Security server system implementation method, medium, security system and device

By analyzing the feature set of network requests using machine learning models, accurately identifying and blocking access to robot programs, the false positive problems and real-time detection problems in the prior art are solved, and the accuracy and real-time detection of detection are improved.

CN119011177APending Publication Date: 2024-11-22F5 NETWORKS INC
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410618809.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2023-05-17
Filing Date
2024-05-17
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

Existing robot program detection methods often produce a large number of false positives, resulting in legal users being misidentified as robot programs, and it is difficult to accurately identify and block access to robot programs in real time.

Method used

Using a machine learning model, a feature set is generated by receiving and associated with the logs of network request features, submitted to the machine learning model for analysis, rejecting requests from the robot program, and allowing legitimate users to access.

Benefits of technology

Accurate identification and blocking of robot programs is achieved, false positives are reduced, normal access experience for legitimate users, and real-time and accuracy of detection are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119011177A_ABST
    Figure CN119011177A_ABST
Patent Text Reader

Abstract

A method implemented by a secure server system, a non-transitory computer-readable medium, a secure server device, and a secure system are shown. With this technique, a request for a web page from a client device to a server device is received. The request has one or more associated logs. A set of features is generated based on one or more associated logs. The set of features is submitted as input to a machine learning model. The request is denied in response to an output of the machine learning model indicating that the request for the web page from the client device is initiated by the robotic program.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The technology relates to methods and systems for automatically detecting web requests originating from bots using machine learning models. Background Art

[0002] A bot is a software application that runs automated tasks on the Internet, often designed to mimic or impersonate human behavior. A bot attack is a malicious attack that leverages a series of connected computers to attack networks, devices, and / or websites. Such attacks often have the effect of significantly degrading the performance of targeted systems and devices. Bots are ubiquitous in web applications, and malicious bots are problematic because they can manipulate information, spread misinformation, and promote unverified information, which can have an adverse impact on end users. As a result, bots can present a variety of issues, such as degraded system performance, information leakage, security threats, and negatively impact the experience of legitimate users of web services.

[0003] Currently, there are many solutions for detecting bots. However, current solutions often generate a large number of false positives, resulting in legitimate users being classified as bots. Detecting bot activity is complicated because many bots actively avoid detection by changing their IP addresses and user agents. A user agent is a special type of software agent that retrieves, presents, and facilitates end-user interaction with web content on behalf of a user, such as a web browser or email reader. Software agents typically identify themselves, their application type, operating system, device model, software vendor, and / or software revision. For example, in the hypertext transfer protocol (HTTP), such identification is transmitted in the header field User-Agent. However, malicious bots may provide header (or other such) information that is intended to mask their true nature. Because legitimate users can provide a large number of header information values, and legitimate values ​​for bots vary constantly, using rule-based approaches to filter network traffic based on header information may result in many legitimate users being incorrectly identified as bots or failing to identify actual bots.

[0004] Therefore, there is a need to create a system to more accurately identify network requests originating from bots, a system that can be deployed in near real-time to identify and block access to bots while correctly identifying and authorizing access to legitimate users. Summary of the invention

[0005] A method implemented by one or more secure server devices, server devices, or client devices, includes receiving a request for a web page from a client device to a server device. The request is associated with a log. A feature set is generated based on one or more associated logs. The feature set of the method is submitted as input to a machine learning model. The request is rejected in response to the machine learning model generating an output indicating that the request for the web page from the client device was initiated by a robot program.

[0006] A secure server device includes a memory and one or more processors, the memory including programming instructions stored thereon, the one or more processors being configured to be able to execute the stored programming instructions to receive a request for a web page from a client device to a server device. The request is associated with a log. A feature set is generated based on one or more associated logs. The feature set of the method is submitted as input to a machine learning model. The request is rejected in response to an output of the machine learning model indicating that the request for the web page from the client device was initiated by a robot program.

[0007] A non-transitory computer-readable medium having instructions for including executable code stored thereon, when the executable code is executed by one or more processors, causes the processor to receive a request for a web page from a client device to a server device. The request is associated with a log. A feature set is generated based on one or more associated logs. The feature set of the method is submitted as input to a machine learning model. The request is rejected in response to an output of the machine learning model indicating that the request for the web page from the client device was initiated by a robot program.

[0008] A secure server system includes one or more secure server devices, a server device, or a client having a memory and one or more processors, the memory including programming instructions stored thereon, and the one or more processors are configured to be able to execute the stored programming instructions to receive a request for a web page from a client device to a server device. The request is associated with a log. A feature set is generated based on one or more associated logs. The feature set of the method is submitted as input to a machine learning model. The request is rejected in response to an output of the machine learning model, the output indicating that the request for the web page from the client device was initiated by a robot program.

[0009] The technology provides numerous advantages, including providing methods, non-transitory computer-readable media, secure server devices, and secure server systems that facilitate controlling access to configuration data in multiple management systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is a block diagram of an exemplary secure server system;

[0011] Figure 2 is a block diagram of an exemplary secure server device;

[0012] Figure 3 is a flow chart of an exemplary method for providing bot detection using a machine learning model;

[0013] Figure 4 is a table of exemplary numerical features and categorical features that may be derived by a secure server device according to an embodiment of the present invention; and

[0014] Figure 5 is a flow chart of an exemplary method for detecting a bot using a machine learning model.

[0015] Although each figure illustrates a specific example, in order to illustrate a clear example, other examples may omit, add, reorder or modify any element shown in the figure. In order to illustrate a clear example, one or more figures may be described with reference to one or more other figures, but the specific arrangement shown in one or more other figures is not required to be used in the example. DETAILED DESCRIPTION

[0016] In the following description, for the purpose of explanation, in order to provide a comprehensive understanding of the present invention, many specific details are set forth. However, it is apparent that the present invention can be practiced without these specific details. In other cases, well-known structures and devices are shown in the form of block diagrams to avoid unnecessary confusion of the present invention.

[0017] It should be further understood that: unless otherwise expressly specified, the term "or" may be inclusive or exclusive; the term "set" may include zero, one, or two or more elements; the terms "first", "second", "a" and "particular" are used as naming conventions to distinguish elements from each other, and unless otherwise specified, do not mean the order, time or any characteristics of the referenced items; the terms "such as", "for example", "as an example" and the like describe one or more examples, but are not limited to the described examples; the terms "include" and / or "comprise" specify the existence of the stated features, but do not exclude the existence or addition of one or more other features.

[0018] A "computer system" refers to one or more computers, such as one or more physical computers, virtual computers, and / or computing devices. For example, a computer system may be or may include one or more server computers, cloud-based computers, cloud-based computer clusters, virtual machine instances or virtual machine computer elements (e.g., virtual processors, storage, and memory), data centers, storage devices, desktop computers, laptop computers, mobile devices, or any other special-purpose computing devices. Unless otherwise expressly stated, a "computer system" referred to herein may refer to one or more computers. When a computer system performs an action, the action is performed by one or more computers of the computer system.

[0019] "Client" (also called "client device") refers to a combination of integrated software components and computing resource allocations, such as memory, computing devices, and processes on the computing devices for executing the integrated software components. The combination of software and computing resources is configured to interact with one or more servers over a network such as the Internet. A client can refer to a combination of components on one or more computers, or it can refer to the one or more computers.

[0020] "Server" (also referred to as "server system", "server computer system", "server device", or "server apparatus") refers to a combination of integrated software components and an allocation of computing resources, such as memory, computing devices, and processes on the computing devices for executing the integrated software components. The combination of software and computing resources provides a particular type of functionality on behalf of the server's clients. A server may refer to a combination of components on one or more computers, or may refer to the one or more computers. A server may include multiple servers; that is, a server may include a first server computing device and a second server computing device, which may provide the same or different functionality to the same or different sets of clients.

[0021] As used herein, the term "website" may refer to a collection of web pages, web resources, and other web content provided through the Internet. For example, a first website may include web content hosted by a web server system such as a first server device, and a second website may include web content hosted by a second web server system such as a second web server device. Typically, the web content of a website shares a single domain name or a collection of related domain names owned or controlled by a common entity. A website may include web page content accessible on the Internet through a browser, and / or web content accessible from another client application through the Internet. Such applications may include web applications executed in a browser and / or local applications installed on a client computing device (also referred to as a "client device").

[0022] This document generally describes systems, methods, devices, and other techniques for automatically determining whether a network request is made by a robot program or a human user of a client device. An entity operating a network server system may wish to prevent robot programs from accessing content hosted by the network server system while allowing legitimate human users to access the content. The disclosed technology may involve using one or more machine learning models to automatically determine whether a given network request is initiated by a robot program or a human. In some embodiments, the disclosed technology may include using multiple machine learning models to make such a determination, for example, by combining various machine learning models into an integrated model and using a majority vote of their outputs to make a final determination about whether the request originated from a robot program or a human. Various embodiments may use an integrated machine learning model composed of one or more basic models, which may include supervised machine learning models (e.g., random forest models and logarithmic regression models), semi-supervised machine learning models (e.g., autoencoders with neural networks, autoencoders with k-means clustering algorithms, autoencoders with k-nearest neighbor algorithms, autoencoders with random forests) and / or unsupervised machine learning models (e.g., autoencoder models, convolutional neural networks with autoencoders).

[0023] The technology disclosed herein relates to deriving features from historical network request logs that can be used to train machine learning models, and deriving similar features from currently received network requests, and then using the similar features as input to the training model to generate an output indicating whether the received network request is generated by a robot program or a human. According to some embodiments, the process of generating features used as input to the machine learning model may include generating a first feature set from the network request log, deriving a second feature set from the first feature set, and then deriving a reduced feature set from the second feature set. The second feature set may include numerical features and categorical features. Part of the process may include converting categorical features into numerical values ​​using techniques such as one hot encoding and / or generating word embeddings using transformer models such as word2vec or bidirectional encoder representation (BERT). In addition, feature scaling may be applied to features to normalize the feature range of the data. When the second feature set has been manipulated so that all features are represented in numbers and have been scaled to normalize the impact of each feature relative to other features, the features may be input into an autoencoder model, which outputs a reduced feature set of a number that a user can select. For example, in some embodiments, the autoencoder model may generate a reduced set consisting of 12 features. This is advantageous because the reduced number of features allows the machine learning model to execute in a relatively fast manner, but the reduced feature set still encapsulates all the information from the larger feature set, which allows the system to make more accurate predictions about whether the request originated from a bot or a human. Additionally, using an ensemble model of different machine learning models using majority voting can also improve the accuracy of the predictions.

[0024] refer to Figure 1 , an exemplary secure server system 100 is shown. The secure server system 100 in this example includes a secure server device 102, which is coupled to a server device 104 and a client device 106 via a communication network 110, although the secure server device 102, the server device and the client device 106 can be coupled together via other topologies. The secure server system 100 may also include other network devices, such as routers or switches, which are known in the art and will not be described herein. The technology provides many advantages, including providing a method for automated robot program detection using one or more machine learning models based on information extracted from logs generated in association with received network requests, non-transitory computer-readable media, secure server systems 100, and secure server devices 102.

[0025] In this particular example, secure server device 102, server device 104, and client device 106 are configured to communicate with one another in a secure manner. Figure 1 100 as dedicated hardware devices. However, one or more of the secure server device 102, the server device 104, and the client device 106 may also be implemented in software within one or more other devices in the secure server system 100. For example, the secure server device 102 may be hosted by one or more of the server devices 104, and other network configurations may also be used. Although the description herein is generally directed to a system distributed across multiple devices, it should be understood that it is contemplated that the techniques disclosed herein may alternatively be adapted for use on a centralized system.

[0026] refer to Figure 1-Figure 2 , the secure server device 102 of the secure server system 100 can perform any number of functions, generate features from network request logs, train one or more machine learning models, and automatically detect robot programs using one or more machine learning models. In this example, the secure server device 102 includes a processor 202, a memory 204, and a communication interface 206 coupled together by a bus, although the secure server device 102 may include other types or numbers of elements in other configurations.

[0027] The processor 202 of the secure server device 102 may execute programmed instructions stored in the memory 204 of the secure server device 102 for any number of functions described and illustrated herein. For example, the processor 202 of the secure server device 102 may include one or more central processing units (CPUs) or a general-purpose processor having one or more processing cores, although other types of processors may also be used.

[0028] The memory 204 of the secure server device 102 stores these programming instructions for one or more aspects of the present technology described and illustrated herein, although some or all of the programming instructions may be stored elsewhere. Various different types of memory storage devices, such as RAM (Random Access Memory), ROM (Read-Only Memory), hard disks, solid-state drives, flash memory, or other computer-readable media, may be used for memory, which other computer-readable media may be read from and written to by a magnetic, optical, or other read-write system coupled to a processor.

[0029] Thus, the memory of the secure server device 102 may store one or more modules, which may include computer executable instructions that, when executed by the secure server device 102, cause the secure server device 102 to perform actions, such as sending, receiving, or otherwise processing network messages, such as and executing the following reference Figure 3-Figure 5 Other actions described and illustrated. Modules can be implemented as components of other modules. In addition, modules can be implemented as applications, operating system extensions, plug-ins, etc.

[0030] Further, the modules may be run in a cloud-based computing environment. The modules may be executed within or as a virtual machine or virtual server managed in a cloud-based computing environment. In addition, the modules, and even the security server device 102 itself, may be located in a virtual server running in a cloud-based computing environment, rather than being bound to one or more specific physical network computing devices. In addition, the modules may be run in one or more VMs executed on the security server device 102. In addition, in one or more examples of the technology, the virtual machines running on the security server device 102 may be managed or supervised by a hypervisor.

[0031] In this particular example, the memory of the secure server device 102 includes a feature generation module 210. According to some examples, the feature generation module 210 can generate, modify, and reduce feature sets that are used to train one or more machine learning models and serve as inputs to the machine learning models to identify whether a given network request was initiated by a human or a robot program. According to ... Figure 3 One or more of steps 302, 303, 304, 306, 308, 310 and / or 318 of process 300 shown in FIG. 3 generate a reduced feature set from logs associated with a network request, which will be described in more detail below.

[0032] In some examples, feature generation module 210 may use a set of logs stored in association with a set of historical network requests (e.g., in a database of secure server device 102) to generate a set of features to be used as training data for a machine learning model. In other words, a database of previously recorded network requests may be stored by secure server device 102 (or another storage device accessible by secure server device 102) along with the logs associated with the stored network requests, and secure server device 102 may utilize these stored logs associated with the set of previous network requests to generate a reduced set of features to train one or more machine learning models.

[0033] In some examples, the stored set of network requests and associated logs may be labeled so that for each network request, it is known whether the network request is initiated by a robot program or a human. In some examples, the set of historical network requests may be unlabeled, in which case the historical network requests may be processed to be labeled before being used as a training data set. In some examples, the historical network requests may be manually labeled by a human who examines each request and determines (e.g., based on the activity associated with the request) whether the request is executed by a robot program or a human. According to some embodiments, the feature generation module 210 may be configured to automatically identify historical network requests made by simple robot programs (e.g., it may determine that a network request made by an uncommon browser is a request made by a robot program) by, for example, applying heuristics, comparing the request with a stored library, and / or applying a rule-based approach. While such an approach helps identify simple robot programs, it may not be able to identify more advanced or complex robot programs, in which case humans may need to review network requests that are not identified as simple robot programs to determine which of the remaining requests are also created by robot programs. After such a determination, the feature generation module 210 may, for each network request in the stored set of historical network requests, mark the associated stored log as being associated with a network request initiated by a human or a robot.

[0034] The memory 204 of the secure server device 102 may also include a machine learning model generation module 212. Figure 3 As described in more detail in steps 312, 314 and 316, the machine learning model generation module 212 can use the reduced feature set generated by the feature generation module 210 as training data to train one or more machine learning models.

[0035] The memory 204 of the secure server device 102 may also include a bot detection module 214 that can automatically make a determination as to whether a newly received network request is initiated by a bot or a human, and take appropriate action in response to the determination, as described below with respect to Figure 3In other words, the bot detection module 214 can automatically detect whether a network request received from the client device 106 directed to the server device 104 is generated by a bot program running on the client device 106, or is generated in response to a human operating the client device 106. If the bot detection module 214 determines that the request is initiated by a human, the bot detection module 214 can allow the request, thereby authorizing access to the requested network resources hosted by the server device 104. Conversely, if the bot detection module 214 determines that the request is initiated by a bot program, the bot detection module 214 can take some security mitigation actions, such as denying the request and preventing the client device 106 from accessing the requested network content on the server device 104 that is the subject of the request, or rate limiting the client device 106 to limit the amount of interaction that the client device 106 can have with the server device 104 in a given time period.

[0036] like Figure 1 As shown in , in some examples, the secure server device 102 can be located between the client device 106 and the server device 104, so that the secure server device 102 can intercept network requests sent to one or more server devices 104 to determine whether the request originates from a robot program or a human. In this way, the secure server device 102 can protect the server device 104 from robot programs accessing its system and network content in near real time, while allowing legitimate human users to access.

[0037] The communication interface 206 of the secure server device 102 is operable to couple and communicate between the secure server device 102, the client device 106, and the server device 104, which are coupled together at least in part via the communication network 110, although other types and numbers of communication networks or systems having other types or numbers of connections or configurations to other devices or elements may also be used.

[0038] By way of example only, the communication network 110 may include a local area network (LAN) or a wide area network (WAN), and may use TCP / IP and industry standard protocols over Ethernet, although other types or numbers of protocols or communication networks may be used. The communication network 110 in this example may employ any suitable interface mechanism and network communication technology, including, for example, telex in any suitable form (e.g., voice, modem, etc.), a Public Switched Telephone Network (PSTN), an Ethernet-based Packet Data Network (PDN) hosting a HyperText Transfer Protocol (HTTP), combinations thereof, and the like.

[0039] The secure server device 102 is shown in this example as including a single device, but in other examples may include multiple devices, each device having a processor (each processor having a processing core) that implements the steps of the technology. In these examples, one or more of the devices may have a dedicated communication interface or memory. Alternatively, one or more of the devices may utilize the memory 204, communication interface 206, or other hardware or software components included in one or more other devices in the secure server device 102.

[0040] Thus, one or more of the devices that together comprise the security server device 102 may be standalone devices, or integrated with one or more other devices or apparatuses (e.g., the server device 104). In addition, for example, one or more of the devices of the security server device 102 may be in the same or different communication networks (including one or more public, private, or cloud networks). In other examples, multiple protection server devices are deployed within and / or outside a local area network associated with the server device 104, which communicates with any number of application servers or end-station computers, and other network topologies may also be used.

[0041] In some examples, the security server device 102 may also be coupled to at least one output device via an I / O subsystem. The output device may be a digital computer display. Examples of displays that may be used in various embodiments include a touch screen display or a light-emitting diode (LED) display or a liquid crystal display (LCD) or an electronic paper display. Alternatively or in addition to a display device, the security server device 102 may include other types of output devices. Examples of other output devices include a printer, a ticket printer, a plotter, a projector, a sound card or a video card, a speaker, a buzzer or a piezoelectric device or other audible device, a lamp or an LED or LCD indicator, a tactile device, an actuator or a servo mechanism.

[0042] In some examples, at least one input device can be coupled to the I / O subsystem of the secure server device 102 for communicating signals, data, command selections, or gestures to the processor 304. Examples of input devices include touch screens, microphones, still and video digital cameras, alphanumeric and other keys, keypads, keyboards, graphic tablets, image scanners, joysticks, clocks, switches, buttons, dials, sliders, and / or various types of sensors such as force sensors, motion sensors, thermal sensors, accelerometers, gyroscopes, and inertial measurement unit (IMU) sensors and / or various types of transceivers such as wireless transceivers such as cellular or Wi-Fi transceivers, radio frequency (RF) or infrared (IR) transceivers, and Global Positioning System (GPS) transceivers.

[0043] Another type of input device is a control device that can perform cursor control or other automated control functions, such as navigation in a graphical interface on a display screen, instead of or in addition to the input function. The control device can be a touch pad, a mouse, a trackball, or cursor direction keys, which are used to transmit direction information and command selections to the processor 202 and to control the movement of the cursor on the display. The input device can have at least two degrees of freedom on two axes (a first axis (e.g., x) and a second axis (e.g., y)), which allows the device to specify a position in a plane. Another type of input device can be a wired, wireless, or optical control device, such as a joystick, a rod, a console, a steering wheel, a pedal, a gear shift mechanism, or other types of control devices. The input device can include a combination of multiple different input devices, such as a camera and a depth sensor.

[0044] In another example, the secure server device 102 may be an Internet of Things (IoT) device or other type of network device, wherein one or more of the output device, input device, and control device are omitted. Alternatively, in such an embodiment, the input device may be one or more cameras, motion detectors, thermometers, microphones, earthquake detectors, other sensors or detectors, measuring devices, or encoders, and the output device may be a dedicated display, such as a single-line LED or LCD display, one or more indicators, display panels, meters, valves, solenoids, actuators, or servo mechanisms.

[0045] Return to reference Figure 1 In this example, the server device 104 of the secure server system 100 may be a commercial web server hosting one or more web pages accessible to the client device 106. The web pages may request network resources from any number of other domains. Example types of requests include XMLHttpRequests, Fetch API requests, WebSocket, WebRTC, Beacon, form submission, resource acquisition (e.g., links, scripts, images, videos, etc.), etc.

[0046] Each of the server devices 104 includes a processor, a memory, and a communication interface, which are coupled together by a bus or other communication link (not shown), although other numbers or types of components may be used. The server devices 104 in this example may include, for example, application servers or database servers, although other types of server devices 104 may also be included in the secure server system 100. In this example, the server devices 104 can exchange communications to facilitate interaction with a hosted web application, which is optionally protected by a user of a client device 106 by the secure server device 102 as described and illustrated herein.

[0047] Thus, in some examples, server device 104 may process network requests (e.g., requests to view network content, log in, and other requests) received from client device 106 via communication network 110 according to, for example, an HTTP-based application RFC protocol. A protected network application may operate on server device 104 and send data (e.g., a file or web page) to client device 106 in response to a request from client device 106. Server device 104 may be hardware or software, or may represent a system having multiple servers in a pool, which may include an internal or external network.

[0048] Although server device 104 is shown as a single device, in other examples, server device 104 may include multiple servers distributed across one or more different network computing devices that bring server device 104 together. In addition, server device 104 is not limited to a particular configuration. Thus, server device 104 may include network computing devices that operate using a master / slave approach, whereby one of the network computing devices of server device 104 operates to manage or otherwise coordinate the operations of the other network computing devices. For example, server device 104 may operate as multiple network computing devices within a cluster architecture, a peer-to-peer architecture, a virtual machine, or within a cloud architecture.

[0049] Therefore, the technology disclosed herein should not be construed as limited to a single environment, and other configurations and architectures are also contemplated. For example, the server device 104 can operate within the secure server device 102 itself, rather than as a separate server that communicates with the secure server device 102 via the network 110. In this example, the server device 104 can operate within the memory 204 of the secure server device 102. Therefore, any number of client devices 106 and server devices 104 can be used. The number of client devices 106 and server devices 104 is Figure 1 It is shown only for the purpose of providing illustrative examples.

[0050] In this example, each of the client devices 106 of the secure server system 100 includes any type of computing device that can exchange network data and interface with web pages, such as a mobile, desktop, laptop, or tablet computing device, a virtual machine (including a cloud-based computer), etc. In this example, each of the client devices 106 includes a processor, a memory, and a communication interface, which are coupled together by a bus or other communication link (not shown), although other numbers or types of components may also be used.

[0051] The client device 106 may run an interface application, such as a standard web browser or a local client application, which may provide an interface to make requests to the server device 106 via the network 110 and receive content stored on the server device 106. The client device 106 may further include a display device, such as a display screen or a touch screen, or an input device, such as a keyboard (not shown).

[0052] Although an exemplary secure server system 100 having a secure server device 102, a server device 104, a client device 106, and a communication network 110 is described and illustrated herein, other types or numbers of systems, devices, components, or elements in other topologies may also be used. It should be understood that the exemplary systems described herein are for exemplary purposes, as many variations of the specific hardware and software used to implement the examples are possible, as will be appreciated by those skilled in the relevant art.

[0053] One or more components described in secure server system 100, such as secure server device 102, server device 104, or client device 106, may be configured to run as virtual instances on the same physical machine. In other words, one or more of secure server device 102, server device 104, or client device 106 may run on the same physical device, rather than as separate devices communicating over a communication network. Figure 1 More or fewer secure server devices 102, server devices 104, or client devices 106 as shown.

[0054] In addition, in any example, two or more computing systems or devices can replace any one of the systems or devices. Therefore, the principles and advantages of distributed processing (such as redundancy and replication) can also be implemented as needed to improve the robustness and performance of the devices and systems of the examples. The example can also be implemented on a computer system extended on any suitable network using any suitable interface mechanism and traffic technology, which computer system only includes wireless traffic networks, cellular traffic networks, packet data networks (Packet Data Network, PDN), the Internet, intranets and combinations thereof by way of example.

[0055] For one or more aspects of the present technology, examples may also be embodied as one or more non-transitory computer-readable media having executable or programmed instructions stored thereon (e.g., in the memory 204, ROM, or storage of the secure server device 102), as described and illustrated by examples herein. The instructions in some examples include executable code that, when executed by one or more processors (e.g., one or more processors 202 of the secure server device 102), causes the processor to perform the steps required to implement the methods of the examples of the present technology described and illustrated herein.

[0056] refer to Figure 3, a flow chart of an example method 300 for providing robot program detection using a machine learning model is shown. In step 302 in this example, the security server device 102 of the security server system 100 obtains raw text features from the log. As will be appreciated by those skilled in the art, a network log (which may be referred to as a "traffic log" or "log") may be created based on and associated with the creation and / or receipt of a network request. The network log provides an audit trail of each network request responded by the server and contains various information about the request. For example, when a client device 106 sends a network request to access the content of a server device 104, a real-time traffic log may be created, which includes raw data, such as the Internet Protocol (IP) address of the client device, the Uniform Resource Locator (URL) of the requested website or webpage, the request path, and other such information.

[0057] According to some examples, one or more raw text features can be read from the log. In an example, the raw text features read from the log may include one or more of an IP address, HTTP version, request path, HTTP method, user agent, and HTTP header. The IP address is the IP address of the requesting device (e.g., client device 106). The HTTP version is the version of HTTP that the requesting device is using (e.g., HTTP / 1.0 or HTTP / 1.1). The request path is the information string following the top-level domain name in the URL. The HTTP method indicates an HTTP request that requests the action (e.g., GET, POST, etc.) performed by the server. The user agent is a header that identifies the type of software agent that the requesting device is using (e.g., the type of browser, the type of email reader, etc.). For example, the user agent can indicate whether the requesting device is using Chrome, Firefox, Safari, or some other browser when making a network request. The HTTP header is a field of the HTTP request that conveys additional context and metadata about the request. As will be appreciated by those skilled in the art, reading raw text features from a log associated with a network request may involve some cleaning and / or reformatting of the data. For each newly received request, the secure server device 102 may store an associated log and a determination of whether the request was initiated by a bot or a human.

[0058] When used to create features to train a machine learning model, in step 303, the secure server device 102 can obtain raw text features from stored logs associated with multiple past network requests based on accessing historical request data. The historical request data may include stored logs associated with previous requests, which may be stored in the memory 204 of the secure server device 102 or in some other data storage location accessible by the secure server device 102. The logs may correspond to multiple previous network requests made by one or more client devices 106. In some examples, the historical request data is marked as having been initiated by a human or a robot. In other words, in some examples, for each network request in the stored historical data, it is known whether the network request was initiated by a robot or a human.

[0059] When used to attempt to classify the new incoming request as being initiated by a robot or a human, the secure server device 102 can obtain raw text features from one or more logs associated with the new request received in step 318. In step 318, a new network request can be generated by the client device 106 and sent to the server device 104. In some examples, the network request can be redirected to the secure server device 102 or otherwise intercepted by the secure server device 102 before reaching the target server device 104. In some examples, the target server device 104 can receive the request and forward it to the secure server device 102 for processing before responding to the request. In this way, the secure server device 102 can make a determination as to whether the newly received request is initiated by a robot or a human to prevent the server device 104 from responding to the request to provide access to the requested content if the request is determined to be initiated by a robot.

[0060] In step 304, the security server device 102 extracts categorical features and numerical features from the raw text features obtained in step 302. A numerical feature is a feature that is a number (e.g., _timestamp, path_ / _count, etc.). A categorical feature is a feature that is not a numerical value (e.g., IP address, user agent, city / country name, etc.), although it may include numbers. According to some embodiments, user agent features, data, or aspects thereof may be obtained as raw text features of a log, and may also be considered as extracted categorical features and processed. For example, in some embodiments, user agent features (or aspects thereof) may be read from a log, and may be extracted as categorical features by transcribing / translating the data into a form used as one or more categorical features (e.g., browser type may be extracted from user agent raw text features as user agent categorical features). The security server device 102 extracts categorical features and numerical features from the raw text features using various rules and / or algorithms. For example, in step 302, the raw text feature is a path (e.g., "examplepath.php") that can be obtained from a log of a network request (e.g., by scanning and copying it from the log), and from the raw text feature, a categorical feature / numerical feature can be extracted by, for example, extracting an "is_script" feature from the path and converting it into a numerical feature, the "is_script" feature can represent the final or extension type of the path (e.g., ".php", ".jsp", ".asp", etc.). For example, in some embodiments, if a predetermined list of extensions does not exist in the subject path, it can be converted to a numerical feature "0", and if the subject path includes an extension / final type that matches one in the predetermined list, it can be converted to a numerical feature "1". Alternatively, in some examples, the "is_script" feature can be converted into a numerical feature by converting the extension type into a number based on a predetermined correlation (e.g., when "is_script" is ".php", the numerical feature can be "1", when it is ".jsp", the numerical feature can be "2", when it is ".asp", the numerical feature can be "3", and so on). After the process of extracting categorical features and numerical features from raw text features, the number of features may be greatly increased because, in some cases, multiple features may be generated using the same raw text features or a combination of raw text features.

[0061] Figure 4The table shown in shows an exemplary non-exhaustive list of example categorical features and numerical features that can be extracted from the raw text features, and provides an explanation of each feature. Other features that can be extracted may include "ua_error" (which indicates an error in parsing the UA using (e.g., when using Python)), "host_error" (which indicates an error in parsing the host header (e.g., when using Python)), "is_env" (which indicates that the path ends with "env"), "path_ends_ / " (which indicates that the path ends with " / "), "path_._count" (which indicates the amount of "." in the path), "ua_parser_class" (which provides an indication that the Python module has detected that this UA is a robot UA), "old_ua" (which indicates that the browser is a very old version (e.g., based on various specified statistics, values, and / or features associated with the browser), "mobile_ua" (which indicates that the US is a mobile device (e.g., on a smartphone)), "ua_*" (which indicates a user agent (e.g., a browser type, such as Chrome, Firefox, Safari, etc.)), "os" (which indicates an operating system (e.g., Windows, Linux, Mac OS, etc.)), “http_ver_*” which indicates the HTTP version (e.g., 1.0, 1.1, or none)), “http_method_cat_*” which indicates the category of HTTP methods used, which may include common methods (e.g., GET, POST, HEAD, PUT), unusual methods (e.g., DELETE, OPTIONS, PATCH, SEARCH, DEBUG, INDEX, CONNECT, REGISTER, PROPFIND), tracking / tracing methods (e.g., TRACK or TRACE), and unused methods (e.g., remaining methods)), and “ip_cat_*” which indicates the category of Internet protocol being used (e.g., IPv6, IPv4 global, IPv4 dedicated, etc.). It should be understood that the various raw text features, digital text features, and categorized text features described herein are merely exemplary, and in different examples, the security server device 102 may obtain, derive, and / or utilize more, fewer, or different features.

[0062] In step 306, the security server device 102 converts the categorical features into numerical features. Different methods can be used to convert categorical features into numerical features, such as one-hot encoding and creating word embeddings (e.g., using a transformer model). The numerical features can be vectors. For categorical features that are expected to have a relatively small number of categories, the security server device 102 can convert the categorical features into numerical features using one-hot encoding. As will be appreciated by those skilled in the art, one-hot encoding is a process involving assigning numbers to each possible category. For example, if the feature representing the time zone of the requesting device is one-hot encoded, "Atlantic Standard Time" may be represented by "1", "Eastern Standard Time" may be represented by "2", Central Standard Time may be represented by "3", and so on. As will be further appreciated by those skilled in the art, categorical features can be converted to numerical features by creating word embeddings using a transformer model (e.g., word2vec or bidirectional encoder representation (BERT)), which can take a string of words and convert them into vectors encoding their meaning. This technique is particularly useful for converting categorical features (such as user agents) because there are a large number of possible variations of legitimate user agents (e.g., revisions / new versions of browsers) that are constantly changing, as well as variations / misspellings of bots attempting to impersonate legitimate user agents, which may include subtle differences in wording / spelling. In one example, the user agent feature is converted to a numerical feature by creating word embeddings using a transformer model, and all other categorical features are converted to numerical features using one-hot encoding. However, it should be understood that this is merely an example, and in different embodiments, different methods may be used to convert different categorical features to numerical features.

[0063] In step 308, the security server device 102 performs feature scaling on the digital features and the numerical features to generate scaled digital features and scaled numerical features. As will be appreciated by those skilled in the art, feature scaling is a known process of normalizing the range of a feature relative to other features so that a particular feature does not inadvertently have an excessive impact on the model. In some embodiments, this step may be optional or may include other aspects, such as intentionally weighting the selected feature as being more important or more influential than other selected features.

[0064] In step 310, the security server device 102 generates a reduced feature set based on the scaled digital features and the scaled numerical features. The security server device 102 can generate a reduced feature set by inputting the scaled digital features and the scaled numerical features into an autoencoder model, which outputs a reduced feature set represented as a vector. The autoencoder uses a nonlinear transformation to encode a given vector into a smaller dimension, and it can generate a complex encoding based on the feature set, which can accurately capture various patterns of robot programs and humans. In some examples, the user can input a selected number of reduced features for the autoencoder model to output. In one example, the autoencoder model is set to output a reduced feature set consisting of a total of 12 features. Therefore, in this case, after the above steps, the network request initially represented by the data captured in the log in step 302 can eventually be represented by a 12-dimensional vector in step 310, although it should be understood that a different number of features can be selected in other examples. In experiments, it was found that 12 dimensions have a high level of accuracy while also maintaining a very fast processing time. Once all features associated with a given request are converted into vectors, they are used as input to one or more machine learning models in a suitable format for training the model or for classification of the associated network request.

[0065] In step 312, the security server device 102 optionally inputs the reduced feature set into one or more unsupervised machine learning models. In step 314, the security server device 102 optionally inputs the reduced feature set into one or more semi-supervised machine learning models. In step 316, the security server device 102 optionally inputs the reduced feature set into one or more supervised machine learning models. In various embodiments, the security server device 102 can train and use one or more different machine learning models to classify new requests as having been initiated by a robot program or a human, for example, a random forest model, a logarithmic regression model, an autoencoder model, an autoencoder with a neural network model (i.e., the output of the autoencoder model is used as the input of the neural network model), an autoencoder with a k-means clustering algorithm, an autoencoder with a k-nearest neighbor algorithm, an autoencoder with a random forest model, a convolutional neural network with an autoencoder model, a support vector machine, a feedforward neural network, and / or any other suitable machine learning model. It is desirable to use a model that maximizes recall, which is a measure of how many bots are correctly recalled, proportional to the number of true positives, and inversely proportional to the number of false negatives. In some examples, to maximize recall, an autoencoder with a random forest can be used to maximize recall for supervised learning, while for unsupervised learning, an autoencoder with a mean clustering algorithm can be used to maximize recall. As will be appreciated by those skilled in the art, each and any of these different types of machine learning models can be trained to predict whether a vector representing features of a newly received web request represents a web request initiated by a bot or a human based on the training that the model receives using a training data set (i.e., a set of vectors representing historical web requests and whether they are bots or humans).

[0066] For example, a K-means clustering algorithm can be trained by taking as input a plurality of vectors, each of which represents a reduced feature set of corresponding historical network requests known to be initiated by a robot program or a human. The K-means clustering algorithm includes drawing all vectors and then clustering them into a selected number of clusters, such as 100 clusters, wherein each cluster represents a plurality of network requests that are most similar to each other based on the feature information of the network requests represented by the vectors. Each cluster can then be labeled as a robot program or a human based on the labels of the majority of vectors within the cluster. For example, if more than half of the vectors in a given cluster are labeled as being associated with a network request initiated by a human, the cluster will be labeled as "human", and a cluster consisting of most vectors labeled as "robot program" will be labeled as "robot program". These labeled clusters represent a trained model, and now when a new network request is received, it can be predicted at step 324 whether the new request is initiated by a robot program or a human.

[0067] For example, when a new network request is received at step 310, it can be converted into a vector representing the reduced feature set as described above in steps 302, 304, 306, 308, and 310, and the vector representing the new network request can be input into one or more trained machine learning models to predict whether the request is initiated by a robot program or a human. For example, in the case of the above-mentioned K-means clustering algorithm, the vector can be plotted, and it can be determined whether the vector is located in a cluster labeled "human" or "bot program". If the security server device 102 determines that the new vector is located in a cluster labeled "human", the security server device 102 will output a prediction that the newly received request is initiated by a human, or if the new vector is located in a cluster that has been labeled as "bot program", the security server device 102 will output a prediction that the newly received request is initiated by a robot program.

[0068] In some examples, the machine learning model that makes the prediction can be an integrated model composed of multiple different machine learning models, which have been trained using a reduced feature set (represented as a vector) of historical network request data. In this case, the vector representation of the reduced feature set of the newly received network request can be provided as the input of each basic model of the integrated model, and each basic model will output its own prediction about whether the new request is initiated by a robot program or a human. The integrated model can determine the final prediction by a majority vote on all basic models. For example, if the integrated model consists of 3 basic models, each of which outputs "robot program" or "human", the integrated model will count the results, and if 2 or more basic models predict that the new request is initiated by a "robot program", the integrated model will output the prediction that the new request is initiated by a "robot program", otherwise, the integrated model will output the prediction that the new request is initiated by a "human". It is conceivable that any number or type of machine learning models described herein or otherwise suitable for making such predictions can be used in the integrated model. By utilizing many different types of machine learning models, the accuracy of the system and / or the confidence level of the generated predictions can be improved. For example, if the system uses an ensemble model consisting of 7 different base models that all generate the same output (e.g., “bot”), the system can have a high confidence that the prediction is correct, rather than, for example, only 4 models predicting “bot” and 3 models predicting “human”.

[0069] As will be appreciated by those skilled in the art, a random forest model is a tree structure that can be trained based on a target variable (i.e., whether the request originated from a robot program or a human), and the tree structure can be trained using supervised learning to determine whether a new request is initiated by a robot program or a human. In addition, as will be appreciated, an autoencoder is an unsupervised model that can determine a threshold that can be used to determine whether the features of the new request (as represented by the vector) represent a robot program or a human based on whether the features of the new request exceed the determined threshold. In various examples, one or more of different types of models can be used to make predictions, such as transformer models (e.g., Bert, RoBERTa, etc.), pre-trained models, and generative models.

[0070] As previously described, in step 318 , secure server device 102 receives a network request from client device 106 requesting access to a web page hosted by server device 104 .

[0071] In step 320, the secure server device 102 determines whether the request received in step 318 originated from a human operating the client device 106, or whether it originated from a bot running on the client device 106 (or posing as the client device 106). For example, if, when the vector representing the reduced set of features of the new request is input into one or more trained machine learning models, the one or more trained machine learning models generate an output that provides a prediction that the network request was initiated by a bot, then the secure server device 102 will determine that the network request was initiated by a bot, or similarly, if the output predicts that the network request was initiated by a human, then the secure server device 102 will determine that the network request was initiated by a human.

[0072] In step 322, in response to determining that the network request was initiated by a human, the secure server device 102 allows the client device 106 to access the requested web page hosted by the server device 106. In some embodiments, this can be achieved by forwarding the received request to the target server device 104 to allow the server device to respond to the request. In the event that the target server device 104 receives the request from the client device 106 and forwards it to the secure server 102, allowing the client device 106 to access the requested web page hosted by the server device 106 may include sending a message to the server device 104 indicating that the request is allowed, which may cause the server device 104 to respond to the request. In other words, in this case, once the initial request is received, the server device 104 may forward it to the secure server device 102 and wait for a response from the secure device 102 indicating that the server device 104 may allow the request before responding to the client device 106.

[0073] In step 324, in response to determining that the network request was initiated by a bot, the secure server device 102 denies the client device 106 access to the requested web page hosted by the server device 106. For example, the secure server device 102 may block communications from the server device 104 from proceeding to the client device 106, or may send a communication to the server device 104 instructing it not to provide the client device with access to the requested web page.

[0074] refer to Figure 5 , a flowchart of an example method 500 for detecting a robot program using a machine learning model is shown. In step 502 of this example, the secure server device 102 of the secure server system 100 receives a request for a web page from the client device 106 to the server device 104. For example, as described above with respect to Figure 1 , Figure 2 and Figure 3 As described, the secure server device 102 may receive a request for a web page from the client device 106. The request will have an associated log containing the original text functionality.

[0075] In step 504 , the secure server device 102 generates a feature set based on the log.

[0076] According to some examples, generating a feature set based on a log may include the following: Figure 3 Extracting original text features from the log as described in step 302, as described above Figure 3 The step 304 described above generates digital features and classification features based on the original text features, as described above with respect to Figure 3 The step 306 of generating a reduced number of features based on the numerical features and the categorical features, and the step 307 of generating a reduced number of features based on the numerical features and the categorical features, and the step 308 of generating a reduced number of features based on the numerical features and the categorical features, and Figure 3 The scaled feature set is input to the autoencoder model as described in step 310, and the autoencoder model outputs a reduced feature set.

[0077] In some examples, generating a reduced feature set may include converting categorical features to numerical features, generating a scaled feature set by applying feature scaling to the numerical features and the numerical features, and inputting the scaled feature set to an autoencoder model that outputs the reduced feature set. One or more of the categorical features may be converted to numerical features using a one-hot encoding. In addition, one or more of the categorical features may be converted to numerical features by generating word embeddings using a transformer such as word2vec or BERT. In some examples, the reduced number of features may be selected to be a set of 12 features represented as vectors, although a different number of reduced features may be selected in other examples.

[0078] In some examples, the step of scaling features to create scaled features may involve scaling only one, some, or all of the digital features and numerical features. In some examples, the step of scaling features to create scaled features may be excluded, and the unscaled features may be used to create a reduced feature set.

[0079] In step 506, the secure server device 102 submits the feature set as input to the machine learning model. In the case where the reduced feature set is generated from historical network request data, the input features can be used to train one or more machine learning models, and in the case where the reduced feature set is generated from newly received network requests, the input features can be used to generate a prediction about whether the newly received network request is initiated by a human or a robot program.

[0080] In some examples, the machine learning model may be an integrated model with an output determined by a majority vote. The integrated model may include one or more supervised machine learning models, unsupervised machine learning models, and / or semi-supervised machine learning models. According to some embodiments, it may be beneficial to select the type and / or number of models in the integrated model based on the amount of available computing resources, because some models require a heavier computational load than other models. For example, where limited computing resources are available, the integrated model may include an autoencoder with a simple machine learning model. In some embodiments, if a large amount of computing resources are available, the integrated model may include a transformer model (e.g., a BERT model) or any other pre-trained model. It will be understood that these are merely examples, and it is conceivable that any combination of suitable models may be used in an integrated model according to various embodiments.

[0081] In some examples, the first machine learning model of the ensemble model can be a K-means clustering algorithm that is applied to a set of training data features represented by a vector to generate a plurality of clusters. Figure 3 Each feature of the reduced feature set generated in step 310 can be represented as a vector that can be plotted. The security server device 102 can generate a training data feature set by obtaining historical logs associated with previous requests, and using these historical logs to generate a reduced feature set according to the techniques described herein. For the previous requests, it is known whether each request was generated by a robot program or a human. For each of the multiple clusters generated by the K-means clustering algorithm, when most of the vectors within the cluster have labels corresponding to humans, the security server device 102 can determine that the cluster represents a human, and when most of the vectors within the cluster have labels corresponding to robots, it can be determined that the cluster represents a robot program.

[0082] In step 508 , in response to the output of the machine learning model indicating that the request for the web page received from the client device was initiated by a bot, the secure server device 102 denies the request.

[0083] According to some embodiments, method 500 may further include, in response to an output of the machine learning model indicating that the request for the web page received from the client was initiated by a human, secure server device 102 allowing the request.

[0084] Utilizing this technology, automated bot detection can be achieved in near real-time. The system can advantageously encapsulate large amounts of feature information, which allows it to utilize rule-based approaches to discover subtleties in the data that are difficult to detect or accurately classify. The disclosed feature extraction and reduction process allows the system to retain most of the feature information while not incurring a large and time-consuming processing load. As a result, the system can process and predict newly received network requests in near real-time with little to no latency, allowing it to filter out requests made by bots without significantly negatively impacting the user experience of legitimate users. Additionally, the use of an integrated machine learning model can improve the accuracy and confidence of the predictions made by the system, significantly reducing the number of false positives, compared to other approaches currently in use.

[0085] Having described the basic concepts of the present invention, it will be very clear to those skilled in the art that the above detailed disclosure is presented by way of example only and not by way of limitation. Various changes, improvements and modifications will occur, and these changes, improvements and modifications are directed to those skilled in the art, although not expressly stated herein. These changes, improvements and modifications are intended to be proposed herein and are within the spirit and scope of the present invention. In addition, the order of the listed processing elements or sequences, or the use of numbers, letters or other names therefor, is not intended to limit the claimed processes to any order, unless otherwise specified in the claims. Therefore, the present invention is limited only by the following claims and their equivalents.

Claims

1. A method for implementing a secure server system, the secure server system comprising one or more secure server devices, server devices or client devices, the method comprising: receiving a request for a web page from a client device to a server device, the request having one or more associated logs; generating a feature set based on the one or more associated logs; Submitting the feature set as input to a machine learning model; as well as The request is denied in response to an output of the machine learning model, the output indicating that the request for the web page from the client device was initiated by a robot program.

2. The method according to claim 1, wherein: Generating the feature set includes: extracting raw text features from the one or more associated logs; Generating numerical features and classification features based on the original text features; and A reduced number of features is generated based on the numerical features and the categorical features.

3. The method according to claim 2, wherein: Generating a reduced feature set includes: Convert the categorical features into numerical features by using one-hot encoding or using a transformer to generate word embeddings; generating a scaled feature set by applying feature scaling to the digital features and the numerical features; and The scaled feature set is input to an autoencoder model, which outputs the reduced feature set.

4. The method according to claim 1, wherein: The machine learning model includes an ensemble model having an output determined by majority voting, the ensemble model including one or more of: Supervised machine learning models; Unsupervised machine learning models; or Semi-supervised machine learning models.

5. The method according to claim 4, wherein: The first machine learning model of the ensemble model is trained by the following steps: Applying a K-means clustering algorithm to the set of training data features represented by the vector to generate a plurality of clusters; as well as For each cluster of the plurality of clusters, the cluster is determined to represent a human when a majority of vectors within the cluster have labels corresponding to humans, and the cluster is determined to represent a robot when a majority of vectors within the cluster have labels corresponding to robots.

6. A secure server device comprising a memory and one or more processors, the memory including programming instructions stored thereon, the one or more processors being configured to be capable of executing the stored programming instructions to: receiving a request for a web page from a client device to a server device, the request having one or more associated logs; generating a feature set based on the one or more associated logs; Submitting the feature set as input to a machine learning model; as well as The request is denied in response to an output of the machine learning model, the output indicating that the request for the web page from the client device was initiated by a robot program.

7. The security server device according to claim 6, wherein: To generate the feature set, the processor is further configured to execute stored programming instructions to: extracting raw text features from the one or more associated logs; Generate numerical features and classification features based on the original text features; as well as A reduced number of features is generated based on the numerical features and the categorical features.

8. The security server device according to claim 7, wherein: To generate the reduced feature set, the processor is further configured to execute stored programming instructions to: Convert the categorical features into numerical features by using one-hot encoding or using a transformer to generate word embeddings; generating a scaled feature set by applying feature scaling to the digital features and the numerical features; as well as The scaled feature set is input to an autoencoder model, which outputs the reduced feature set.

9. The security server device according to claim 6, wherein: The machine learning model includes an ensemble model having an output determined by majority voting, the ensemble model including one or more of: Supervised machine learning models; Unsupervised machine learning models; or Semi-supervised machine learning models.

10. The security server device according to claim 9, wherein: The processor is further configured to execute the stored programming instructions to train the first machine learning of the integrated model by: Applying a K-means clustering algorithm to the set of training data features represented by the vector to generate a plurality of clusters; as well as For each cluster of the plurality of clusters, the cluster is determined to represent a human when a majority of vectors within the cluster have labels corresponding to humans, and the cluster is determined to represent a robot when a majority of vectors within the cluster have labels corresponding to robots.

11. A non-transitory computer readable medium having stored thereon instructions including executable code which, when executed by one or more processors, causes the processors to: receiving a request for a web page from a client device to a server device, the request having one or more associated logs; generating a feature set based on the one or more associated logs; Submitting the feature set as input to a machine learning model; as well as The request is denied in response to an output of the machine learning model, the output indicating that the request for the web page from the client device was initiated by a robot program.

12. The non-transitory computer readable medium of claim 11, wherein: To generate the feature set, the executable code, when executed by the processor, further causes the processor to: extracting raw text features from the one or more associated logs; Generate numerical features and classification features based on the original text features; as well as A reduced number of features is generated based on the numerical features and the categorical features.

13. The non-transitory computer readable medium of claim 12, wherein: To generate the reduced feature set, the executable code, when executed by the processor, further causes the processor to: Convert the categorical features into numerical features by using one-hot encoding or using a transformer to generate word embeddings; generating a scaled feature set by applying feature scaling to the digital features and the numerical features; as well as The scaled feature set is input to an autoencoder model, which outputs the reduced feature set.

14. The non-transitory computer readable medium of claim 11, wherein: The machine learning model includes an ensemble model having an output determined by majority voting, the ensemble model including one or more of: Supervised machine learning models; Unsupervised machine learning models; or Semi-supervised machine learning models.

15. The non-transitory computer readable medium of claim 14, wherein: The executable code, when executed by the processor, further causes the processor to train a first machine learning model of the integrated model by: Applying a K-means clustering algorithm to the set of training data features represented by the vector to generate a plurality of clusters; as well as For each cluster of the plurality of clusters, the cluster is determined to represent a human when a majority of vectors within the cluster have labels corresponding to humans, and the cluster is determined to represent a robot when a majority of vectors within the cluster have labels corresponding to robots.

16. A security system comprising one or more security server devices, server devices, or clients having a memory and one or more processors, the memory including programming instructions stored thereon, the one or more processors configured to be able to execute the stored programming instructions to: receiving a request for a web page from a client device to a server device, the request having one or more associated logs; generating a feature set based on the one or more associated logs; Submitting the feature set as input to a machine learning model; as well as The request is denied in response to an output of the machine learning model, the output indicating that the request for the web page from the client device was initiated by a robot program.

17. The security system of claim 16, wherein: To generate the feature set, the processor is further configured to execute stored programming instructions to: extracting raw text features from the one or more associated logs; Generate numerical features and classification features based on the original text features; as well as A reduced number of features is generated based on the numerical features and the categorical features.

18. The security system of claim 17, wherein: To generate the reduced feature set, the processor is further configured to execute stored programming instructions to: Convert the categorical features into numerical features by using one-hot encoding or using a transformer to generate word embeddings; generating a scaled feature set by applying feature scaling to the digital features and the numerical features; as well as The scaled feature set is input to an autoencoder model, which outputs the reduced feature set.

19. The security system of claim 16, wherein: The machine learning model includes an ensemble model having an output determined by majority voting, the ensemble model including one or more of: Supervised machine learning models; Unsupervised machine learning models; or Semi-supervised machine learning models.

20. The security system according to claim 19, characterized in that The processor is further configured to execute the stored programming instructions to train the first machine learning model of the ensemble model by: Applying a K-means clustering algorithm to the set of training data features represented by the vector to generate a plurality of clusters; as well as For each cluster of the plurality of clusters, the cluster is determined to represent a human when a majority of vectors within the cluster have labels corresponding to humans, and the cluster is determined to represent a robot when a majority of vectors within the cluster have labels corresponding to robots.

Citation Information

Cited By

  • Data security monitoring method and system based on USB network communication

    CN119945763A