Pith. sign in

REVIEW 4 major objections 6 minor 43 references

PhishKey: A Novel Centroid-Based Approach for Enhanced Phishing Detection Using Adaptive HTML Component Extraction

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read PhishKey combines a URL CNN with a centroid-based HTML extractor to reach 98.70% F1 and resist injection attacks.

desk verdict CAPE's test-time selection rule is either circular or unspecified, so the claimed accuracy and robustness are unsubstantiated. read the letter →

arxiv 2506.21106 v1 pith:Q4QTDZ6G submitted 2025-06-26 cs.CR cs.AI

Felipe Castaño , Eduardo Fidalgo , Enrique Alegre , Rocio Alaiz-Rodríguez , Raul Orduna , Francesco Zola This is my paper
classification cs.CRcs.AI
keywords phishingdetectioncentroid-basedextractionHTMLtokenizationcharacter-levelCNNWord2Vecbagofwordssoft-votingensembleinjectionattacks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

PhishKey is a phishing detector that reads both the URL and the raw HTML of a page, with no hand-crafted features and no cropping of long pages. Its HTML branch, CAPE, picks the 2,000 most representative words of each page by cosine similarity to per-class centroids learned from word embeddings, then classifies that condensed bag of words. The paper reports up to 98.70% F1 across four public datasets and, under a 2,000-word injection attack, performance drops of only 1.32 to 8.53 percentage points, against 8.80 to 27.88 for the cropping baseline. If that holds, phishing detectors can use the whole page without sacrificing input length limits.

What carries the argument

The central object is CAPE (Centroid-Based Key Component Phishing Extractor), a module that converts an arbitrarily long HTML document into a fixed set of salient tokens. Its defined pipeline is: tokenize the HTML; embed every token into $\mathbb{R}^{100}$ with Word2Vec; compute one centroid per class as the mean embedding of the tokens in that class; for each sample select the top-2,000 tokens by cosine similarity to the centroid of that sample's class; build a bag-of-words vector over the selected vocabulary; and classify that vector with a Random Forest. The companion machinery is a character-level CNN on the URL and a soft-voting ensemble over the two modules' probability outputs. CAPE's role is to let the HTML be analyzed in full, with length normalized by selection rather than by cropping, which the paper argues is why performance degrades little when misleading words are prepended.

What would settle it

Re-run the full pipeline on a held-out test set, but pick the tokens by cosine similarity to the nearest of the two training-set centroids instead of to the sample's true class centroid; if F1 drops materially, the reported 98.70% depended on test labels being known. A simpler check is to inspect the inference code for any call that passes a ground-truth label into the token-selection step.

Watch

Extended reading notes

Core claim

The central claim is that representing an HTML page by the words closest to its class's centroid—rather than by a fixed-length crop—preserves the evidence a classifier needs while discarding noise, and that combining this with a character-level URL model makes phishing detection both more accurate and harder to fool with injected content. In PhishKey, the HTML branch tokenizes a page, embeds each word with Word2Vec into 100 dimensions, computes a centroid for the legitimate class and one for the phishing class, keeps the 2,000 tokens per sample that are closest to that sample's class centroid under cosine similarity, builds a bag-of-words vector, and classifies with a Random Forest. A soft-voting ensemble fuses that prediction with a character-level convolutional network on the URL. Across the four evaluation datasets, the paper reports F1 scores up to 98.70% with lower cross-validation standard deviations than the baseline, and when 2,000 words from the opposite class are prepended to each test sample, PhishKey degrades by 1.32 to 8.53 percentage points while the baseline degrades by 8.80 to 27.88.

Load-bearing premise

The selection of the 2,000 key tokens uses each sample's true phishing-or-legitimate label to choose which class centroid to measure against, and the paper does not say how that label is obtained when classifying a new page.

Editorial extensions

If this is right

  • Pages that place phishing content after the first 2,000 words are no longer automatically truncated out of the analysis.
  • Updating the detector to new attack patterns requires retraining the embeddings and centroids rather than redesigning hand-crafted features.
  • An attacker who injects decoy words at the start of a page must make those words semantically close to the target class, not merely frequent.
  • With only 5% of the training data, PhishKey still reports higher F1 than the cropping baseline on the larger datasets, suggesting the approach is usable when labeled pages are scarce.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: if token selection truly uses the sample's true class centroid at inference, then a deployment version must use a nearest-centroid rule instead, and its F1 should be measured explicitly.
  • Beyond the paper: an attacker who injects paraphrased legitimate text rather than raw words from another dataset may be harder to defeat; this variant is not tested.
  • Beyond the paper: since the bag-of-words representation discards word order, feeding the selected tokens to a sequence-aware model could improve the datasets where the cropping baseline still wins.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes PhishKey, a phishing detector that combines a character-level URL CNN with CAPE, an HTML word-level extractor. CAPE tokenizes HTML, maps tokens to Word2Vec embeddings, computes class centroids for phishing and legitimate pages, selects the 2,000 most representative tokens per sample by cosine similarity to the centroid of the sample's class, builds a Bag-of-Words vector, and classifies it with a Random Forest. The URL and HTML predictions are combined by soft voting. Experiments on four datasets report F1 scores up to 98.70% and claim that under a 2,000-word injection attack the method degrades by only 1.32 to 8.53 percentage points, compared with 8.80 to 27.88 points for WebPhish. The central selection step, however, requires the ground-truth class of each sample, which is unavailable at test time; the reported results are therefore either circular or unreproducible as described.

Significance. If the method worked as claimed, the paper would offer a useful practical contribution: automatic feature extraction from HTML without manual feature engineering, evaluation across four datasets, and an explicit adversarial injection protocol. The centroid-based extractive selection idea is interesting, and the experimental design includes useful controls such as training-set reductions, 5-fold cross-validation, and cross-dataset injection. These strengths cannot rescue the central flaw: Equation (4) makes the feature representation depend on the label being predicted. No label-free inference rule is stated and no code is released, so the accuracy and robustness claims are not supported for a deployable system. The paper also internally contradicts its 'complete sample processing, no cropping' contribution by selecting only 2,000 tokens per sample.

major comments (4)
  1. [Section 3.1.3, Eq. (4)] The selection rule T_selected = top-m{sim(v_i,j, c_c) | v_i,j in d_i} uses c_c for 'their respective class' c in {C_phishing, C_legitimate}. In the experimental setup each test sample has a known true label; if that label is used to choose c_c, the BoW input encodes the answer before the classifier runs, making the accuracy and F1 values in Tables 4 and 5 circular. At inference time the true class is unknown, so the rule cannot be applied. The paper nowhere specifies a label-free rule such as nearest-centroid selection, a preliminary classifier, or a rule based on the URL module, and no code is released. This is the central mechanism of CAPE, not a presentation issue.
  2. [Section 5, Table 5 (injection experiment)] The robustness claim is a direct consequence of the same leak. When 2,000 words of the opposite class are prepended to a test sample, a selector that knows the true class will tend to discard those injected words because they are far from the centroid of the true class. This explains the reported degradation of only 1.32 to 8.53 points versus 8.80 to 27.88 points for WebPhish. The experiment therefore does not demonstrate robustness of a deployable method; it demonstrates that the feature extractor was given the label.
  3. [Section 3.1.1 and Section 4.3] The paper does not state whether the Word2Vec embeddings are trained on the training split only or on the full dataset before the 74:16:20 split and 5-fold cross-validation. If the latter, test-set information enters through the unsupervised embedding, creating a second leakage channel. This must be clarified, and if necessary the evaluation must be rerun with embeddings fit only on the training folds.
  4. [Section 3.1 and Section 3.1.4] The paper repeatedly claims 'complete sample processing' and 'without losing information' (Abstract and Section 3.1), yet CAPE discards all tokens outside the selected 2,000 before BoW construction. Selecting 2,000 tokens is itself a cropping operation, and the claim that no information is lost is not supported. The text should describe the selection as lossy and explain why the chosen m preserves the class-relevant content.
minor comments (6)
  1. [Section 3.1.2] K-means is mentioned but not used in the centroid calculation; the centroid in Eq. (2) is simply the arithmetic mean of the class embeddings. Either remove the k-means reference or explain how clustering is incorporated into the method.
  2. [Tables 4 and 5] The column header 'PhishKey Extractor' is ambiguous: it is unclear whether the results are for CAPE alone or for the full soft-voting ensemble with the URL CNN. The text in Section 5 uses 'PhishKey' and 'PhishKey Extractor' inconsistently.
  3. [Section 4.3] The description of the data split is confusing: '74:16:20' sums to 110, and the text then says 20% of the training samples are extracted for validation. Please state the split unambiguously.
  4. [Abstract and Section 6] The headline 'up to 98.70% F1' is reported without stating that it is the result on the Opara et al. dataset at the 100% training reduction, as shown in Table 4; the abstract should tie the number to its experimental condition.
  5. [Table 2] There are minor typos: 'Phising Source' should be 'Phishing Source', and 'Stuf Gate42' should be 'Stuff Gate42' or the form used in the original reference.
  6. [Section 4.3, injection attack] The description of the injection attack should specify whether the 2,000 injected words are prepended to the existing HTML content or replace the first 2,000 tokens, and how the custom tokenizer treats the boundary between injected and original content.

Circularity Check

2 steps flagged · score 8.0 of 10

CAPE's token selection (Eq. 4) uses the ground-truth class centroid at inference, making the reported F1 and injection robustness either label-leaked or unreproducible.

  1. self definitional [Section 3.1.2-3.1.3, Eqs. (2)-(4)]
    "Each sample d_i belongs to a class c ∈ {C_phishing, C_legitimate}. ... For each sample d_i, the m most representative tokens are selected based on their proximity to the centroid of their respective class. ... T_selected = top-m{sim(v_i,j, c_c)|v_i,j ∈ d_i}."

    Eq. (2) computes the centroid from embeddings of samples of the true class c, and Eq. (4) defines the selected token set using that same c ('their respective class'). For an unlabeled test sample, c is exactly the label the classifier is supposed to output. If the true label is used in Eq. (4), the BoW input already encodes the target, so the reported 98.70% F1 is forced: the classifier is given tokens chosen from the centroid of the correct class before it makes any prediction. The paper never states a label-free inference rule (e.g., nearest-centroid selection), so either the evaluation leaks labels or the method is underspecified and unreproducible.

  2. self definitional [Section 4.3 and Section 5, Experiment 3, Table 5]
    "we introduce an injection attack by placing 2,000 words from a different class at the beginning of each sample; the data injected is selected from a different dataset than the one used to train the model, ensuring that the prediction of the model is not influenced by prior knowledge of the injected data."

    The injected words come from the opposite class. Under Eq. (4), the selector keeps only tokens near the centroid of the sample's true class, so the opposite-class injections are discarded before the BoW representation reaches the Random Forest. The claimed injection resistance (drops of only 1.32-8.53 points vs 8.80-27.88 points for WebPhish) is therefore a direct consequence of label-based token selection, not of learned robustness. Unless a label-free selection rule is specified and applied at test time, the comparison in Experiment 3 is invalid by construction.

full rationale

The central derivation chain of CAPE is not self-contained. Equation (4) makes the feature representation depend on the true class centroid 'c_c' of the sample, which is the classification target itself; if applied literally at test time, the BoW input is defined in terms of the label, so the headline accuracy and the injection-robustness results reduce to the ground-truth label. If the authors instead intended a label-free rule, such as nearest-centroid selection or a preliminary classifier, that rule is absent from the paper, making the evaluation unreproducible. The remaining components (character-level URL CNN, Word2Vec embeddings, Random Forest, and soft-voting ensemble) are standard and are not circular; the paper's self-citations to the PILWD dataset and Phikita are not load-bearing for the main derivation. However, because the core CAPE mechanism either leaks labels or is unspecified, the main claims are materially compromised.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

All components are standard ML building blocks (Word2Vec, BoW, RF, CNN, soft voting). No new physical or conceptual entities are introduced beyond the CAPE algorithmic pipeline.

free parameters (6)
  • m (number of selected tokens) = 2000
    The number of most representative words kept per HTML sample. Chosen by the authors without an ablation or sensitivity analysis.
  • Word2Vec embedding dimension = 100
    Dimensionality of token embeddings. Stated in Section 3.1 but no justification or tuning.
  • URL CNN embedding dimension = 16
    Embedding size for the character-level URL model, copied from Opara et al. [30] without tuning.
  • Soft-voting ensemble weights = Not reported
    Grid-searched on validation data (Section 3.2). The resulting weights are not disclosed, so the ensemble is fitted to the validation split.
  • Random Forest hyperparameters = Not reported
    No hyperparameters are given for the RF classifier selected in Experiment 1.
  • BoW vocabulary size M = Not reported
    The vocabulary generated from the top-m tokens is not quantified; the construction of B is underspecified.
assumptions (5)
  • domain assumption Word2Vec embeddings capture semantic similarity of HTML tokens and JavaScript/CSS fragments
    The method trusts that cosine proximity in embedding space identifies representative words for phishing and legitimate pages (Section 3.1).
  • domain assumption Class centroids computed on the training split generalize to unseen pages
    The centroids are used to select test-set tokens, but no distribution-shift analysis is provided.
  • ad hoc to paper A fixed 2,000-word selection preserves enough information for classification
    The choice of m=2000 is not derived from data or theory; it mirrors the truncation length in Opara et al. [30].
  • domain assumption The custom HTML tokenizer correctly isolates tags, attributes, JavaScript, and CSS
    The tokenizer is described only functionally; its correctness and losslessness are not demonstrated.
  • domain assumption The injection attack (prepending 2000 opposite-class words) is representative of adversarial manipulations
    Only one attack pattern is tested; the paper generalizes from it to 'robustness against adversarial manipulations'.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PhishKey: A Novel Centroid-Based Approach for Enhanced Phishing Detection Using Adaptive HTML Component Extraction." pith.science (2026). https://pith.science/paper/Q4QTDZ6G

@misc{pith2026250621106,
  author       = {Pith},
  title        = {Pith review of: PhishKey: A Novel Centroid-Based Approach for Enhanced Phishing Detection Using Adaptive HTML Component Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q4QTDZ6G}},
  note         = {Machine review of arXiv:2506.21106}
}
read the original abstract

Phishing attacks pose a significant cybersecurity threat, evolving rapidly to bypass detection mechanisms and exploit human vulnerabilities. This paper introduces PhishKey to address the challenges of adaptability, robustness, and efficiency. PhishKey is a novel phishing detection method using automatic feature extraction from hybrid sources. PhishKey combines character-level processing with Convolutional Neural Networks (CNN) for URL classification, and a Centroid-Based Key Component Phishing Extractor (CAPE) for HTML content at the word level. CAPE reduces noise and ensures complete sample processing avoiding crop operations on the input data. The predictions from both modules are integrated using a soft-voting ensemble to achieve more accurate and reliable classifications. Experimental evaluations on four state-of-the-art datasets demonstrate the effectiveness of PhishKey. It achieves up to 98.70% F1 Score and shows strong resistance to adversarial manipulations such as injection attacks with minimal performance degradation.

Figures

Figures reproduced from arXiv: 2506.21106 by the authors.

Figure 1
Figure 1. PhishKey components interaction. The red box contain the details of CAPE. [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 39 canonical work pages

  1. [1]

    Abawajy, J., Richard, A., Aghbari, Z.A., 2018. Securing websites against ho- mograph attacks, in: Security and Privacy in Communication Networks: Se- cureComm 2017 International Workshops, ATCS and SePrIoT, Niagara Falls, ON, Canada, October 22–25, 2017, Proceedings 13, Springer. pp. 47–59

    work page 2018

  2. [2]

    An effective detection approach for phishing websites using url and html fea- tures

    Aljofey, A., Jiang, Q., Rasool, A., Chen, H., Liu, W., Qu, Q., Wang, Y., 2022. An effective detection approach for phishing websites using url and html fea- tures. Scientific Reports 12, 8842

    work page 2022

  3. [3]

    Combining long-term recurrent convolutional and graph convolutional networks to detect phishing sites using url and html

    Ariyadasa, S., Fernando, S., Fernando, S., 2022. Combining long-term recurrent convolutional and graph convolutional networks to detect phishing sites using url and html. IEEE Access 10, 82355–82375

    work page 2022

  4. [4]

    Asif, A.U.Z., Shirazi, H., Ray, I., 2023. Machine learning-based phishing detec- tion using url features: A comprehensive review, in: International Symposium on Stabilizing, Safety, and Security of Distributed Systems, Springer. pp. 481– 497

    work page 2023

  5. [5]

    A survey of intelligent detection designs of html url phishing attacks

    Asiri, S., Xiao, Y., Alzahrani, S., Li, S., Li, T., 2023. A survey of intelligent detection designs of html url phishing attacks. IEEE Access 11, 6421–6443

    work page 2023

  6. [6]

    Classifying phishing urls using recurrent neural networks, in: 2017 APWG sym- posium on electronic crime research (eCrime), IEEE

    Bahnsen, A.C., Bohorquez, E.C., Villegas, S., Vargas, J., González, F.A., 2017. Classifying phishing urls using recurrent neural networks, in: 2017 APWG sym- posium on electronic crime research (eCrime), IEEE. pp. 1–8

    work page 2017

  7. [7]

    Bashir, H.M., Li, Q., Hou, J., 2020. A high capacity text steganography utilizing unicode zero-width characters, in: 2020 International Conferences on Internet of Things (iThings) and IEEE Green Computing and Communications (Green- Com) and IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE 20 Smart Data (SmartData) and IEEE Congress on Cybermati...

    work page 2020

  8. [8]

    Text summarization using large language models: a comparative study of mpt-7b-instruct, falcon-7b-instruct, and openai chat-gpt models

    Basyal, L., Sanghvi, M., 2023. Text summarization using large language models: a comparative study of mpt-7b-instruct, falcon-7b-instruct, and openai chat-gpt models. arXiv preprint arXiv:2310.10449

    arXiv 2023

Show all 43 references
  1. [9]

    Text summarization for big data analytics: a comprehensive review of gpt 2 and bert approaches

    Bharathi Mohan, G., Prasanna Kumar, R., Parathasarathy, S., Aravind, S., Hanish, K., Pavithria, G., 2023. Text summarization for big data analytics: a comprehensive review of gpt 2 and bert approaches. Data Analytics for Internet of Things Infrastructure , 247–264

    2023

  2. [10]

    Grambeddings: a new neural network for url based identification of phishing web pages through n-gram em- beddings

    Bozkir, A.S., Dalgic, F.C., Aydos, M., 2023. Grambeddings: a new neural network for url based identification of phishing web pages through n-gram em- beddings. Computers & Security 124, 102964

    2023

  3. [11]

    Phikita: Phishing kit attacks dataset for phishing websites identification

    Castaño, F., Fernañdez, E.F., Alaiz-Rodríguez, R., Alegre, E., 2023. Phikita: Phishing kit attacks dataset for phishing websites identification. IEEE Access 11, 40779–40789

    2023

  4. [12]

    Single document automatic text summarization using term frequency-inverse document frequency (tf-idf)

    Christian, H., Agus, M.P., Suhartono, D., 2016. Single document automatic text summarization using term frequency-inverse document frequency (tf-idf). ComTech: Computer, Mathematics and Engineering Applications 7, 285–294

    2016

  5. [13]

    Tracking phishing attacks over time, in: Proceedings of the 26th International Conference on World Wide Web, pp

    Cui, Q., Jourdan, G.V., Bochmann, G.V., Couturier, R., Onut, I.V., 2017. Tracking phishing attacks over time, in: Proceedings of the 26th International Conference on World Wide Web, pp. 667–676

    2017

  6. [14]

    A multi-layered security model to counter social engineering attacks: a learning-based approach

    Edwards, L., Zahid Iqbal, M., Hassan, M., 2024. A multi-layered security model to counter social engineering attacks: a learning-based approach. International Cybersecurity Law Review , 1–24

    2024

  7. [15]

    Internet organised crime threat assessment (iocta) 2024

    Europol, 2024. Internet organised crime threat assessment (iocta) 2024. Most , 1–12

    2024

  8. [16]

    Decoding the threat landscape: Chatgpt, fraudgpt, and wormgpt in social engineering attacks

    Falade, P.V., 2023. Decoding the threat landscape: Chatgpt, fraudgpt, and wormgpt in social engineering attacks. arXiv preprint arXiv:2310.05595

    2023 arXiv

  9. [17]

    Phishing url detection using url ranking, in: 2015 ieee international congress on big data, IEEE

    Feroz, M.N., Mengel, S., 2015. Phishing url detection using url ranking, in: 2015 ieee international congress on big data, IEEE. pp. 635–638

    2015

  10. [18]

    Phishing activity trends report 3 quarter

    Group, A.P.W., 2024. Phishing activity trends report 3 quarter. 21

    2024

  11. [19]

    Cutting through the confusion: A measurement study of homograph attacks., in: USENIX Annual Technical Conference, General Track, pp

    Holgers, T., Watson, D.E., Gribble, S.D., 2006. Cutting through the confusion: A measurement study of homograph attacks., in: USENIX Annual Technical Conference, General Track, pp. 261–266

    2006

  12. [20]

    The state of phishing attacks

    Hong, J., 2012. The state of phishing attacks. Communications of the ACM 55, 74–81

    2012

  13. [21]

    Phishing detection system through hybrid machine learning based on url

    Karim, A., Shahroz, M., Mustofa, K., Belhaouari, S.B., Joga, S.R.K., 2023. Phishing detection system through hybrid machine learning based on url. IEEE Access 11, 36805–36822

    2023

  14. [22]

    Extractive based text summarization using k-means and tf-idf

    Khan, R., Qian, Y., Naeem, S., 2019. Extractive based text summarization using k-means and tf-idf. International Journal of Information Engineering and Electronic Business 12, 33

    2019

  15. [23]

    A hybrid phishing detection system using deep learning-based url and content analysis

    Korkmaz, M., Kocyigit, E., Sahingoz, O., Diri, B., 2022. A hybrid phishing detection system using deep learning-based url and content analysis. Elektronika ir Elektrotechnika 28

    2022

  16. [24]

    Deep neural net- work based phishing classification on a high-risk url dataset, in: International conference on soft computing and pattern recognition, Springer

    Korkmaz, M., Kocyigit, E., Sahingoz, O.K., Diri, B., 2020. Deep neural net- work based phishing classification on a high-risk url dataset, in: International conference on soft computing and pattern recognition, Springer. pp. 648–657

    2020

  17. [25]

    A stacking model using url and html features for phishing webpage detection

    Li, Y., Yang, Z., Chen, X., Yuan, H., Liu, W., 2019. A stacking model using url and html features for phishing webpage detection. Future Generation Computer Systems 94, 27–39

    2019

  18. [26]

    Phishstorm: Detecting phishing with streaming analytics

    Marchal, S., François, J., State, R., Engel, T., 2014. Phishstorm: Detecting phishing with streaming analytics. IEEE Transactions on Network and Service Management 11, 458–471

    2014

  19. [27]

    Mensah, P., Blanc, G., Okada, K., Miyamoto, D., Kadobayashi, Y., 2015. Ajna: anti-phishing js-based visual analysis, to mitigate users’ excessive trust in ssl/tls, in: 2015 4th International Workshop on Building Analysis Datasets and Gath- ering Experience Returns for Security...

    2015

  20. [28]

    Efficient estimation of word representations in vector space

    Mikolov, T., Chen, K., Corrado, G., Dean, J., 2013. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781

    2013 arXiv

  21. [29]

    Phishmon: A machine learn- ing framework for detecting phishing webpages, in: 2018 IEEE International Conference on Intelligence and Security Informatics (ISI), IEEE

    Niakanlahiji, A., Chu, B.T., Al-Shaer, E., 2018. Phishmon: A machine learn- ing framework for detecting phishing webpages, in: 2018 IEEE International Conference on Intelligence and Security Informatics (ISI), IEEE. pp. 220–225. 22

    2018

  22. [30]

    Look before you leap: Detecting phishing web pages by exploiting raw url and html characteristics

    Opara, C., Chen, Y., Wei, B., 2024. Look before you leap: Detecting phishing web pages by exploiting raw url and html characteristics. Expert Systems with Applications 236, 121183

    2024

  23. [31]

    Htmlphish: Enabling phishing web page detection by applying deep learning techniques on html analysis, in: 2020 In- ternational Joint Conference on Neural Networks (IJCNN), IEEE

    Opara, C., Wei, B., Chen, Y., 2020. Htmlphish: Enabling phishing web page detection by applying deep learning techniques on html analysis, in: 2020 In- ternational Joint Conference on Neural Networks (IJCNN), IEEE. pp. 1–8

    2020

  24. [32]

    Phiusiil: A diverse security profile empowered phishing url detection framework based on similarity index and incremental learning

    Prasad, A., Chandra, S., 2024. Phiusiil: A diverse security profile empowered phishing url detection framework based on similarity index and incremental learning. Computers & Security 136, 103545

    2024

  25. [33]

    Comparative assessment of extractive summariza- tion: textrank tf-idf and lda

    Rani, U., Bidhan, K., 2021. Comparative assessment of extractive summariza- tion: textrank tf-idf and lda. Journal of scientific research 65, 304–311

    2021

  26. [34]

    Catchphish: detection of phishing websites by inspecting urls

    Rao, R.S., Vaishnavi, T., Pais, A.R., 2020. Catchphish: detection of phishing websites by inspecting urls. Journal of Ambient Intelligence and Humanized Computing 11, 813–825

    2020

  27. [35]

    Centroid-based text summarization through compositionality of word embeddings

    Rossiello, G., Basile, P., Semeraro, G., 2017. Centroid-based text summarization through compositionality of word embeddings

    2017

  28. [36]

    Machine learning based phishing detection from urls

    Sahingoz, O.K., Buber, E., Demir, O., Diri, B., 2019. Machine learning based phishing detection from urls. Expert Systems with Applications 117, 345–357

    2019

  29. [37]

    Dephides: Deep learning based phishing detection system

    Sahingoz, O.K., Buber, E., Kugu, E., 2024. Dephides: Deep learning based phishing detection system. IEEE Access

    2024

  30. [38]

    Phishing websites detection using a novel multipurpose dataset and web tech- nologies features

    Sánchez-Paniagua, M., Fidalgo, E., Alegre, E., Alaiz-Rodríguez, R., 2022. Phishing websites detection using a novel multipurpose dataset and web tech- nologies features. Expert Systems with Applications 207, 118010

    2022

  31. [39]

    Sánchez-Paniagua, M., Fidalgo, E., González-Castro, V., Alegre, E., 2021. Im- pact of current phishing strategies in machine learning models for phishing de- tection, in: 13th International Conference on Computational Intelligence in Se- curity for Information Systems (CISIS 2...

    2021

  32. [40]

    Get to the point: Summarization with pointer-generator networks

    See, A., Liu, P.J., Manning, C.D., 2017. Get to the point: Summarization with pointer-generator networks. arXiv preprint arXiv:1704.04368

    2017 arXiv

  33. [41]

    Http header based phishing attack detection using machine learning

    Shukla, S., Misra, M., Varshney, G., 2024. Http header based phishing attack detection using machine learning. Transactions on Emerging Telecommunica- tions Technologies 35, e4872. 23

    2024

  34. [42]

    Sequence level contrastive learning for text summarization, in: Proceedings of the AAAI conference on artificial intelligence, pp

    Xu, S., Zhang, X., Wu, Y., Wei, F., 2022. Sequence level contrastive learning for text summarization, in: Proceedings of the AAAI conference on artificial intelligence, pp. 11556–11565

    2022

  35. [43]

    A survey on large language model (llm) security and privacy: The good, the bad, and the ugly

    Yao, Y., Duan, J., Xu, K., Cai, Y., Sun, Z., Zhang, Y., 2024. A survey on large language model (llm) security and privacy: The good, the bad, and the ugly. High-Confidence Computing , 100211. 24

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.