 3 objects.

Brief Description of the Drawings

In FIG. Figure 1 shows the result of the operation "union", which was used in the synthesis of a new object based on a cube and a ball.

In FIG. Figure 2 shows the result of the operation “taking the complement,” when the part that was inside the ball at some intersection of the cube and the ball was “bitten off” from the cube.

In FIG. 3 shows an image of a captcha test according to paragraph 2 of the claims of this application, in which the user is given the task of indicating (or entering from the keyboard) the number of different objects perceived by him in the image.

In FIG. 4 shows an image of a captcha test according to paragraph 3 of the claims of this application, in which the user is given the task of pointing to an object in one of two images, moreover, to that object that is not in the other, neighboring image.

In FIG. 5 shows an image of a captcha test according to paragraph 4 of the claims of this application, in which the user is given the task of pointing to some object in the image, which is between two identical objects

The present invention is directed to the development of such a captcha, which is based on the ability of a person to highlight areas related to different objects on an image, a person can easily perceive and describe the relative position of objects displayed on the image. It is assumed that the objects are not spaced in the frame, but can partially obscure each other, or have partial self-blocking. Moreover, in humans this ability is completely independent of whether objects known or unknown to a person are shown in the image. Currently, there is not even a statement about the existence of any reliably working systems (for non-trivial cases of grouping objects in a frame) for preliminary automatic partitioning of a frame into areas related to different objects. For all recognition systems currently on the market, finding in the frame the location of a recognizable object and, in fact, its recognition are always combined operations. That is, if training on its recognition was not carried out on an object of a certain class, then such an object for an automatic system cannot be separated from the background in the image. Some attempts to develop automatic allocation of different and unknown objects in the frame as a preliminary process in the analysis of the content of the frame should be considered ineffective (see http://www.topazk.ru/ssl/Projectl/Projectl.htm). The application of the currently popular neural network algorithms to solve this problem is unknown to the authors of the application. Therefore, it should be considered unlikely that attackers-crackers of the captcha proposed in this application will be able in the foreseeable future to create a unique software system with properties that a person easily demonstrates, but which no one has yet presented as the ability of some automated process.

For the method according to this application, as well as for the method selected as a prototype (US No. 7725395), as well as for most methods based on the visualization of 3D models, the common thing is that at each user request for access to a protected Internet the resource begins to automatically execute the sequence of the following working stages of the response (it is assumed that the user interaction scenario described below is the most typical, but not the only one for the method described in this application):

- choose one or several 3-dimensional models from the base of models or synthesize; if several models are selected, a group is formed from them in three-dimensional space;

- based on the selected or synthesized models, one or more images with random image formation parameters are generated, and some image description parameters are automatically generated and stored (for example, the coordinates of the locations of individual objects in the frame, the number of objects, etc.);

- further, in response to a user’s request, the server returns an image or several images to the user's browser, and also returns text information about the sequence of necessary actions (some task, for example, “enter the number of objects in the image”) that the user must carry out, in order to to confirm that he is not a robot;

- the user looks at the image, reads the text of the task and performs it;

- when the user performs the sequence of response actions, the user's browser generates a set of parameters, this set of parameters characterizes the user's perception regarding the image content, then the user browser sends the values of these parameters to the server;

- the server receives a parametric description of the response actions: the user and performs an automatic comparison of two sets of parameters, i.e. parameters characterizing the user's perception of the image content and the “correct” image description parameters that are stored on the server;

- in case of a “failure” for the user to compare the parameters of his response and the saved “correct” image characteristics (for example, the user incorrectly determined the number of objects in the image) - the server sends a denied access to the resource to the user's browser or vice versa, if the comparison is positive, the user is provided access, for example, to the registration page on the resource.

Our proposed method is intended to make a distinction - who exactly is the computer user - a person or a program running on another computer.

The methods for forming an image (or images) for captcha proposed in this application involve the use of ready-made 3-dimensional models or synthesized based on primitives by applying the following operations: combining (see Fig. 1), taking an intersection, taking an addition (see Fig. 2) .

It is assumed that the images of the CAPTCHA test show rather complex three-dimensional figures, and these figures are shown so that most of them either partially obscure other figures, or partially obscure one of the figures. The complexity of the figure means that the figure can be convex-concave, that is, the figure itself with one part of it can obscure the other part, as; for example, from a certain angle, a person’s hand obscures a part of the body. Complex volumetric figures also mean those that contain edges and faces, and the faces are not necessarily flat. In addition to the complexity of the figures themselves, a factor that significantly complicates the process of computer image analysis are shadows, which when modeling lighting complex objects of the group cast on each other.

According to the method described in this application, with its practical implementation in the form of captcha, in a three-dimensional virtual space a group of 3-dimensional objects is formed, which are quite complex geometric shapes. At the same time, lighting is simulated by a group of 3-dimensional models: from several point and distributed sources, so that shadows from the faces of other objects are observed on the surface of some objects (see Fig. 3). The shape of each object shown in the image arises as a result of a special algorithmic process - synthesis of a complex shape. The synthesis process of each object in the image has many parameters. So, the synthesis of a 3-dimensional model of an object is an algorithmic process of merging or another combination of simple 3-dimensional models.

The number (from three to eight) of complex objects in a group, lighting parameters (location of light sources, their type and number), as well as camera parameters (focal length, angle) - all these parameters will be set in an individual combination for each user request. Because with each access, the parameters are selected in a pseudo-random way (and these are: synthesis parameters of the forms, parameters of the relative positions of objects in the group, lighting parameters, group shooting parameters), in the aggregate, such a variety of parameters leads to the fact that the image that is proposed to be analyzed by the user, and to which questions are raised in the process of passing the captcha, it is actually demonstrated only once and in one single session of passing the captcha.

The implementation of the invention

At one stage of the test (captcha), where the image is shown: the user looks at the image, reads the text of the task and performs it.

A test task for a user may be a proposal to answer a question (and enter an answer using the keyboard, but not exclusively in this way) about the number of different objects that he observes in the group of 3D objects in the displayed image (see Fig. 3). An Internet robot that attackers could have created would probably do the wrong thing. For a person, the answer is not difficult.

Also, two images can be displayed simultaneously to the user. These images are formed for different groups of 3D models and for different lighting and shooting parameters. But, two groups of objects in two images differ only in that the composition of the group in one image differs only in the absence of one of the objects (3D models) from the composition of the group in another image (see Fig. 4). At the same time, the user is given the task of indicating that object on one of the images that is missing on another, “neighboring frame”. Since the objects in each image do not have a name, they are shown only once, and they are all significantly different in shape, and all their forms are quite complex, and in addition, there are partial screenings and self-blocking, and in addition, the objects have a "random" mutual arrangement, and the lighting conditions and the “shooting” perspective on two frames are significantly different, then in these conditions, determining which object is missing on one of the two frames is an insurmountable task for any computer vision program, but it’s completely trivial but for a person.

You can also distinguish between a person performing a captcha task or a software process based on a person’s ability to isolate different objects from a group of 3-dimensional objects, perceive their relative position and understand the simple text in Russian, reflecting the peculiarities of the relative position of objects expressed using prepositions and adverbs, given below. This ability can be expressed by a person by verbal description of what he sees in the image. But it is important for us that a person is quite able to answer complex questions asked in Russian about the relative position of objects in a group. At the stage of the test (captcha), when the user looks at the image and reads the text of the task, the text may contain a requirement, for example: "indicate the object that is located behind the object on the left side of which there is a symbol" A "." Those. the user is given tasks that can be performed not only on the condition of a “correct” perception of the spatial arrangement of individual objects in the frame, but also on the condition of a correct perception of the text of the task, which is formulated by a rather lengthy sentence in ordinary Russian. For example, there may be the following captcha task: “Indicate an object that is between two objects of the same shape” (Fig. 5). CAPTCHAs containing such a task-question, where the relative spatial position of two or more objects in three-dimensional space is indicated in the text, are also protected by this application. Such an indication of the relative position of objects in the text of the task, as a rule, means that they use at least one of the spatial prepositions: “on”, “in”, “under”, “above”, “for”, “before” , “Y”, “with”, “between”. Or else one of the following spatial dialects of the Russian language can be applied: “opposite”, “inside”, “between”, “left of”, “right of”, “left”, “right”, “aside”, “about ”,“ Next to ”,“ closer to ”,“ further from ”,“ behind ”,“ behind ”,“ among ”,“ outside ”.

Also, one or more 3-dimensional models, which include for subsequent visualization in a group of objects, are synthesized from a set of separate primitive forms by applying operations: combining, taking intersection, taking additions.

Claims (12)

1. A method of identifying a computer user "a person or an Internet robot", comprising the steps of: (a) select one or more 3-dimensional models from the library of 3-dimensional models, each of the 3-dimensional models in the library contains geometric shape data in three spatial dimensions, (b) from the models selected in step (a) in a virtual three-dimensional space, a group of models is formed formed by their deformation, location on the stage, rotation, lighting and shooting options, (c) generating at least one 2-dimensional image of the group that is formed in step (b), while automatically describing and generating image description parameters, (d) inform the user about the sequence of necessary actions that the user must take to confirm that the user is not a robot, (e) when the user performs a sequence of response actions, about which they inform him at step (d), describe the response of the user with a set of parameters and save it, (e) carry out an automatic comparison of the stored parameters of the response of the user with the parameters of the image description, which form in step (c), characterized in that the group of 3-dimensional models that are formed in step (a) includes at least three models: at least one model that has a convex-concave shape, and at least two models of the group that are partially obscured by at least one group model. 2. The method according to p. 1, characterized in that when performing step (e), the parameters of the user’s actions are formed and stored so that at least one parameter is uniquely associated with the number of models detected by the user in the image that is generated in step (c), and this "user" parameter in step (e) is compared with the corresponding image description parameter automatically generated in step (c), which characterizes the true number of models in the image. 3. The method according to p. 1, characterized in that when performing step (e), the parameters of the user’s actions are formed so that at least one parameter is uniquely associated with the missing at least one model, in at least one image, in at least at least one of the pairs of images that are formed in step (c) based on one group of models that is formed in step (a). 4. The method according to p. 1, characterized in that they describe the desired sequence of user actions in step (d), using at least two models of the group formed in step (b), words from the series “on”, “ c ”,“ under ”,“ above ”,“ for ”,“ before ”,“ y ”,“ when ”,“ between ”,“ opposite ”,“ inside ”,“ to the left of ”,“ to the right of ”,“ to the left ”,“ to the right ”,“ aside ”,“ near ”,“ next to ”,“ closer to ”,“ further from ”,“ behind ”,“ behind ”,“ among ”,“ outside ”. 5. The method according to p. 1, characterized in that the models used in step (b) are synthesized by applying the operations of combining, intersecting and supplementing from a set of separate primitive forms selected in a pseudo-random manner.

Images