Phishing attack detection

Abstract

A computerized method for analyzing a subject URL to determine whether the subject URL is associated with a phishing attack is disclosed. The method includes steps of detecting keypoints within a screenshot of a webpage corresponding to the subject URL and determining a set of confidences based on an analysis of the detected keypoints with a model. Each confidence within the set of confidences is assigned to feature vector within a set of training feature vectors representing a training set of URLs used in generating the model. The method comprises performing an image comparison between the screenshot and a screenshot corresponding to a feature vector within the set of training feature vectors, the feature vector being assigned a highest confidence. Responsive to determining the image comparison result exceeds a predefined threshold, transmitting an alert indicating that the subject URL is associated with the phishing attack.

Description

FIELD

Embodiments of the disclosure relate to the field of cybersecurity. More specifically, embodiments of the disclosure relate to a system and method for detecting phishing attacks.

GENERAL BACKGROUND

Phishing is a growing problem on the internet. Phishing is the attempt to obtain sensitive information from targets by disguising requests as legitimate. A phishing attack can entail the transmission of an electronic communication, such as an email, to one or more recipients that purports to be from a known institution, such as a bank or credit card company, and seems to have a legitimate intention; however, the email is actually intended to deceive the recipient into sharing its sensitive information. Often the email draws the recipient to a counterfeit version of the institution's webpage designed to elicit the sensitive information, such as the recipient's username, password, etc.

For example, a malware author may transmit an email to a recipient purporting to be from a financial institution and asserting that a password change is required to maintain access to the recipient's account. The email includes a Uniform Resource Locator (URL) that directs the recipient to a counterfeit version of the institution's website requesting the recipient to enter sensitive information in a displayed form in order to change the recipient's password. Neither the email nor the URL are associated with the actual financial institution or its genuine website, although the entail and the counterfeit website may have an official “look and feel” and imitate a genuine email and website of the institution. The phishing attack is completed when the recipient of the email enters and submits sensitive information to the website, which is then delivered to the malware author.

Current solutions for phishing detection include textual search and analysis of entails and a displayed webpage. However, such solutions have a plurality of drawbacks and too often fail to detect phishing attacks. As a first drawback, current textual search-based phishing detection systems may be unable to determine whether a website to which a URL resolves is a phishing website due to an insufficient amount of text displayed on the website. Specifically, when a website contains insufficient text, a textual search analysis may not have enough data to allow an accurate analysis. As a second drawback, current textual search-based current solutions may be unable to perform an analysis on the website to which the URL resolves due to the text of the website being contained within one or more images (e.g., bitmaps, jpegs, etc.), which cannot be processed using a textual search-based analysis. As yet another drawback, current textual search-based solutions may be unable to perform the necessary textual search and analysis in many languages due to an insufficient corpus of data; thus, providing a lackluster solution with respect to the global nature of attacks on businesses today (e.g., the large number of characters in Asian languages makes a textual search-based analysis difficult). Thus, a new phishing detection technique is needed to more efficiently, efficaciously, and reliably detect phishing cybersecurity attacks (“cyberattacks”) of this type.

DESCRIPTION OF THE DRAWINGS

Embodiments of the disclosure are illustrated by way of example and not by way of limitation in the figures of the accompanying drawings, in which like references indicate similar elements and in which:

FIG. 1 is an exemplary block diagram of a logic flow during performance of a training process by a phishing detection and analysis system (PDAS) according to an embodiment of the invention;

FIG. 2 is an exemplary block diagram of a logic flow during performance of a detection process by a phishing detection and analysis system (PDAS) according to an embodiment of the invention;

FIGS. 3A-3B provide a flowchart illustrating an exemplary method for analyzing a URL by the PDAS of FIG. 4 to determine whether the URL is associated with a phishing attack; and

FIG. 4 is an exemplary embodiment of a logical representation of the phishing detection and analysis system of FIG. 1.

DETAILED DESCRIPTION

I. Overview Summary

Embodiments of systems and methods for detecting phishing attacks are described. The phishing, detection and analysis system (PDAS) is configured to detect a phishing attack through the use of computer vision techniques that leverage a graphic representation (i.e., the representation expressing the “look and feel”) of a webpage to determine whether the webpage is attempting to mimic a legitimate webpage. Some specific embodiments attempt to determine whether a webpage is attempting to mimic a webpage through which a user enters sensitive information, such as login credentials, for purposes of stealing such sensitive information.

As a general overview, the PDAS described herein includes (i) a training process and (ii) a detection process. The training process generates a machine learning model (“model”), the model including a set of correlation rules, that is used in the detection process. The detection process receives a URL and analyzes the URL based on the model to make a determination as to whether the URL is part of a phishing cyberattack.

The training process involves the generation of a model using machine learning techniques, the model representing a categorization of a training set of URLs into one or more webpage families, the training set of URLs known to be associated with genuine (non-phishing) websites (in some embodiments, known phishing URLs may be provided to improve the model). The training process includes retrieval of a screenshot associated with each URL of the training set of URLs, processing of each screenshot to (i) detect a set of keypoints and (ii) generate feature vectors corresponding to the detected set of keypoints. A feature may be interpreted as a keypoint and corresponding keypoint descriptors, e.g., parameters of the keypoint enabling identification of the keypoint and its location within a screenshot. After generation of the feature vectors, the feature vectors are labeled based on a known webpage family and the model is generated, the model being a digitized representation of the correlation or the feature vectors corresponding to the URLs within the training set of URLs. In some embodiments, the webpage families may represent a set of URL domains.

The model is generated for use in a detection process, discussed below, to identify the keypoints within a feature vector corresponding to at least one of the URLs within the training set to which a screenshot under analysis most closely correlates. A URL under analysis, a webpage under analysis or a screenshot under analysis may be referred to as a subject URL, a subject webpage and a subject screenshot, respectively. More specifically, the analysis of a subject screenshot using the model results in a set of confidences, each confidence corresponding to a feature vector corresponding to a URL within the training set. The highest confidence indicating the highest correlation between the subject screenshot and the feature vector corresponding to a URL within the training set. The correlation is based on the generated feature vectors of (i) the subject screenshot, and (ii) the screenshot(s) corresponding to the URLs within the training set.

A webpage family refers to a set of webpages associated with a particular company or other organization that shares a webpage design system reflecting the branding and design elements (including logos, layout and visual landscape (e.g., color, contrast, etc.)) of the organization to provide site visitors with a consistent and recognizable visual experience. The webpage family may include one or more webpage members, though generally it will include plural webpages for each family of most interest in the practice of the invention, i.e., those generally used in phishing attacks. The members of the webpage family may and generally will differ from one another, for example in message content (such as textual content, user-interactive elements, and pictorial elements) and even graphical elements; hence, they will generally exhibit variations called “variances” across the family. Often the webpage family may share a domain name and/or other URL components, but that is not the necessary and sufficient determinant of membership in the family since similar domain names and other URL components may mislead visitors as to the “owner” of the website. However, domain name sharing may be used as one aspect in determining family membership. Accordingly, the invention may use computer vision to determine family membership.

It should be noted that any variances will be such that webpages within a webpage family have a consistent layout and visual landscape. As discussed below, webpages belonging to a particular owner (e.g., a single company sharing a domain name) may differ in some aspects of the layout and visual landscape (e.g., differ in number of input types and/or input forms, for example, textboxes in a first webpage and radio dials in a second webpage). In such an embodiment, the webpages may be divided into two webpage families based on detected keypoints with both webpage families being linked to the single owner for the detection process.

In particular, during the training process, for each screenshot corresponding to a URL within the training set of URLs, a set of keypoints is detected. Each keypoint detected within the screenshot identifies a “point of interest.” Points of interest reflect regions of an image which are observable. A “keypoint” may be defined as an image region, e.g., a set of pixels, within the screenshot (e.g., any shaped-region such as a circular-shaped region). Known keypoint detection techniques such as rule sets that detect keypoints based on pixel density. Scale-Invariant Feature Transform (SIFT), Features from Accelerated Segment Test (FAST) and/or Binary Robust Invariant Scale Keypoints (BRISK) may be utilized to detect keypoints.

Subsequently, keypoint descriptors corresponding to the detected keypoints are determined. A keypoint descriptor may include a set of one or more parameters that describe the keypoint such as keypoint center coordinates x and y relative to the screenshot, a scale (e.g., being a radius of a circular image region, when applicable), and/or an orientation determined by the gradient of the pixel greyscale within the keypoint. The parameters enable the generated model to be invariant to orientation or scale differences in the subject screenshot. Each keypoint descriptor provides the ability to reliably identify a keypoint within a sereenshot. The keypoints and/or keypoint descriptors of the processed screenshot may be stored in a data store (e.g. a database).

More specifically, the training process begins upon receipt of a list of labeled URLs. The list of URLs resolves to webpages that are generally known to be typically targeted for use in phishing attacks such as login webpages of banks or other online accounts of well-known companies such as Apple iTunes®, Spotify®, Netflix®, etc. The list of URLs (wherein the set of URLs is referred to as the “training set”) may be obtained or updated periodically or aperiodically for training of the PDAS classifier logic so as to reflect commonly visited websites. The PDAS may obtain a plurality of screenshots corresponding to a webpage associated with a URL, each such screenshot corresponding to a browser/operating system combination. A screenshot of the webpage to which each URL resolves is obtained by the PDAS, which then utilizes computer vision techniques to detect keypoints, determine keypoint descriptors and generate a feature vector for each screenshot. A feature may be interpreted as a set of keypoints and their corresponding keypoint descriptors that indicate a point of interest within the screenshot (e.g., a logo or a portion thereof). A feature vector includes the plurality of features detected within a screenshot. In some embodiments, methods other than a vector may be used to store and organize the features, such as a matrix or other data structure within memory. As an example, the features may be distinctive aspects of a webpage that enable the PDAS, during the detection process, to determine whether a subject webpage is attempting to mimic a webpage included in the training set. The features of each screenshot are inserted into separate vectors and labeled according to the webpage family to which the URL corresponding to the feature vector belongs. The plurality of labeled feature vectors are then used by the PDAS to generate a model using machine learning. As mentioned above, the model is a digitized representation of the correlation of the feature vectors corresponding to the URLs within the training set of URLs. More specifically, the model may be a collection of keypoints corresponding to the training set described as a function implemented programmatically in code where the function is tuned using the training set and keypoints selected (digital sampling) by machine learning techniques. Once the function is generated and tuned, it can be applied to other (unknown) image keypoint sets during the detection process to analyze a subject screenshot. One example of the model may be a hyperplane.

During the machine learning in generating the model, the training typically involves a plurality of webpages from the same webpage family and the system is trained to recognize family membership through identifying keypoints shared (high correlation) across those “labeled” webpages. The detection of those keypoints, including their location within the corresponding webpage, is key to later classification of an “unlabeled” webpage as being a member of the family to which it purports to be a member (through visual similarity). After all, a webpage can have a large number of keypoints (e.g., hundreds or thousands for example), and the training, in some embodiments, may go to selection of the keypoints that together are unique to the corresponding screenshot and can be used for training analysis to accurately identify members of the labeled webpage families and later, after training, with respect to the unlabeled webpages. Moreover, the keypoints can be selected so as to capture the common branding, and design elements of a webpage family rather than variances across the members of the family so that membership in the family can be accurately determined with minimal or no false positives or false negatives.

In some embodiments, the generation of the model involves detecting keypoints within each screenshot corresponding to the URLs within the training set. The detected keypoints are then used to extract features within each screenshot and generate a feature vector for each screenshot. The detection of keypoints and generation of feature vectors are performed using computer vision techniques such as those mentioned above. Each feature vector within the set of feature vectors is labeled according to a webpage family to which it belongs and the set of feature vectors are then used to generate the model using machine learning techniques. The machine learning techniques may include, but are not limited or restricted to, the generation of (i) support vector machines (SVMs), (ii) distribution functions such as a naïve bayes classifier, (iii) a K-nearest neighbor pattern detection algorithm, and/or (iv) a decision tree.

Machine learning techniques rely on a machine learning model, which is executable by a processor to implement a computerized approach to learning from and making predictions based on data sets. These include stored known (labelled) data sets used to train (or tune) the machine learning model in reaching conclusions regarding the data sets, e.g., classify and verify the classification by comparison with the labels. The data sets also include one or more unknown (unlabeled) data sets, for which the machine learning model is to reach conclusions (classify) by applying its acquired “learning” with, if trained properly, a high degree of confidence.

In recent years, machine learning technology has seen development and application diverse fields (such as computer vision) of a great many widely-used, executable machine learning computer programs. These implement any of a variety of different machine learning techniques commonly referred to as machine learning functions or algorithms. For purposes of this invention, the details of the machine learning functions and their implementation in machine learning software programs need not be described here though those of skill in the art would be readily able to choose from many commercially or publicly available (“open source”) alternatives to implement this invention, as mentioned above.

Herein, machine learning is used to recognize membership and non-membership in a family of webpages as a strong indication of phishing attacks. More specifically the machine learning model represents the correlation of the feature vectors corresponding to the screenshots based on data sets associated with the screenshots. It can be understood that each data set is a collection of images information, which can be computationally processed pursuant to or by a machine learning function implemented programmatically. The machine learning function operates on keypoints expressed as keypoint descriptors and is generated or tuned using the training set to both select (digital sample) keypoints and use their descriptors, formed into feature vectors, in classifying the webpage images. The machine learning function itself is generated and tuned during the training phase to classify the data sets, and then is stored in association with the stored representation of the screenshot images in memory for later use. Once the function is generated and tuned, it can be applied to other (unknown) image data sets for their classification.

The detection process involves receipt of a URL for analysis to determine whether the URL is associated with a phishing cyberattack (“subject URL”). The detection process involves retrieval of a subject screenshot corresponding to the subject URL and detection of the keypoints of the subject screenshot. Keypoint descriptors are then generated that correspond to the detected keypoints. The detection process includes an analysis of the generated keypoint descriptors based on the model generated during the training process to determine a correlation between (i) the keypoints corresponding to the subject URL and (ii) keypoints corresponding to the URLs within the training set that have been categorized into webpage families. One or more screenshots corresponding to the webpage family being the most highly correlated to the keypoints of the subject screenshot is selected. In some embodiments there may be a plurality of webpage families closely correlated with the keypoints of the subject screenshot; the remainder of the system would be processed relative to that plurality. The keypoints of the subject screenshot are compared, via known image comparison techniques in the field of computer vision, to the keypoints of the selected screenshot. In some embodiments, a plurality of screenshots may be associated with a single webpage family as mentioned above, in which case an image comparison would be performed between the subject screenshot and each screenshot of the plurality of screenshots corresponding to the most highly correlated webpage family. If this image comparison exceeds a threshold, the subject URL is determined to be associated with a phishing cyberattack. Upon determination of the subject URL being associated with a phishing cvberattack, an alert and/or a report is issued to an administrator or a cybersecurity analyst.

More particularly, the detection process includes the (i) generation of a subject screenshot of a webpage retrieved from a subject URL, (ii) processing the subject screenshot to identify a set of keypoints, (iii) correlating the set of keypoints to a set of known benign or known phishing pages using the model, and (iv) if the correlation exceeds a threshold, classifying the subject URL as part of a phishing cyberattack. In some embodiments, the retrieval of the subject screenshot, or content associated therewith, may be via a centrally located system using an internet browser as discussed below or via accessing a data caching system that has stored therein previously captured screenshots. In some embodiments, the PDAS performs a pre-filtering process, which may include static scanning of the subject URL (e.g., blacklist or white list analysis, namely heuristics, exploit signature checks and/or vulnerability signature checks for example). If the subject URL is not determined to be either malicious (i.e. related to a phishing attack) or benign based on the static scanning, a subject screenshot of the subject webpage to which the subject URL resolves is obtained by the PDAS.

The screenshot may result from the processing of the webpage (the webpage associated with the URL) based on the characteristics (e.g. selected internet browser applications, operating systems, etc.). A logic module of the PDAS utilizes computer vision techniques to detect keypoints within the subject screenshot and generates a feature vector based on the detected keypoints in the same manner as discussed above with respect to the training process. The feature vector of the subject screenshot is analyzed using the model to determine a set of confidences, with each confidence corresponding to a separate labeled feature vector corresponding to the training set; thus, providing an understanding of the webpage family having the highest confidence (e.g., which webpage family, and specifically, which webpage, is most likely being mimicked by the subject screenshot). For example, a first confidence corresponds to the likelihood a screenshot within a first webpage family is being mimicked and a n^th confidence corresponds to the likelihood a screenshot within a n^th webpage family is being mimicked. A screenshot of at least a first webpage of the webpage family having the highest confidence is then used in an image comparison operation with the subject screenshot. The image comparison may include a comparison of detected keypoints of the subject screenshot and the webpage(s) of the webpage family having the highest confidence. When the image comparison results in a match above a predefined threshold, the PDAS determines that the subject webpage and the subject URL are part of a phishing attack.

Specifically, in contrast to alternative phishing detection systems that may merely perform image comparisons—e.g., comparisons of detected keypoints in a brute force manner between a webpage under analysis and hundreds or thousands of webpage screenshots—the disclosure provides novel systems and methods that enable a detection process involving computer vision techniques to avoid performing image comparisons between a subject webpage and hundreds or thousands or webpage screenshots while providing a determination that limits false positives and false negatives through the use of a model trained using the detected keypoints of hundreds or thousands of webpage screenshots prior to the detection process. Specifically, a brute force image comparison of hundreds or thousands of webpage screenshots within a training set to the subject webpage is avoided by generating a model that represents the detected keypoints of each of the webpages within the training set and utilizing the model to obtain a set of confidences, each confidence indicating the likelihood a webpage of a webpage family is being mimicked by the subject webpage.

To achieve higher efficiencies during analysis compared to alternative systems, the systems and methods described below only conduct an image comparison with respect to the webpage(s) corresponding to the webpage family having the highest confidence of visual similarity to the subject screenshot. Thus, an image comparison of (i) the subject screenshot and (ii) the webpage(s) corresponding to the webpage family having the highest confidence is more efficient with time and resources than a brute force method of performing image comparisons between the subject screenshot and hundreds or thousands of screenshots. Thus, with respect to the detection process, by performing the feature generation and classification processes discussed in detail below prior to the image comparison, the disclosure provides systems and methods for detecting phishing URLs and webpages that efficiently use resources and save processing time previously needed to perform such a determination.

II. Terminology

In the following description, certain terminology is used to describe various features of the invention. For example, each of the tents “logic” and “component” may be representative of hardware, firmware or software that is configured to perform one or more functions. As hardware, the term logic (or component) may include circuitry having data processing and/or storage functionality. Examples of such circuitry may include, but are not limited or restricted to a hardware processor (e.g., microprocessor, one or more processor cores, a digital signal processor, a programmable gate array, a microcontroller, an application specific integrated circuit “ASIC”, etc.), a semiconductor memory, or combinatorial elements.

Additionally, or in the alternative, the logic (or component) may include software such as one or more processes, one or more instances, Application Programming Interface(s) (API), subroutine(s), function(s), applet(s), servlet(s), routine(s), source code, object code, shared library/dynamic link library (dll), or even one or more instructions. This software may be stored in any type of a suitable non-transitory storage medium, or transitory storage medium (e.g., electrical, optical, acoustical or other form of propagated signals such as carrier waves, infrared signals, or digital signals). Examples of a non-transitory storage medium may include, but are not limited or restricted to a programmable circuit; non-persistent storage such as volatile memory (e.g., any type of random access memory “RAM”); or persistent storage such as non-volatile memory (e.g., read-only memory “ROM”, power-backed RAM, flash memory, phase-change memory, etc.), a solid-state drive, hard disk drive, an optical disc drive, or a portable memory device. As firmware, the logic (or component) may be stored in persistent storage.

Herein, a “communication” generally refers to related data that is received, transmitted, or exchanged within a communication session. The data may include a plurality of packets, where a “packet” broadly refers to a series of bits or bytes having a prescribed format. Alternatively, the data may include a collection of data that may take the form of an individual or a number of packets carrying related payloads, e.g., a single webpage received over a network.

The term “computerized” generally represents that any corresponding operations are conducted by hardware in combination with software and/or firmware.

According to one embodiment of the disclosure, the term “malware” may be broadly construed as any code, communication or activity that initiates or furthers a cyberattack. Malware may prompt or cause unauthorized, anomalous, unintended and/or unwanted behaviors or operations constituting a security compromise of information infrastructure. For instance, malware may correspond to a type of malicious computer code that, as an illustrative example, executes an exploit to take advantage of a vulnerability in a network, network device or software, for example, to gain unauthorized access, harm or co-opt operation of a network device or misappropriate, modify or delete data. Alternatively, as another illustrative example, malware may correspond to information (e.g., executable code, script(s), data, command(s), etc.) that is designed to cause a network device to experience anomalous (unexpected or undesirable) behaviors. The anomalous behaviors may include a communication-based anomaly or an execution-based anomaly, which, for example, could (1) alter the functionality of a network device executing application software in an atypical manner; (2) alter the functionality of the network device executing that application software without any malicious intent; and/or (3) provide unwanted functionality which may be generally acceptable in another context.

A “characteristic” includes data associated with an object under analysis that may be collected without execution of the object such as metadata associated with the object (e.g., size, name, path, grey scale, etc.) or content of the object (e.g., portions of code) without execution of the selected object.

The term “object” generally relates to content (or a reference to access such content) having a logical structure or organization that enables it to be classified for purposes of analysis for malware. The content may include an executable (e.g., an application, program, code segment, a script, dynamic link library “dll” or any file in a format that can be directly executed by a computer such as a file with an “.exe” extension, etc.), a non-executable (e.g., a storage file; any document such as a Portable Document Format “PDF” document; a word processing document such as Word® document; an electronic mail “email” message, web page, etc.), or simply a collection of related data. In one embodiment, an object may be a URL or list of URLs. The object may be retrieved from information in transit (e.g., one or more packets, one or more flows each being a plurality of related packets, etc.) or information at rest (e.g., data bytes from a storage medium).

The term “network device” may be construed as any electronic computing system with the capability of processing data and connecting to a network. Such a network may be a public network such as the Internet or a private network such as a wireless data telecommunication network, wide area network, a type of local area network (LAN), or a combination of networks. Examples of a network device may include, but are not limited or restricted to, an endpoint device (e.g., a laptop, a mobile phone, a tablet, a computer, etc.), a standalone appliance, a server, a router or other intermediary communication device, a firewall, etc.

The term “transmission medium” may be construed as a physical or logical communication path between two or more network devices or, between components within a network device. For instance, as a physical communication path, wired and/or wireless interconnects in the form of electrical wiring, optical fiber, cable, bus trace, or a wireless channel using radio frequency (RF) or infrared (IR), may be used. A logical communication path may simply represent a communication path between two or more network devices or between components within a network device.

Finally, the terms “or” and “and/or” as used herein are to be interpreted as inclusive or meaning any one or any combination. Therefore, “A, B or C” or “A, B and/or C” mean “any of the following: A; B; C; A and B: A and C; B and C; B and C.” An exception to this definition will occur only when a combination of elements, functions, steps or acts are in some way inherently mutually exclusive.

As this invention is susceptible to embodiments of many different forms, it is intended that the present disclosure is to be considered as an example of the principles of the invention and not intended to limit the invention to the specific embodiments shown and described.

III. General Architecture

Referring to FIG. 1, an exemplary block diagram of a logic flow during performance of a training process by a phishing detection and analysis system (PDAS) according to an embodiment of the invention is shown. The logic flow 100 of the training process illustrates the flow of data among logic modules of the PDAS 400, as seen in FIG. 4, in order to train a classifier 112 (e.g., CV classifier) for use in detecting URLs that resolve to phishing websites.

As an overview the training process involves receipt of a list of URLs for use the detection of phishing websites. The list of URLs may be based on internal analytics, a third-party source, or the like. The URLs included in the list of URLs may be either known, benign websites (e.g., those that are often used in carrying out phishing attacks) and/or known phishing websites. The screenshot of the website of each URL of the list of URLs is obtained by the content fetcher 104, and the feature generation logic 106 utilizes computer vision techniques to generate keypoint descriptors, also referred to as “features” as mentioned above, based on each screenshot, as discussed below. The features of each screenshot are inserted into separate vectors (“feature vectors”). The feature domain mapper 108 receives the feature vectors and labels each vector according to the website family of the screenshot to which the feature vector corresponds. The plurality of labeled feature vectors are then provided to the training module 110, which uses the plurality of feature vectors to generate a model that categorizes the plurality of labeled feature vectors. In one embodiment, the model represents a plurality of hyperplanes onto which the features of each vector may be categorized. The model is then provided to the classifier 112 for use in the detection process, as discussed with respect to at least FIG. 2.

More specifically, the training process of FIG. 1 begins with the content fetcher 104 receiving a list of URLs from a source. The source may include, but is not limited or restricted to a third-party website, an administrator and/or a cybersecurity analyst (hereinafter referred to as “an analyst”). Specifically, the list of URLs provided to the content fetcher 104 may be used to fetch data used in training the classifier 112. As one non-limiting example, a detection process may be focused on detecting phishing websites attempting to mimic banking websites, which are often used in phishing, attacks. However, the disclosure should not be limited to banking websites, instead, any website may be used by the PDAS 400. For purposes of clarity the examples discussed herein will involve banking websites.

Upon receiving the list of URLs, the content fetcher 104 obtains a screenshot of the website to which each URL provided resolves. The content fetcher 104 obtains a screenshot by utilization of an internet browser to access a URL to render the webpage to which the URL resolves and, after waiting a specified timeout period during which the webpage rendering is completed, a screen shot is captured and saved as an object (e.g., an image file such as a JPEG). Alternatively, as discussed above, the content fetcher 104 may obtain the screenshot via data caching system in a situation when the screenshot has been previously obtained and stored therein.

Upon obtaining a screenshot of the website to which each URL resolves, one or more screenshots may be provided to the feature generation logic 106. It should be noted that the content fetcher 104 may provide the one or more screenshots (or identifiers thereof, such as file names) to the feature generation logic 106 as other screenshots from the list of URLs are being collected, as opposed to obtaining the screenshots prior to passing the screenshots along to the feature generation logic 106. For each screenshot, the feature generation logic 106 is responsible for: (1) detecting keypoints within the screenshot, (2) generating keypoint descriptors based on the detected keypoints, and (3) generating a feature vector that includes the generated keypoint descriptors. The feature generation logic 106 uses computer vision techniques to detect the keypoints. According to one embodiment of the disclosure, the computer vision techniques may include detection of groupings of pixels wherein a grouping of pixels includes predetermined characteristics such as specified changes in grey scale. The feature generation logic 106 may utilize the computer vision techniques to detect edges and corners in images in the screenshot or more generally to perform density location operations, which detect groupings of pixels within the screenshot that include a high density of pixels (e.g., non-white space). The feature generation logic 106 may detect keypoints of the screenshots, the keypoints related to one another based on geometric measurements (e.g. distance between sets of keypoints, angle of intersection between sets of keypoints, etc.). Specific examples of the keypoint detection procedure will be well known to those skilled in the art. Additionally, various computer vision techniques may be utilized to detect keypoints. One example of a computer vision technique that may be utilized includes blob detection based on one or more matrices, e.g., the Hessian matrix, to determine change in the greyscale of pixels and generation of keypoints that exceed a predefined threshold of change according to a calculated determinant of one or more of the matrices. The term “blob” may refer to a region of pixels. Further, in one embodiment, a computer vision technique may be used to detect keypoints, which may then be generated and placed in a feature vector as discussed herein, such that the edge, corner and/or blob detection is dependent on detection of properties of a screenshot such as brightness, color, or greyscale compared to surrounding regions.

Upon detecting a plurality of keypoints within a sereenshot, the feature generation logic 106 determines a keypoint descriptor for each keypoint. A keypoint descriptor may be generated by extracting a predefined-sized block of pixels including the keypoint, dividing the block of pixels into sub-blocks, and taking into account a plurality of possible orientations of the pixels and storing such information (a keypoint descriptor may be referred to herein as a feature). In one embodiment, a vector (“a feature vector”) may then be created for each screenshot, the feature vector storing the plurality of keypoint descriptors for a particular screenshot.

The set of feature vectors for the screenshots are then provided to the feature domain mapper 108 by the feature generation logic 106. With respect to a first feature vector, the feature domain mapper 108 labels the first feature vector according to the webpage family of the webpage to which the URL resolves to which the generated features correspond. As a non-limiting example, a feature vector is generated for a URL that resolves to a Bank of America webpage (e.g., a log-in webpage). The feature vector (containing the generated features of the Bank of America webpage) is then labeled as “Bank of America.” Herein, labeling may correspond to appending to or otherwise associating an identifier with a feature vector. The feature domain mapper 108 performs the labeling process for each feature vector, wherein, in one embodiment, the webpage family may be provided along with the screenshot from the content fetcher 104 and further passed along with the feature vector from the feature generation logic 106.

The plurality of labeled feature vectors are provided to the training module 110, which generates a model, based on the plurality of feature vectors, to associate feature vectors based on labeling. The association of feature vectors may be based on a correlation of the plurality of feature vectors above a predefined (or variable) threshold. As mentioned above, in one embodiment, the model may represent modeling of a plurality of hyperplanes into which the features of each vector may be categorized. In such an embodiment, each URL may be representative of a webpage family with each webpage family having its own hyperplane. Each of the plurality of hyperplanes may be generated by the training module 110, based on the key point descriptors discussed above, as well as the keypoints themselves (i.e., keypoint center coordinates randy, a scale of the keypoint and an orientation of the keypoint). Additionally, in such an embodiment, the training module 110 may then generate a model that represents the plurality of hyperplanes. The model is then provided to the classifier 112 for use in the detection process, as discussed with respect to at least FIG. 2. In alternative embodiments, the model may be generated to represent the categorization of the plurality of feature vectors in other forms, such as in a model representing a histogram, wherein each bin of the histogram includes the feature vector corresponding to a webpage within the training set.

Referring now to FIG. 2, an exemplary block diagram of a logic flow during performance of a detection process by a PDAS according to an embodiment of the invention is shown. The logic flow 200 of the detection process illustrates the flow of data among logic modules of the PDAS 400, as seen in FIG. 4, in detecting URLs that resolve to phishing websites.

As a general overview, the detection process begins when the PDAS 400 receives a subject URL. In one embodiment, the pre-filter 116 is provided the subject URL and performs a pre-filtering step, discussed below. However, according to another embodiment, the PDAS 400 may receive an object and, in such an embodiment, an optional URL extractor 114 may first extract the subject URL (e.g., from an email or other object) and provide the extracted subject URL to the pre-filter 116 for pre-filtering. The pre-filter 116 performs a pre-filtering process, such as one or more static scans, on the URL, which may include performing whitelist/blacklist comparisons. When the subject URL is not found to be either malicious or benign, the subject URL is provided to the content fetcher 104, which obtains, in some embodiments, generates, a screenshot of the webpage to which the URL resolves, as discussed above.

The content fetcher 104 retrieves from the URL then provides the subject screenshot of the contents of the subject webpage (e.g., an image file, or an identifier, enabling retrieval of the image file), rendered by an interne browser, to the feature generation logic 106. As discussed above, the feature generation logic 106 detects keypoints within the subject screenshot and generates a feature vector based on the detected keypoints. The feature vector corresponding to the subject screenshot is provided to the classifier 112 for webpage family classification based on the model generated by the training module. As discussed above with respect to FIG. 1, each webpage family may correspond to a URL domain (e.g., each webpage family may correspond to a domain of a bank website such that the webpage families may be, for example, Bank of America, Wells Fargo, First Republic, etc.). A confidence may be determined for webpage family based on an analysis of the detected keypoint descriptors of the subject screenshot in accordance with the model. The confidence determined for a webpage family indicates the likelihood that the subject screenshot is attempting to mimic the webpage corresponding to the webpage family.

The webpage family having the highest confidence may be passed to an image comparator 120 (e.g., CV image comparator) which performs an image comparison between the subject screenshot and the webpage corresponding to the webpage family with the highest confidence. In some embodiments, one or more webpage families having the highest confidences are passed to the image comparator 120, which performs the image comparison for screenshots of webpages corresponding to the one or more webpage families. During training, a feature vector is determined for each webpage family and each feature vector is utilized in generating the model. The model provides confidences for each feature vector corresponding to URLs within the training; thus, the set of confidences provide an indication as to both the webpage family corresponding to the highest confidence and the feature vector, corresponding to a particular screenshot, having highest confidence. In a second embodiment, a webpage within a webpage family may be predefined as the webpage within a webpage family to be used in an image comparison when the webpage family is determined to have the highest confidence. In another embodiment, two or more (or all) webpages within a webpage family may be indicated as having the highest correlation and/or two or more (or all) webpages within the webpage family may be predefined as those to be used in an image comparison.

When the image comparison results in a match (e.g., correlation value) above a predefined threshold, the PDAS 400 determines that the subject webpage and the subject URL itself are part of a phishing attack. Where two or more webpages within a webpage family are used, the comparison with the subject webpage may be made separately for each of the webpages and, in alternatively embodiments (i) if any or a prescribed number of the resulting correlation values exceed the threshold, the URL is declared part of a phishing attack, or (ii) if the correlation value determined by statistically combining the separate correlation values (e.g., as by determining the mean, median, or mode of the separate correlation values) exceeds the threshold, the URL is declared part of a phishing attack.

In contrast to performing image comparisons on a large body of screenshots, the above-described detection process involving computer vision techniques analyzes only a relevant (based on a level of confidence) subset of screenshots associated with the subject screenshot.

More specifically, the detection process of FIG. 2 begins when the PDAS 400 receives a URL. As discussed above, the PDAS 400 may receive an object and, in such an embodiment, an optional URL extractor 114 may first extract the URL from the object and provide the extracted URL to the pre-filter 116 for pre-filtering. In another embodiment, the PDAS 400 may be provided with a URL, which may be passed directly to the pre-filter 116.

The pre-filter 116 performs a pre-filtering process on the URL, which may include one or more static scans such as whitelist/blacklist comparisons. In particular, the whitelist/blacklist database 118 stores data corresponding to whitelisted URLs (indicators determined to be benign) as well as blacklisted URLs (indicators determined to be associated with cyberattacks, e.g., phishing attacks). Comparisons performed by the pre-filter 116 between the whitelisted and blacklisted URLs stored in the whitelist/blacklist database 118 seek to remove any URLs known to be either benign or malicious. As a result of removing known benign or malicious URLs from the analysis, URLs passed on by the pre-filter 116 as not being knowingly benign or malicious and that resolve to webpages that very closely resemble known benign webpages (e.g., those of Bank of America, Wells Fargo, etc.) or malicious (e.g., known phishing webpages) are determined to be phishing webpages. Specifically, known benign URLs may be removed from the detection analysis by the pre-filter 116 (e.g., legitimate URLs of Bank of America, Wells Fargo, etc.) thus, URLs that are not removed by the pre-filter 116 and resolve to a webpage that very closely resembles the “look and feel” (graphic representation) of a benign webpage may be determined to be a phishing URL.

When the URL is not found to be either malicious or benign URL is not present in the blacklist or whitelist), the URL is provided to the content fetcher 104, which obtains a screenshot of the webpage to which the URL resolves, as discussed above with respect to the training process in accordance with FIG. 1. The content fetcher 104 then provides the screenshot of the webpage (e.g., an image file, or an identifier enabling, retrieval of the image file) to the feature generation logic 106. The feature generation logic 106 uses computer vision techniques to detect keypoints within the screenshot. The feature generation logic 106 extracts blocks of pixels from the screenshot having a predetermined size, e.g., a 16×16 block, that includes the keypoint. Each block of pixels is then used to generate a keypoint descriptor for the keypoint included within the block of pixels as discussed above. The plurality of keypoint descriptors describing the keypoints detected within a screenshot is stored in a vector, referred to herein as a “feature vector.” Specifically, the feature vector represents a description of the keypoints of the subject screenshot. The feature vector is then provided to the classifier 112.

The classifier 112 uses the feature vector of the subject screenshot as an input to the model generated during training. Analyzing the feature vector of the subject screenshot using the model results in a plurality of confidences. Each confidence of the plurality of confidences corresponds to a separate webpage family of the URLs provided to the PDAS 400 during training (“the training set”). As an illustrative example, when the training set includes URLs for Bank of America, Wells Fargo, First Republic, and other known banking webpages for a total of twenty (20) banking webpages in the training set, the analysis of the feature vector of the subject screenshot during the detection process may result in 20 confidences. Specifically, a first confidence may correspond to the Bank of America webpage, a second confidence may correspond to the Wells Fargo webpage, etc., with each confidence indicating the likelihood that the subject webpage is attempting to mimic the webpage corresponding to the webpage family. Continuing the example, the first confidence indicates the likelihood that the subject webpage is attempting to mimic the Bank of America webpage based on how closely the subject webpage resembles the “look and feel” of the Bank of America webpage.

The webpage family having the highest confidence may be passed to the image comparator 120, which performs an image comparison between the subject screenshot and the webpage corresponding to the webpage family with the highest confidence. The image comparison may perform an in-depth comparison of keypoints according to the keypoint descriptors within the feature vector of the subject screenshot with the keypoints of the webpage corresponding to the webpage family having the highest confidence to determine how closely the subject screenshot matches the webpage corresponding to the webpage family having the highest confidence. When the image comparison results in a match above a predefined threshold, the PDAS 400 determines that the subject webpage and the subject URL itself are part of a phishing attack.

When the subject URL and the subject webpage are determined to be part of a phishing attack, the reporting engine 122 generates an alert to a cybersecurity analyst, an administrator, and/or users of one or more endpoints indicating that the subject URL and subject webpage are part of a phishing attack.

In additional embodiments, a webpage family may include a plurality of webpages (e.g., Bank of America login webpages) that vary slightly. In such an embodiment, during the training process, the feature domain mapper 108 may label the feature vectors of the two or more webpages with the same webpage family and the feature vectors may be mapped to the same hyperplane during the generation of the model by the training module 110.

In some embodiments, two webpage families may correspond to the same overall webpage “owner.” For example, as Bank of America may have multiple login webpages for which the “look and feel” differs, a first Bank of America login webpage may include two text boxes corresponding to an entry of a customer's username and password, while a second Bank of America login webpage may include three text boxes corresponding to an entry of a customer's email address, social security number and birthday. Thus, for purposes of the training and detection processes, the first and second Bank of America differ in terms of their “look and feel” and may be afforded separate webpage families. However, both webpage families may be linked to Bank of America for the detection process.

Referring to FIGS. 3A-3B, a flowchart illustrating an exemplary method for analyzing a URL by the PDAS of FIG. 4 to determine whether the URL is associated with a phishing attack is shown. Each block illustrated in FIGS. 3A-3B represents an operation performed in the method 300 of detecting whether a URL is associated with a phishing attack by the phishing detection and analysis system (PDAS). Herein, the method 300 starts when the PDAS receives an object for phishing analysis (block 302). In one embodiment, the object may be a URL for analysis. However, in an alternative embodiment, the object may be, for example, an email message (email) wherein the content of the email includes a URL. In such an embodiment in which the object is an email, the method 300 may include an optional step of extracting the URL from the email (optional block 304).

Subsequently, the method 300 includes an operation of performing a pre-filter check on the URL (block 306). In one embodiment, the pre-filter check includes a static analysis of the URL, which may include, but is not limited to, a comparison with one or more entries of a whitelist and/or a comparison with one or more entries of a blacklist. In some embodiments, when the object is deemed suspicious and/or cannot be determined to be either benign or phishing, the method 300 continues analysis of the object by obtaining a screenshot of the webpage to which the URL resolves (“URL screenshot” as mentioned above) (block 308).

Upon obtaining the URL screenshot, the method 300 detects keypoints within the URL based on computer vision techniques and determines keypoint descriptors. Based on the keypoints and the determined keypoint descriptors, a feature vector is generated that includes the keypoints and their keypoint descriptors (block 310). The keypoints may include, inter alia, regions on the URL screenshot that are high-contrast regions of the URL screenshot. In one embodiment, a high-contrast region may refer to a set of two or more pixels such that a change in greyscale value is greater than or equal to a predefined threshold of two neighboring pixels. Additionally, in some embodiments, the URL screenshot may be in color and in such embodiments, the detection may include detection of a variance in color hues above a predefined threshold (e.g., a change in red, green, yellow, and blue values defining the pixel).

The feature vector is provided to a classifier, e.g., the classifier 112 as seen in FIGS. 1-2, and analyzed according to the model generated during the training process as discussed above. The analysis of the feature vector with the model results in a determination of a confidence for each feature vector included in the training set of URLs (block 312).

The webpage having the highest confidence based OD the analysis using the model is provided to, e.g., the CV image comparator 120 as seen in FIG. 2, which performs an image comparison between the URL screenshot and a screenshot of the webpage corresponding to the feature vector having the highest confidence (block 314). When the result of the image comparison is less than a predefined threshold, e.g., indicating a match of the two screenshots does not meet the predefined threshold (no at block 316), the method 300 determines the subject URL is not a phishing URL (block 318).

However, when the result of the image comparison is greater than or equal to the predefined threshold e.g., indicating a match of the two screenshots meets or exceeds the predefined threshold (yes at block 316), the method 300 determines the subject URL is a phishing URL (block 320) and subsequently generates and issues an alert (block 322). The alert may be issued to, for example, a user attempting to access the URL using an endpoint device, a network administer and/or a cybersecurity analyst.

FIG. 4 is an exemplary embodiment of a logical representation of the phishing detection and analysis system of FIG. 1. The phishing detection and analysis system (PDAS) 400, in an embodiment, may be stored on a non-transitory computer-readable storage medium of an endpoint device that includes a housing, which may be made entirely or partially of a hardened material (e.g., hardened plastic, metal, glass, composite or any combination thereof) that protects the circuitry within the housing, namely one or more processors 402 that are coupled to a communication interface 404 via a first transmission medium 406. The communication interface 404, in combination with a communication logic 412, enables communications with external network devices and/or other network appliances to receive updates for the PDAS 400. According to one embodiment of the disclosure, the communication interface 404 may be implemented as a physical interface including one or more ports for wired connectors. Additionally, or in the alternative, the communication interface 404 may be implemented with one or more radio units for supporting wireless communications with other electronic devices. The communication interface logic 412 may include logic for performing operations of receiving and transmitting one or more objects via the communication interface 404 to enable communication between the PDAS 400 and network devices via a network (e.g., the internet) and/or cloud computing services, not shown.

The processor(s) 402 is further coupled to a persistent storage 410 via a second transmission medium 408. According to one embodiment of the disclosure, the persistent storage 410 may include the following logic as software modules: the pre-filter 116, the URL extractor 114, the content fetcher 104, the feature generation logic 106, the feature domain mapper 108, the training module 110, the classifier 112, the image comparator 120, the reporting engine 122, and the communication interface logic 412. The operations of these software modules, upon execution by the processor(s) 402, are described above. The whitelist/blacklist database 118 is stored data for access by the pre-filter 116. Of course, it is contemplated that some or all of this logic may be implemented as hardware, and if so, such logic could be implemented separately from each other.

In the foregoing description, the invention is described with reference to specific exemplary embodiments thereof. It will, however, be evident that various modifications and changes may be made thereto without departing from the broader spirit and scope of the invention as set forth in the appended claims.

Claims

1. A computerized method for analyzing a subject Uniform Resource Locator (URL) by a phishing detection and analysis system (PDAS) to determine whether the subject URL is associated with a phishing attack, the PDAS including one or more processors and a storage medium, the method comprising: performing, by the PDAS, a first set of operations including: detecting a plurality of keypoints within a subject screenshot of a subject webpage corresponding to the subject URL;providing the plurality of keypoints to a machine learning model, wherein the machine learning model is a representation of relationships between a set of training feature vectors representing a training set of URLs, each of the training feature vectors corresponds to a separate webpage family within a set of webpage families;executing the machine learning model using the plurality of keypoints as input to the machine learning model to determine a set of confidences, wherein each confidence within the set of confidences reflects a potential relationship between the subject screenshot and a webpage family within the set of webpage families, and wherein a first webpage family within the set of webpage families is associated with a highest confidence of the set of confidences;performing, by the PDAS, a second set of operations including: performing an image comparison between the subject screenshot and one or more screenshots corresponding to webpages within the first webpage family;determining whether a result of the image comparison exceeds a predefined threshold;responsive to the determining that the result of the image comparison exceeds the predefined threshold, generating an alert or report indicating that the subject URL is associated with the phishing attack.

2. The computerized method of claim 1, further comprising: receiving, by the PDAS, the subject URL; andobtaining, by the PDAS, the subject screenshot of the subject webpage corresponding to the subject URL.

3. The computerized method of claim 1, further comprising: performing, by the PDAS, a pre-filter check on the subject URL, the pre-filter check including one or more static analyses.

4. The computerized method of claim 1, further comprising: determining, by the PDAS, from the subject screenshot, one or more keypoint descriptors for each of the plurality of keypoints, each of the one or more keypoint descriptors including one or more parameters of a keypoint of the plurality of keypoints; andgenerating, by the PDAS, a feature vector including the plurality of keypoints and the one or more keypoint descriptors for each of the plurality of keypoints,wherein determining, by the PDAS, the set of confidences includes analyzing the feature vector accordingly to the machine learning model.

5. The computerized method of claim 1, wherein the image comparison includes: retrieving, by the PDAS, previously determined keypoints of a first screenshot of the one or more screenshots; andcorrelating, by the PDAS, the previously determined keypoints of the first screenshot with the plurality of keypoints of the subject screenshot.

6. The computerized method of claim 1, wherein the subject screenshot includes image data of the subject webpage that is configured to be displayable on a computer screen.

7. The computerized method of claim 1, wherein each keypoint of the plurality of keypoints includes an image region having a known orientation within a set of pixels representing the subject screenshot.

8. The computerized method of claim 1, wherein a first confidence corresponds to a likelihood that the subject webpage is mimicking a first screenshot of a first webpage of the first webpage family, the first confidence included within the set of confidences.

9. The computerized method of claim 1, wherein the machine learning model is generated and trained with a plurality of screenshots of each of a plurality of webpages using a combination of an internet browser and an operating system.

10. The computerized method of claim 1, wherein the performing of the image comparison includes: performing the image comparison between the subject screenshot and one or more screenshots corresponding to webpages of the first webpage family until a result of an image comparison is greater than or equal to the predefined threshold.

11. The computerized method of claim 1, wherein the responsive to the determining that the result of the image comparison exceeds the predefined threshold includes transmitting the alert or report to one or more of a user of an endpoint device or an analyst.

12. The computerized method of claim 1, wherein the executing of the machine learning model indicates that a plurality of webpage families are closely correlated with the subject screenshot and the performing the image comparison is between the subject screenshot and a screenshot of at least one webpage from each of the set of webpage families including the one or more screenshots corresponding to the webpages within the first webpage family.

13. The computerized method of claim 1, further comprising: responsive to the determining that the result of the image comparison is below the predefined threshold, indicating that the subject URL is not associated with the phishing attack.

14. A non-transitory computer-readable medium, when processed by one or more processors, analyzes a subject Uniform Resource Locator (URL) to determine whether the subject URL is associated with a phishing attack, the non-transitory computer readable medium comprising: a feature generation logic module that, when executed by the one or more processors, detects a plurality of keypoints within a subject screenshot of a subject webpage corresponding to the subject URL and generates a subject feature vector that includes the detected plurality of keypoints;a classifier logic module that, when executed by the one or more processors, executes a machine learning model using the detected the plurality of keypoints as input to the machine learning model to determine a set of confidences, wherein each confidence of the set of confidences reflects a potential relationship between the subject screenshot and a first webpage family within a set of webpage families, and wherein the first webpage family within the set of webpage families is associated with a highest confidence of the set of confidence, wherein the machine learning model is a representation of relationships between a set of training feature vectors representing a training set of URLs, each of the training feature vectors corresponds to a separate webpage family within the set of webpage families;an image comparator logic module that, when executed by the one or more processors, (i) performs an image comparison between the subject screenshot and one or more screenshots corresponding to webpages within the first webpage family, and (ii) determines whether a result of the image comparison exceeds a predefined threshold; anda reporting logic module that, when executed by the one or more processors, responsive to the image comparator logic module determining that the result of the image comparison exceeds the predefined threshold, generates an alert or report indicating that the subject URL is associated with the phishing attack.

15. The computer-readable medium of claim 14, wherein the subject screenshot is captured as an image file.

16. The computer-readable medium of claim 14, further comprising: a content fetcher logic module that, when executed by the one or more processors, obtains the subject screenshot by accessing a data caching system that stores one or more previously captured screenshots.

17. The computer-readable medium of claim 14, further comprising: a pre-filter logic module that, when executed by the one or more processors, performs static scanning the subject URL including an analysis of one or more of a blacklist or white list.

18. The computer-readable medium of claim 14, wherein the image comparison includes: retrieving previously determined keypoints of a first screenshot of a first webpage; andcorrelating the previously determined keypoints of the first screenshot with the detected keypoints of the subject screenshot.

19. The computer-readable medium of claim 14, wherein each keypoint of the detected keypoints includes an image region having a known orientation within a set of pixels representing the subject screenshot.

20. The computer-readable medium of claim 14, wherein the machine learning model includes a representation of a correlation of the set of training feature vectors, wherein a first feature vector corresponds to a first webpage corresponding to a first webpage family of a set of webpage families.

21. The computer-readable medium of claim 14, wherein the machine learning model is generated and trained with a plurality of screenshots of each of a plurality of webpages using a combination of an internet browser and an operating system.

22. The computer-readable medium of claim 14, wherein the performing of the image comparison includes: performing the image comparison between the subject screenshot and one or more screenshots corresponding to webpages of the first webpage family until a result of an image comparison is greater than or equal to the predefined threshold.

23. The computer-readable medium of claim 14, wherein the responsive to the determining that the result of the image comparison exceeds the predefined threshold includes transmitting the alert or report to one or more of a user of an endpoint device or an analyst.

24. The computer-readable medium of claim 14, wherein the executing of the machine learning model indicates that a plurality of webpage families are closely correlated with the subject screenshot and the performing the image comparison is between the subject screenshot and a screenshot of at least one webpage from each of the set of webpage families including the one or more screenshots corresponding to the webpages within the first webpage family.

25. The computer-readable medium of claim 14, further comprising: responsive to the determining that the result of the image comparison is below the predefined threshold, indicating that the subject URL is not associated with the phishing attack.

26. A non-transitory computer-readable medium, when processed by one or more processors, generates a machine learning model used in determining whether a subject Uniform Resource Locator (URL) is associated with a phishing attack, the non-transitory computer readable medium comprising: a feature generation logic module that, when executed by the one or more processors, for each screenshot corresponding to a URL within a set of training URLs detects keypoints within each screenshot and generates a feature vector for each screenshot that includes the detected keypoints of the corresponding screenshot;a domain mapper logic module that, when executed by the one or more processors, receives each feature vector generated by the feature generation logic module and labels each feature vector according to a webpage family of the screenshot to which the feature vector corresponds to generate a plurality of labeled feature vectors; anda training module logic that, when executed by the one or more processors, generates the machine learning model including a digitized representation of a correlation of the plurality of labeled feature vectors corresponding to the set of training URLs, wherein the machine learning model is a representation of relationships between a set of training feature vectors representing the set of training URLs, each of the training feature vectors corresponds to a separate webpage family within a set of webpage families, and wherein execution of the machine learning model using a plurality of keypoints of a subject screenshot as input determines a set of confidences, wherein each confidence within the set of confidences reflects a potential relationship between the subject screenshot and a webpage family within the set of webpage families, and wherein a first webpage family within the set of webpage families is associated with a highest confidence of the set of confidences; andan image comparison logic module that, when executed by the one or more processors, (i) performs an image comparison between the subject screenshot and one or more screenshots corresponding to webpages within the first webpage family, and (ii) determines whether a result of the image comparison exceeds a predefined threshold.

27. The computer-readable medium of claim 26, wherein the subject screenshot is captured as an image file.

28. The computer-readable medium of claim 26, wherein the feature generation logic module uses at least one computer vision technique to detect the keypoints within each screenshot.

29. The computer-readable medium of claim 26, wherein the machine learning model represents one or more hyperplanes onto which features of each feature vector are categorized.

30. The non-transitory computer-readable medium of claim 26, further comprising: a content fetcher logic module that, when executed by the one or more processors, obtains a screenshot of a web site corresponding to each URL within the set of training URLs, the set of training URLs including URLs determined as likely to be targeted for phishing attacks, wherein each screenshot is obtained using a different combination of an internet browser and an operating system.

31. The computer-readable medium of claim 30, wherein the content fetcher logic module obtains a first screenshot by accessing a data caching system that stores one or more previously captured screenshots.

32. The non-transitory computer-readable medium of claim 26, further comprising: a reporting logic module that, when executed by the one or more processors, responsive to the image comparator logic module determining that the result of the image comparison exceeds the predefined threshold, generates an alert or report indicating that the subject URL is associated with the phishing attack.