Technique of accurate phishing detection
A combined semantic, structural, and visual inspection system using perceptual hashing and deep learning improves phishing detection accuracy by identifying deceptive URLs and pages with high precision, addressing the limitations of existing methods.
Patent Information
- Application Number
- PCT/EP2025/062764
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-05-09
- Publication Date
- 2026-11-12
AI Technical Summary
Existing phishing detection technologies suffer from high false positive and negative rates, particularly when attackers create visually and semantically convincing clones of legitimate websites, and are computationally expensive or require continuous updates.
A system that combines semantic, structural, and visual inspection techniques, using perceptual hashing and deep learning-based image hashing to analyze web content for sensitive intents and visual similarity with a registry of legitimate sites, classifying URLs as deceptive or non-deceptive.
Enhances phishing detection accuracy by identifying visually misleading pages with high precision and minimal latency, reducing false positives and negatives, and adapting to evolving attack strategies.
Smart Images

Figure EP2025062764_12112026_PF_FP_ABST
Abstract
Description
[0001] TECHNIQUE OF ACCURATE PHISHING DETECTION
[0002] TECHNICAL FIELD AND BACKGROUND
[0003] The present disclosure, optionally, relates to a network security system and, more particularly, but not exclusively, to classifying phishing web content in relation to user interaction and authentication.
[0004] Phishing is a form of cyberattack in which an attacker impersonates a legitimate entity in order to deceive users into disclosing sensitive information. A typical phishing attack may begin with the attacker creating a replica of a legitimate website — such as a financial institution's login page — and disseminating emails that spoof the identity of the institution. These emails may prompt recipients to follow a link to allegedly refresh security credentials or take similar action. Phishing emails are often crafted to exploit the risk-averse nature of human psychology and to evoke a sense of urgency, thereby increasing the likelihood of interaction occurring before the recipient has an opportunity to critically assess the communication.
[0005] For purposes of better understanding some embodiments of the present disclosure, as illustrated in Figures 2-6 of the drawings, reference is first made to a typical scenario of a typical phishing attack as illustrated in Figure 1. According to the background of the disclosure, a user first receives a phishing email 100. The email 100 may comprise a sender signature 101, crafted to conceal the fact that the true sender email is not related to the service being impersonated, a recipient 102 or a plurality thereof, and a subject 103 crafted to induce interaction of the targeted recipient with the email 100. Upon clicking a link 104 in an email, the recipient is directed to an adversarial website 110 impersonating the alleged sender using techniques such as a visually similar domain name 111, textual page content 113 and image content 112. On the adversarial website 110, the user may be further asked to input login credentials 114, 115 and other personal data 116. Depending on the nature of the attack and the impersonated service, the adversarial website 110 may invite the user to log in, verify personal details, update credentials and so on. Once the attacker collects the information (for example, after targeted user clicks a submit button 117), the adversarial site 110 may simulate a service outage or unresponsiveness, encouraging the user to try again later. In the meantime, the victim might be redirected to the legitimate website that was impersonated, unaware that their credentials have already been compromised. The attacker, having harvested credentials from numerous users, can then execute a large-scale fraudulent transaction, such as transferring funds and executing immediate withdrawals.
[0006] Various technical approaches have been developed in an effort to detect and mitigate phishing attacks, each with its own strengths and limitations. One of the earliest and most widely adopted methods involves the use of phishing URL blacklists. These systems rely on curated databases of known malicious web addresses that are flagged and blocked by security software. However, such an approach is inherently reactive. Blacklists must be continuously updated to remain effective, often lag behind newly emerging threats, and are susceptible to maintenance overhead and performance issues. Additionally, blacklist reliance can lead to false positives and, in some cases, may introduce vulnerabilities such as denial-of-service (DoS) scenarios due to overly broad filtering.
[0007] To improve detection beyond known URLs, content-based analysis techniques have been employed. These methods attempt to identify structural and transactional patterns commonly found in phishing websites. For example, a site may be flagged if it contains certain forms, buttons, or text indicative of credential harvesting or financial manipulation. Other common heuristics used in such systems include checking for unusual or suspicious URL structures such as misspellings, deceptive domain names, or utilization of HTTPS protocol. However, the effectiveness of content matching is frequently constrained by the high dependency on application context and domain-specific behavior. As a result, its accuracy can be inconsistent across distinct categories of web content.
[0008] Visual similarity analysis has also emerged as a detection mechanism, particularly targeting the graphical fidelity of phishing sites. These techniques are based on the observation that many fake websites replicate the visual style of authentic ones using low-resolution or poorly designed images. Detection systems may flag pages with visually degraded assets or discrepancies in layout. In some implementations, systems examine the quality of page design, including the presence of broken links — especially on rarely accessed informational pages — as well as evaluating the authenticity of contact information. Indicators of suspicious content include vague or incorrect phone numbers, missing physical addresses, or inconsistencies inprovided credentials. Other layers of heuristic detection may consider the quality of textual content on the page. Poor grammar, misspellings, or unprofessional phrasing may suggest a hastily constructed or automated phishing attempt. Nonetheless, as attackers become more adept at creating high-quality clones, this method has proven insufficient — especially when phishing pages closely mimic legitimate sites with high visual accuracy.
[0009] Machine learning (ML) classifiers represent a more dynamic approach, employing statistical models trained on large datasets to differentiate between benign and malicious web content. These classifiers typically analyze features such as URL structure, metadata, textual content, and user interaction patterns. While ML techniques can adapt to evolving attack strategies, they are often computationally expensive to train and deploy; their performance is sensitive to the quality of the input data, the feature selection process, and the thoroughness of model training. Consequently, error rates remain a challenge, and false negatives are a persistent risk.
[0010] Natural language processing (NLP) and deep learning techniques, including Long Short-Term Memory (LSTM) models, have been applied to analyze the semantics of web content. These methods are capable of understanding the intent behind textual elements, providing nuanced detection capabilities that can outperform simpler heuristic approaches. However, such models are often effective only when phishing pages contain low-quality or non-standard content and may be easily bypassed by replicating authentic content with high fidelity.
[0011] Another important vector in phishing detection is the analysis of ingress emails, which are a frequent initial delivery mechanism for phishing attempts. A common feature of phishing emails is the presence of an untrusted or ambiguous reference source, for example, an unfamiliar sender address or a domain that does not correspond with the organization purportedly represented. In many cases, hyperlinks embedded within phishing emails may not match the displayed hyperlink text. In other cases, phishing emails often direct recipients to websites that do not belong to the legitimate domain of the alleged sender but are chosen to graphically resemble it to the point of confusion. A further detection pattern may be the inclusion of alleged sensitive personal or account-related information within the body of the message itself, as legitimate organizations typically do not request or disclose private details through unsolicited email correspondence, and any such requests are considered suspicious under standard security practices.
[0012] Typical phishing detection systems rely heavily on one or more machine learning classifiers trained to analyze URLs and textual content extracted from messages and web pages. These classifiers are designed to assign a likelihood score indicating whether a given input is potentially deceptive, and to trigger a user alert or block access if a phishing threat is detected. While this approach has significantly advanced the field, it remains constrained by its reliance on training data and feature engineering. Recent state-of-the-art classifiers continue to exhibit considerable error rates (up to 10%) even under favorable conditions. In this context, there is a demand for new and improved techniques possessing higher precision and reliability.
[0013] SUMMARY
[0014] Optionally, the present disclosure introduces a mechanism for classifying potentially deceptive web pages using a combination of semantic, structural, and visual inspection techniques.
[0015] According to an aspect of some embodiments of the present disclosure there is provided a computer system for phishing detection, comprising a processor configured to perform the following: identify one or more sensitive intents in a web content addressable via a URL; compute a visual representation of the web content; compute a visual similarity between the visual representation and one or more reference visual representations in a registry; determine absence of the URL in the registry; and classify the URL as a deceptive URL based on the visual similarity, absence of the URL in the registry and presence of the one or more sensitive intents in the web content.
[0016] Optionally, computing a visual representation comprises applying at least one of (i) a perceptual hashing algorithm and (ii) a deep learning-based image hashing algorithm.
[0017] Optionally, the processor is further configured to, based on identifying no sensitive intents in the web content, classify the URL as a not deceptive URL.Optionally, identifying one or more sensitive intents in the web content comprises applying a large language model (LLM) to the web content.
[0018] Optionally, the one or more sensitive intents comprise at least one of a: authentication intent; one-time password acquisition intent; answer to a security question acquisition intent; financial account data acquisition intent; payment intent; outbound communication establishment intent; credential update intent; personally identifiable information acquisition intent; and biometric data acquisition intent.
[0019] Optionally, the computer system further comprises a non-transitory computer readable medium storing a URL classification database, wherein the processor is further configured to: responsive to determining presence of the URL in the URL classification database, output the classification associated with the URL in the URL classification database; record a deceptive classification associated with the URL in the URL classification database based on the visual similarity and identifying one or more sensitive intents in the web content; record a not deceptive classification associated with the URL in the URL classification database based on at least one of (i) identifying no sensitive intents in the web content and (ii) no visual similarity.
[0020] Optionally, each URL comprises one or more of a scheme, a top level domain, a registered domain, a subdomain, a port, a path, a query, and a fragment.
[0021] According to an aspect of some embodiments of the present disclosure there is provided a processor-implemented method for phishing detection, comprising: identifying one or more sensitive intents in a web content addressable via a URL; computing a visual representation of the web content; computing a visual similarity between the visual representation and one or more reference visual representations in a registry; determining absence of the URL in the registry; and classifying the URL as a deceptive URL based on the visual similarity, absence of the URL in the registry and presence of the one or more sensitive intents in the web content.
[0022] Unless otherwise defined, all technical and / or scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the disclosure pertains. Although methods and materials similar or equivalent to those described herein can be used in the practice or testing of embodiments of the disclosure, exemplary methods and / or materials are described below. In case of conflict, the patent specification, including definitions, will control. In addition, the materials, methods, and examples are illustrative only and are not intended to be necessarily limiting.
[0023] Implementation of the method and / or system of embodiments of the disclosure can involve performing or completing selected tasks manually, automatically, or a combination thereof. Moreover, according to actual instrumentation and equipment of embodiments of the method and / or system of the disclosure, several selected tasks could be implemented by hardware, by software or by firmware or by a combination thereof using an operating system.
[0024] For example, hardware for performing selected tasks according to embodiments of the disclosure could be implemented as a chip or a circuit. As software, selected tasks according to embodiments of the disclosure could be implemented as a plurality of software instructions being executed by a computer using any suitable operating system. In an exemplary embodiment of the disclosure, one or more tasks according to exemplary embodiments of method and / or system as described herein are performed by a data processor, such as a computing platform for executing a plurality of instructions. Optionally, the data processor includes a volatile memory for storing instructions and / or data and / or a non-volatile storage, for example, a magnetic hard-disk and / or removable media, for storing instructions and / or data. Optionally, a network connection is provided as well. A display and / or a user input device such as a keyboard or mouse are optionally provided as well.
[0025] BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS
[0026] Some embodiments of the disclosure are herein described, by way of example only, with reference to the accompanying drawings. With specific reference now to the drawings in detail, it is stressed that the particulars shown are by way of example and for purposes of illustrative discussion of embodiments of the disclosure. In this regard, the description taken with the drawings makes apparent to those skilled in the art how embodiments of the disclosure may be practiced.
[0027] In the drawings:FIG. 1 is an illustration of a typical phishing attack, according to the background of the disclosure;
[0028] FIG. 2 is an illustration of a phishing website classification decision-making module, according to some embodiments of the disclosure;
[0029] FIG. 3 is a flowchart of a phishing website classification system, according to some embodiments of the disclosure; FIG. 4 is a sequence diagram illustrating a phishing website classification system operation, according to some embodiments of the disclosure;
[0030] FIG. 5 is a schematic diagram showing a logical layout of a phishing website classification system, according to some embodiments of the disclosure; and
[0031] FIG. 6 is an illustration of integration modes of a phishing website classification system, according to some embodiments of the disclosure.
[0032] DESCRIPTION
[0033] The present disclosure, optionally, relates to a network security system and, more particularly, but not exclusively, to classifying phishing web content in relation to user interaction and authentication.
[0034] As used herein, the terms “authentic website” and “legitimate website” are used interchangeably to refer to a website that is known to a user and that performs its advertised or intended function.
[0035] As used herein, the terms “phishing website” and “adversarial website” are used interchangeably to refer to a website comprising one or more web pages created by an attacker for the purpose of impersonating an authentic website.
[0036] As used herein, the term “URL” means “uniform resource locator.”
[0037] In accordance with various embodiments of the present disclosure, there is disclosed a method and system for the accurate detection and classification of phishing websites, based on a coordinated analysis of semantic, structural, and visual attributes. Central to this disclosure is a new principle of phishing classification that relies on the joint evaluation of perceptual similarity between a suspicious page and a known legitimate version, and the inferred intent behind the content of the page. This dual evaluation enables a deeper and more nuanced assessment of web content, identifying pages that may visually or linguistically mislead users into disclosing sensitive information.
[0038] According to one use case disclosed in the present application, a user accessing a web page through a secure browsing environment may be deceived to navigate to an adversarial website visually replicating a legitimate website and hosted on a different domain. The adversarial website may be convincingly designed and contain motivational cues prompting the user to enter account credentials, share personal information, or perform other sensitive actions. In this use case, the aim of the present disclosure is to classify, responsive to the user attempting to access a web page, the web page as legitimate or adversarial, and deliver the web page content to the user or block it, respectively.
[0039] The approach described hereinbelow is particularly advantageous in contexts where attackers create visually and semantically convincing clones of legitimate websites, and where traditional techniques based on blacklist matching or URL heuristics produce an unacceptable rate of false negatives or false positives.
[0040] The method described hereinbelow is particularly useful in systems where phishing attempts must be identified with high precision and minimal latency, such as in real-time browsing environments, secure email gateways, or digital banking platforms. The disclosure further supports integration with legacy phishing protection systems, enabling flexible usage patterns to further improve web content loading latency and reduce total computational costs.
[0041] Before explaining at least one embodiment of the disclosure in detail, it is to be understood that the disclosure is not necessarily limited in its application to the details of construction and the arrangement of the components and / or methods set forth in the following description and / or illustrated in the drawings. The disclosure is capable of other embodiments or of being practiced or carried out in various ways.
[0042] Referring now to the drawings, Figure 2 illustrates the phishing website classification decision-making module. According to some embodiments of the disclosure, the decision-making module 200 receives a plurality of legitimate websites210, each comprising a first URL 211 and a first web page content 212, and an adversarial website 110, comprising a second URL 221 and a second web page content 222.
[0043] Optionally, each of the web page contents 212, 222 may comprise a respective plurality of sensitive intent elements 213, 223. Examples of sensitive intent elements include, but are not limited to, HTML text edit boxes, text areas or password edit boxes, web application components implemented with technologies such as canvas and emulating the functionality of edit boxes, date selection components and so on. Other examples of sensitive intent elements include informational messages inviting the user to call a phone number or install a mobile phone application.
[0044] Optionally, responsive to determining a presence of one or more sensitive intents in a first web page content, visual similarity of the first web page content with at least one second web page content exceeding a configured threshold, and determining a mismatch of the first URL with the respective second URL, the decision-making module 200 classifies the first URL as deceptive 231, otherwise classifies the first URL as non-harmful 232.
[0045] Reference is now made to Figure 3, illustrating a flowchart of a phishing website classification system. According to some embodiments of the disclosure, the process proceeds through the following steps.
[0046] At a step 301, the phishing website classification system receives a website URL. The URL may be supplied directly by a web browser, by an enterprise security system, or by other software components. Next, the process proceeds to a decision step 302.
[0047] At the decision step 302, the system makes a determination whether the URL is present in a classification decision cache. If the URL is present in the classification cache, the process proceeds to a step 306, otherwise the process proceeds to a step 303.
[0048] At the step 303, the system submits the URL into a baseline phishing classification subsystem, thereby obtaining a preliminary computationally inexpensive classification decision and a respective confidence level. Next, the process proceeds to a decision step 304.
[0049] At the decision step 304, the system makes a determination whether the confidence level of the decision produced by the baseline phishing classification subsystem exceeds a configured threshold. If the confidence level exceeds the configured threshold, the process proceeds to a step 305, otherwise the process proceeds to a decision step 307. This enables the system to activate the deep inspection methods solely if the baseline system does not achieve a result with a satisfactory confidence level, and a more precise analysis is required.
[0050] At the step 305, the system saves the URL and the respective decision into a cache for future reuse. For example, the cache may comprise a persistent cache or reside in random-access memory. This enables the system to identify the respective decision of the URL at step 302 during subsequent executions, avoiding repeatedly executing classification for a same URL.
[0051] Optionally, the cache comprises a programmatic interface, enabling purging of decision entries from the cache upon request.
[0052] Optionally, the cache comprises one of a time-based or a rule-based entry expiration trigger, enabling purging of decision entries from the cache responsive to timeouts or external events.
[0053] At the decision step 307, the system makes a determination whether the web page content at the received URL comprises at least one sensitive intent. If the web page content comprises at least one sensitive intent, the process proceeds to a decision step 309, otherwise the process proceeds to a step 308a. A sensitive intent describes a functionality of a page that is associated with elevated risks and may be utilized by an attacker. This check enables the system to avoid spending computational and memory resources on classifying trivial web page that cannot be practically utilized for phishing purposes. In an example, a trivial web page comprises text and image content and does not comprise certain interactive elements.
[0054] Optionally, the system performs a check for a presence of at least one of a login and password authentication intent, a one-time password acquisition intent, including variations involving single-use authentication hyperlinks, an answer to a security question acquisition intent, a financial account data acquisition intent, an initiation or completion of a payment procedure intent, including a bank card payment, a e-wallet payment or a cryptocurrency payment, an outbound communicationestablishment intent, including a call-to-action to dial a phone number or initiate a messenger dialogue, a credential update intent, a personally identifiable information acquisition intent, and a biometric data acquisition intent.
[0055] Optionally, the system applies a large language model (LLM) to the web page content, thereby performing a check for a presence of sensitive intent. For example, the system may apply NuExtract LLM.
[0056] Optionally, the system computes embedding vectors similarity of the web page content to a reference web page content, thereby performing a check for a presence of sensitive intent.
[0057] Optionally, the system applies one of a hierarchical navigable small world graph, a k-dimensional tree, a scalable nearest neighbors search, an approximate nearest neighbors search, a term frequency-inverse document frequency approach, a Jaccard similarity metric, a Levenshtein distance metric, a clustering in latent semantic space approach, and a graph-based similarity analysis to the web page content and a plurality of true positive samples, thereby performing a check for a presence of sensitive intent.
[0058] At the step 308a, the system produces a classification decision that the web page content at the received URL is non-harmful. Next, the process proceeds to the step 305.
[0059] At the step 309, the system makes a determination whether the web page content at the received URL is visually similar to at least one web content entry in a plurality of web content entries stored in a legitimate web page registry. As an adversarial web page impersonating a legitimate website by definition can only be created after the respective legitimate web page being impersonated has already been created. Therefore, maintaining a registry of visual representations of legitimate websites enables the system to compare a visual representation of an examined web page content against each entry in the registry. If the maximum visual similarity metric value between the web page content and each web content entry in the legitimate web page registry is below a configured threshold, the process proceeds to a step 308b, otherwise the process proceeds to a step 312.
[0060] At the step 308b, the system produces a classification decision that the web page content at the received URL is authentic. Next, the process proceeds to a step 310.
[0061] At the step 310, the system computes a visual representation of the web page content. For example, the visual representation may be computed using a perceptual hashing algorithm or a deep learning based image hashing algorithm. Next, the process proceeds to a step 311.
[0062] At the step 311, the system creates a new entry in the legitimate web page registry, comprising the received URL and a computed visual representation of the web page content. This enables the system to update and refresh the registry based on added information. Next, the process proceeds to the step 305.
[0063] Optionally, computing a visual representation of a web page content comprises applying at least one perceptual hashing algorithm.
[0064] Optionally, computing a visual representation of a web page content comprises applying at least one deep learning based image hashing algorithm.
[0065] Optionally, the perceptual hashing algorithm comprises one of discrete cosine transform based hashing, discrete wavelet transform based hashing, Marr-Hildreth operator based hashing, radial variance based hashing, singular value decomposition based hashing, histogram-based hashing, color moments based hashing, feature point based hashing, and block mean value based hashing.
[0066] Optionally, the deep learning based image hashing algorithm comprises one of convolutional neural network based hashing, convolutional autoencoder based hashing, and generative adversarial networks based hashing.
[0067] Optionally, the legitimate web page registry comprises one of a time-based or a rule-based entry expiration trigger, enabling purging of web page visual representation entries from the registry responsive to timeouts or external events. Optionally, the legitimate web page registry comprises a plurality of visual representations of web page semantic blocks. For example, the web page registry may store visual representations of individual components of web pages such as data entry areas, header and footer bars, background areas, material panes and so on. This enables the system to perform checksof individual semantic block components of an examined web page against individual semantic block components in the registry, thereby further improving detection robustness.
[0068] At the step 312, the system produces a classification decision that the web page content at the received URL is deceptive. Next, the process proceeds to the step 305.
[0069] Optionally, an URL comprises one or more of a scheme, a top level domain, a registered domain, a subdomain, a port, a path, a query, and a fragment. The system
[0070] Reference is now made to Figure 4, illustrating sequence diagram of a phishing website classification system operation. According to some embodiments of the disclosure, the system comprises a payload classifier 401, a URL decision cache 402, a URL hashing module 403, a URL evaluation module 404, a rendering module 405, and a deep inspector module 406. The system further interacts with a user 400 via a user agent such as a web browser, a web server 420 and a legitimate web page registry 430.
[0071] Optionally, a payload classifier 401 receives 301 from a user 400 a URL to be inspected, and submits the URL to the URL hasher 403. The URL hasher computes 407 and returns the computed URL hash to the payload classifier 401.
[0072] Optionally, the payload classifier submits 408 the hash to the URL cache 402 and receives a response comprising a decision or a marker of absence thereof. If a deceptive classification decision is received, the payload classifier 401 outputs the decision in a configured manner - for example, by redirecting 306 the user 400 to a warning page, blocking loading of the requested page. If a legitimate classification is received, the payload classifier outputs the decision in a configured manner - for example, by requesting 409 the URL from a respective web server 420 and enabling the user 400 to receive 410 the web content addressed by the URL.
[0073] Optionally, if a marker of absence of a cached decision is received, the payload classifier 401 submits 303 the URL for classification to the baseline URL evaluator 404 and receives the response.
[0074] Optionally, the payload classifier checks the classification response confidence 304. If the received baseline classification response comprises a confidence value exceeding a configured threshold, the payload classifier 401 submits 305a the URL, the hashed URL, and the classification response to the URL cache 402 for caching, and outputs the classification decision in a configured manner. For example, if a deceptive classification decision is received, the payload classifier 401 redirects 306 the user 400 to a warning page, blocking loading of the requested page. For example, if a legitimate classification is received, the payload classifier requests 409 the URL from a respective web server 420 and enables the user 400 to receive 410 the web content addressed by the URL.
[0075] Optionally, if the received baseline classification response comprises a confidence value below the configured threshold, the payload classifier 401 requests 409 the web content addressed by the URL from the web server 420 and receives 410 the response. The payload classifier 401 subsequently submits 411 the URL and the respective web content to the deep inspector 406.
[0076] Optionally, the deep inspector 406 submits 412 the web content to the Tenderer 405 and receives a rendered visual representation of the web content.
[0077] Optionally, the deep inspector 406 detects 413 elements associated with a sensitive intent in the web content. Optionally, the deep inspector 406 computes 414 a fingerprint of the rendered visual representation of the web content and submits 415 the fingerprint to the legitimate web page registry 430.
[0078] Optionally, based on the results of the previous steps, the deep inspector 406 classifies 416 the web content and returns 417 the classification decision to the payload classifier 401.
[0079] Optionally, if a legitimate classification is computed, the deep inspector 406 submits 310 the URL and one of a visual representation of the web page addressed by the URL or the fingerprint of the visual representation to the legitimate web page registry 430.
[0080] Optionally, the deep inspector 406 submits 305b the URL and the classification result to the URL cache 402. Optionally, the payload classifier 401 updates 305c the respective entry at the URL cache 402 with the URL hash value.Optionally, the payload classifier 401 performs a configured classification action 418. For example, if a deceptive classification decision is received, the payload classifier 401 redirects 306 the user 400 to a warning page, blocking loading of the requested page. For example, if a legitimate classification is received, the payload classifier returns to the user 400 the previously received web content.
[0081] Reference is now made to Figure 5, illustrating a schematic diagram of a phishing website classification system logical layout. According to some embodiments of the disclosure, the deep inspector 406 comprises a flow controller 500, an intent inspector 510 and a visual similarity evaluator 520, and the payload classifier 401 comprises the URL hasher 403. The flow controller 500 is configured to interface with an existing phishing classifier and to selectively trigger a deep inspection flow when the confidence level of the preliminary classification falls below a configured threshold. The intent inspector 510 is adapted to analyze the semantic content of the web page and identify sensitive intents. In some embodiments, the intent inspector may utilize natural language processing (NLP) techniques to evaluate the linguistic cues present on the page. The visual similarity evaluator 520 is configured to generate a fingerprint of the visual content of the page, using techniques such as perceptual hashing or deep learning-based image analysis. Such comparison enables detection of visually deceptive content even when the underlying images have been resized, compressed, or otherwise modified.
[0082] Optionally, the system monitors aggregate statistics such as false-positive rate, false-negative rate, and distributional drift in perceptual-hash distances. Responsive to a drift metric exceeding a predefined threshold, the flow controller 500 automatically lowers the deep-inspection trigger threshold or configures the visual similarity evaluator 520 with an expanded visual similarity search radius, thereby adapting to previously unseen spoofing methods.
[0083] Optionally, the system further comprises a registry of adversarial websites, enabling the system to recognize and block derivative attacks that reuse known visual or semantic patterns.
[0084] Optionally, the system further logs each classification decision pertaining to an examined website, the respective confidence score and the plurality of embedding vectors of the visual representation of the website to a feedback store. A scheduled background process periodically, episodically or responsive to an external trigger adjusts the baseline URL-classification subsystem and the sensitive intent detection model on the accumulated feedback set, thereby incrementally refining detection accuracy without interrupting real-time operation.
[0085] Reference is now made to Figure 6, illustrating integration modes of a phishing website classification system. According to some embodiments of the disclosure, the deep inspector 406 may be integrated with at least one of an enterprise web access filtering system 600A or with a web hosting provider system 600B.
[0086] Optionally, in the enterprise web access filtering system scenario 600A the deep inspector 406 is invoked by the payload classifier 401, which is in turn invoked by an Intrusion detection system / Intrusion prevention system (IDS / IPS) 610. All web content requests and responses are routed through the IDS / IPS 610 according to an enterprise security policy.
[0087] Optionally, in the web hosting provider scenario 600B the deep inspector 406 is integrated with a web hosting system 620, enabling the web hosting provider to detect and act on phishing websites attempting to utilize its infrastructure.
[0088] The terms "comprises", "comprising", "includes", "including", “having” and their conjugates mean "including but not limited to".
[0089] The term “consisting of’ means “including and limited to.”
[0090] The term "consisting essentially of' means that the composition, method, or structure may include additional ingredients, steps, and / or parts, but only if the additional ingredients, steps, and / or parts do not materially alter the basic and novel characteristics of the claimed composition, method, or structure.
[0091] It is appreciated that certain features of the disclosure, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the disclosure, which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable subcombination or as suitable in any other described embodiment of the disclosure. Certain features described in the context of various embodiments are not to be considered essential features of those embodiments unless the embodiment is inoperative without those elements.Although the disclosure has been described in conjunction with specific embodiments thereof, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art. Accordingly, it is intended to embrace all such alternatives, modifications and variations that fall within the spirit and broad scope of the appended claims.
[0092] It is the intent of the Applicant s) that all publications, patents and patent applications referred to in this specification are to be incorporated in their entirety by reference into the specification, as if each individual publication, patent or patent application was specifically and individually noted when referenced that it is to be incorporated herein by reference. In addition, citation or identification of any reference in this application shall not be construed as an admission that such reference is available as prior art to the present disclosure. To the extent that section headings are used, they should not be construed as necessarily limiting. In addition, any priority documents) of this application is / are hereby incorporated herein by reference in its / their entirety.
Claims
CLAIMS1. A computer system for phishing detection, comprising:a processor configured to perform the following:identify one or more sensitive intents in a web content addressable via a URL;compute a visual representation of the web content;compute a visual similarity between the visual representation and one or more reference visual representations in a registry;determine absence of the URL in the registry; andclassify the URL as a deceptive URL based on the visual similarity, absence of the URL in the registry and presence of the one or more sensitive intents in the web content.
2. The computer system according to claim 1, wherein computing a visual representation comprises applying at least one of (i) a perceptual hashing algorithm and (ii) a deep learning based image hashing algorithm.
3. The computer system according to claim 1, wherein the processor is further configured to, based on identifying no sensitive intents in the web content, classify the URL as a not deceptive URL.
4. The computer system according to claim 1, wherein identifying one or more sensitive intents in the web content comprises applying a large language model (LLM) to the web content.
5. The computer system according to claim 1, wherein the one or more sensitive intents comprise at least one of a: authentication intent;one-time password acquisition intent;answer to a security question acquisition intent;financial account data acquisition intent;payment intent;outbound communication establishment intent;credential update intent;personally identifiable information acquisition intent; andbiometric data acquisition intent.
6. The computer system according to claim 1, further comprising:a non-transitory computer readable medium storing a URL classification database,wherein the processor is further configured to:responsive to determining presence of the URL in the URL classification database, output the classification associated with the URL in the URL classification database;record a deceptive classification associated with the URL in the URL classification database based on the visual similarity and identifying one or more sensitive intents in the web content;record a not deceptive classification associated with the URL in the URL classification database based on at least one of (i) identifying no sensitive intents in the web content and (ii) no visual similarity.
7. The computer system according to claim 1,wherein each URL comprises one or more of a scheme, a top level domain, a registered domain, a subdomain, a port, a path, a query, and a fragment.
8. A processor-implemented method for phishing detection, comprising:identifying one or more sensitive intents in a web content addressable via a URL;computing a visual representation of the web content;computing a visual similarity between the visual representation and one or more reference visual representations in a registry;determining absence of the URL in the registry; andclassifying the URL as a deceptive URL based on the visual similarity, absence of the URL in the registry and presence of the one or more sensitive intents in the web content.
