Pith. sign in

REVIEW 3 major objections 7 minor 1 cited by

WebGuard++:Interpretable Malicious URL Detection via Bidirectional Fusion of HTML Subgraphs and Multi-Scale Convolutional BERT

T0 review · 3 major / 7 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 WebGuard++ claims that splitting the HTML DOM into subgraphs and fusing them bidirectionally with multi-scale URL features lifts true-positive rates by 1.1x to 7.9x at fixed false-positive rates of 0.001 and 0.0001.

desk verdict The headline TPR@FPR gains are not computable from the declared test sets, so the central empirical claim collapses; the subgraph-voting idea is salvageable but the paper needs a full re-evaluation. read the letter →

arxiv 2506.19356 v1 pith:ZJNQXOPH submitted 2025-06-24 cs.CR cs.LG

Ye Tian , Zhang Yumin , Yifan Jia , Jianguo Sun , Yanbin Wang This is my paper
classification cs.CRcs.LG
keywords maliciousURLdetectionphishingHTMLsubgraphlearningbidirectionalfusionConvBERTvotingmechanisminterpretabilitylowfalse-positiverate
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

The paper sets out to show that combining URL text with the page's HTML structure, at both fine and coarse scales, gives a malicious-page detector that works better than URL-only or HTML-only systems. Its model, WebGuard++, reads URLs through a ConvBERT encoder with pyramid attention, splits the DOM tree into hash-partitioned subgraphs so that sparse malicious nodes are not washed out by benign ones, and connects the two modalities with bidirectional cross-attention. A biased voting module flags a page when sampled subgraph batches agree on maliciousness, and in doing so localizes the suspicious DOM regions. The reported payoff is a 1.1x to 7.9x higher true-positive rate at fixed false-positive rates of 0.001 and 0.0001, which would matter for defending users at web scale without flooding them with false alarms.

What carries the argument

The load-bearing mechanism is the coupling of subgraph-aware HTML encoding with a biased voting module. The DOM is parsed into a graph; nodes are assigned to $T_f$ groups by a hash function; batches of subgraphs are sampled and processed through block-diagonal message passing and pooling, so malicious nodes can reinforce each other without being diluted by dominant benign content. The bidirectional coupling module stacks self- and cross-attention layers so URL semantics and HTML structure exchange information in both directions. Finally, the biased voting policy declares a URL-HTML pair malicious when more than one of the sampled subgraph rounds predict malicious, and the winning subgraphs are returned as the localized evidence.

What would settle it

Count the negative samples in the test split and inspect the threshold schedule used for the ROC curve; if the test set has N negatives, no threshold can produce a false-positive rate below 1/N without tied scores or resampling, so a reported TPR at FPR=0.0001 for the Abdelhakim set (with roughly 117 negatives) would need an interpolation procedure that the paper does not state. A reader could rerun the published model and check whether the ROC curve actually crosses FPR=0.0001 and 0.001, or whether those points are extrapolations.

Watch

Extended reading notes

Core claim

On its own terms, the paper discovers that subgraph-level rather than whole-graph processing of HTML DOMs preserves the weak signal of malicious nodes, and that bidirectional URL-HTML attention lets each modality disambiguate the other. Concretely, WebGuard++ reports TPR@FPR(0.001) of 0.7939 on MTLP and 0.3802 on Abdelhakim, and TPR@FPR(0.0001) of 0.6848 and 0.3802 respectively, outperforming all listed baselines by 1.1x to 7.9x at those operating points. The same voting mechanism also yields component-level attribution, because the subgraphs that trigger the malicious vote are themselves the evidence.

Load-bearing premise

The headline numbers assume that true-positive rate can be reliably measured at false-positive rates of 0.0001 and 0.001, but the test sets contain only about 1010 (MTLP) and 117 (Abdelhakim) negative samples, so the finest observable nonzero FPR is about 0.001 and 0.0085; the paper does not explain how thresholds below that resolution were obtained.

Editorial extensions

If this is right

  • At a fixed false-positive rate of 0.001 or 0.0001, WebGuard++ would block more malicious pages than URL-only or HTML-only systems, reducing missed attacks during high-volume crawling.
  • Because the voting module operates on subgraphs, every flagged page comes with a shortlist of suspicious DOM regions, giving analysts a starting point for triage without needing a separate explanation model.
  • Subgraph-based encoding prevents the signal of rare malicious nodes (e.g., obfuscated scripts, fake forms) from being averaged away by abundant benign nodes, which is the property that makes the low-FPR gains possible.
  • Bidirectional coupling lets an ambiguous URL (e.g., a lookalike domain) be overruled or confirmed by structural anomalies in the HTML, and vice versa, so evasion that targets only one modality is less likely to succeed.
  • On cross-domain data, the model reached near-perfect accuracy within two epochs while baselines needed five or more, suggesting faster adaptation when deployed on new distributions.

Reading between the lines

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

  • The same partition-and-vote recipe could be applied to other graph-structured abuse signals, such as malicious JavaScript abstract syntax trees or email HTML bodies, where suspicious nodes are also sparse relative to benign content.
  • Because the reported TPR@FPR figures may depend on an unstated interpolation below the test set's resolution, a natural extension is to rerun the evaluation on a much larger negative sample or to report interpolated ROC-AUC with confidence intervals instead of point values.
  • If subgraph attribution is stable across runs, the voting module could generate pseudo-labels for training a cheaper single-pass detector that does not need the multi-round voting loop at inference time.
  • The robustness test uses random edge deletion only; a stronger check would be adversarial URL perturbation or DOM restructuring, which the paper does not yet cover.
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

3 major / 7 minor

Summary. WebGuard++ proposes a fused URL/HTML detector for malicious webpages. It combines a ConvBERT-based cross-scale URL encoder with spatial pyramid attention, a subgraph-aware GNN HTML encoder that partitions DOM graphs and pools subgraph-level features, a bidirectional cross-attention coupling module, and a biased voting mechanism that aggregates predictions over randomly sampled subgraphs to provide component-level attribution. The authors evaluate on the MTLP and Abdelhakim phishing datasets, compare against several URL-only and multimodal baselines, and report strong gains at very low false-positive rates, plus ablation, cross-dataset, and robustness experiments.

Significance. If the reported results were reliable, the method would be a meaningful step toward phishing detection at very low false-positive rates with interpretable subgraph-level evidence. The architecture is clearly described, the components are modular, and the use of public datasets is a strength. However, the central quantitative claim—1.1x–7.9x higher TPR at FPR=0.0001 and 0.001—is not supported by the reported evaluation setup, because the test sets are far too small to resolve those operating points. In addition, the cross-dataset generalization experiment uses an unrelated course/non-course labeling task, and the coverage probability calculation for the voting mechanism is mathematically inconsistent with the stated voting rule. These are load-bearing issues, so the paper cannot be accepted in its current form.

major comments (3)
  1. [Section IV-A, Tables I–IV] The reported test sets cannot realize the claimed TPR@FPR operating points. In Table I, the WebGuard++ row (TN=989, FP=21, FN=24, TP=966) sums to 2000, not the 10,000 stated in the table header and in Section IV-A; the number of negatives is 1010, so the smallest nonzero FPR is 1/1010 ≈ 0.00099, which is already above 0.0001. In Table II, the WebGuard++ row (TN=109, FP=8, FN=12, TP=59) sums to 188, giving 117 negatives and a smallest nonzero FPR of 1/117 ≈ 0.00855; therefore no threshold on this test set can produce FPR=0.001, let alone FPR=0.0001. The paper does not describe any interpolation, score smoothing, bootstrapping, or alternative test set that would estimate sub-resolution operating points. If the authors intend the one-sided convention TPR at FPR≤x, that convention is never stated and would still leave the Abdelhakim FPR=0.001 column equal to the zero-FP operating point for every method, which is not the 'fixed FPR' claim made in the abstract. As written, Tables III and IV and the headline 1.1x–7.9x improvement are unsupported.
  2. [Section IV-A and Section IV-B3, Figure 5] The cross-dataset generalization test is not a phishing generalization test. The course-cotrain-data dataset is described in Section IV-A as a course categorization dataset with 230 course and 821 non-course pages from the Web->Kb collection, and Section IV-B3 reports accuracy, recall, ROC-AUC, PR-AUC, and TPR@FPR on this data. The labels are course versus non-course, not malicious versus benign, so near-perfect performance on this task provides no evidence that WebGuard++ generalizes to malicious URL detection. This experiment should be removed or replaced with a genuinely related phishing benchmark; as it stands, the generalization claim in the conclusion is not supported.
  3. [Section III-D, Eqs. (21)–(23), Algorithm 1] The coverage probability argument does not match the voting rule. With num_group=5, iter_per=5, iter_num=4, and one malicious subgraph, the probability that one round misses the malicious subgraph is 20%, and the probability that all five rounds miss it is (0.2)^5 = 0.032%. However, Algorithm 1 declares the page malicious only when 1count ≥ 2, i.e., when at least two of the five rounds contain a malicious prediction. The probability of at least two rounds containing the malicious subgraph is 1 − (0.2)^5 − 5(0.2)^4(0.8) ≈ 0.9933, not the 0.99968 implied by the text's 'basically extract subgraphs with full coverage' statement. The paper conflates 'at least one round contains the malicious subgraph' with 'the voting rule fires,' so the theoretical justification for the voting mechanism is incorrect as stated.
minor comments (7)
  1. [Section IV-B1, text after Table I] The sentence 'advanced models such as PMANet and TransURL outperform our model in F1, MCC, and low FPR metrics' contradicts Table I, where WebGuard++ has higher F1 (0.9772 versus 0.9756 for PMANet and 0.9699 for TransURL) and higher MCC (0.9549 versus 0.9519 and 0.9391); this sentence should be corrected.
  2. [Tables I and II] Several table entries are malformed or numerically inconsistent: 'URLBERT 0.96100.98950.97500.9965' lacks separators, SEMI-GAN's accuracy appears as '09186' instead of '0.9186', and the row sums for BILSTM, URLNET, and SEMI-GAN in Table I do not equal 2000; the authors should verify all arithmetic and align the table headers with the actual sample sizes.
  3. [Algorithm 1] The variable names '0count' and '1count' are not valid identifiers in most programming languages and are hard to read; renaming them to, for example, benign_count and malicious_count would improve clarity.
  4. [Section IV-A, Abdelhakim Dataset] The text states that the Abdelhakim Dataset contains 11,430 URLs, but every row in Table II sums to 188 (or 181 for SEMI-GAN); please clarify the exact evaluation subset size and the train/test split.
  5. [Figure 1] Figure 1 contains untranslated Chinese text ('专家特征') in the diagram; an English version should be provided.
  6. [Section III-D, Eq. (21)] The expression C(4,3)/C(5,4) is an unnecessarily roundabout way to write 4/5; using the simplified form would make the intended probability clearer.
  7. [General, interpretability claim] The paper claims component-level traceability as a contribution, but no experiment quantifies localization accuracy (for example, overlap between flagged subgraphs and ground-truth malicious DOM regions); either add such an evaluation or soften the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the architecture is described directly and the low-FPR figures are a measurement-validity concern, not a construction-level circularity.

full rationale

The paper's derivation chain is self-contained rather than circular. The URL encoder, HTML subgraph encoder, coupling module, and voting rule are specified by the paper's own equations (Eqs. 1-20 and Algorithm 1); they are not obtained by fitting a parameter to the target metric and then renaming it a prediction. The coverage calculation in Eqs. (21)-(23) is an a priori combinatorial argument over assumed group sizes (num_group=5, iter_per=5, iter_num=4) and is not a fitted or predicted result. The evaluation uses public datasets (MTLP and Abdelhakim) with external baselines. Although several baselines and related works (TransURL, PMANet, URLBERT, Fed-urlBERT, PhishBERT) are by the same group, none is used as a load-bearing justification: the paper never invokes a uniqueness theorem or an author-imported constraint to force its design. The reported TPR@FPR(0.0001/0.001) values are statistically questionable because the declared test sets contain only about 1010 and 117 negative samples, making FPR=0.0001 unattainable at any threshold; however, this is an empirical validity gap, not circularity, since the values are not equivalent by construction to the inputs. Therefore no circular step is identified and the score is 0.

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

The paper contributes an architecture, not a derivation. The central claim depends on unverified data-processing choices (cleaning and subsampling, hash partitioning) and hand-set voting parameters, plus several modeling assumptions about embeddings and task transfer that are not independently validated.

free parameters (6)
  • Voting threshold (1count threshold) = 2
    Algorithm 1 (line 21) classifies as malicious only when at least 2 of the sampled rounds predict malicious; prose in Section III-D says any malicious subgraph (>=1) triggers classification. This hand-set threshold directly controls the reported TPR/FPR trade-off.
  • num_group (number of hash subgraphs per HTML graph) = 5 (example)
    Used in the Section III-D probability argument (Eq 21-23); the value used in the experiments is not reported.
  • iter_per (number of voting rounds) = 5 (example)
    Used in the same probability argument; the experimental value is unreported.
  • iter_num (subgraphs sampled per round) = 4 (example)
    Used in the same probability argument; the experimental value is unreported.
  • DSConv expansion rates d1-d4
    Spatial pyramid branch rates in Eq (2); the values are never specified, though the branch outputs are summed in Eq (3).
  • MTLP cleaning and 10,000-sample even subsample = 10,000
    Section IV-A states 'we cleaned the MTLP Dataset and evenly sampled 10,000 data'; the cleaning is not described and the confusion matrix in Table I sums to 2000.
assumptions (5)
  • domain assumption Random hash-based node partitioning into subgraphs preserves malicious signal after per-subgraph GNN aggregation
    Section III-B, Eq (11)-(16): the paper assumes threat-indicative nodes form detectable subgraphs and are not diluted; no edge-preservation analysis or comparison to full-graph GNN is provided.
  • ad hoc to paper A page is malicious if some sampled subgraph is predicted malicious, and the 5-round/4-subgraph sampling covers the malicious subgraph with probability near 1
    Section III-D, Eq (21)-(23): the calculation assumes exactly one malicious subgraph among num_group=5 and uses the >=1 rule while Algorithm 1 uses >=2.
  • domain assumption CharBERT and ConvBERT embeddings, plus Word2Vec node embeddings, provide sufficient features for the detection task
    Section III-A and III-B: pretrained models are invoked without specifying initialization, fine-tuning, or the CharBERT variant; no citation is given for CharBERT.
  • ad hoc to paper The Web->Kb course categorization labels are a valid target for cross-dataset generalization of a phishing detector
    Section IV-B3: the course-cotrain-data task (course vs non-course) is unrelated to phishing, yet it is used to claim generalization.
  • standard math Standard probability and attention formulas are correct as stated
    Eqs (17)-(20) and Eq (21)-(23) use standard formulas, but the voting threshold inconsistency means the probability argument does not match the algorithm.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WebGuard++:Interpretable Malicious URL Detection via Bidirectional Fusion of HTML Subgraphs and Multi-Scale Convolutional BERT." pith.science (2026). https://pith.science/paper/ZJNQXOPH

@misc{pith2026250619356,
  author       = {Pith},
  title        = {Pith review of: WebGuard++:Interpretable Malicious URL Detection via Bidirectional Fusion of HTML Subgraphs and Multi-Scale Convolutional BERT},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZJNQXOPH}},
  note         = {Machine review of arXiv:2506.19356}
}
read the original abstract

URL+HTML feature fusion shows promise for robust malicious URL detection, since attacker artifacts persist in DOM structures. However, prior work suffers from four critical shortcomings: (1) incomplete URL modeling, failing to jointly capture lexical patterns and semantic context; (2) HTML graph sparsity, where threat-indicative nodes (e.g., obfuscated scripts) are isolated amid benign content, causing signal dilution during graph aggregation; (3) unidirectional analysis, ignoring URL-HTML feature bidirectional interaction; and (4) opaque decisions, lacking attribution to malicious DOM components. To address these challenges, we present WebGuard++, a detection framework with 4 novel components: 1) Cross-scale URL Encoder: Hierarchically learns local-to-global and coarse to fine URL features based on Transformer network with dynamic convolution. 2) Subgraph-aware HTML Encoder: Decomposes DOM graphs into interpretable substructures, amplifying sparse threat signals via Hierarchical feature fusion. 3) Bidirectional Coupling Module: Aligns URL and HTML embeddings through cross-modal contrastive learning, optimizing inter-modal consistency and intra-modal specificity. 4) Voting Module: Localizes malicious regions through consensus voting on malicious subgraph predictions. Experiments show WebGuard++ achieves significant improvements over state-of-the-art baselines, achieving 1.1x-7.9x higher TPR at fixed FPR of 0.001 and 0.0001 across both datasets.

Figures

Figures reproduced from arXiv: 2506.19356 by the authors.

Figure 1
Figure 1. Framework diagram of the model structure of WebGuard++. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Biased voting mechanism process. Where B denotes the batch size, G1 represents the 1 st subgraph, l denotes the l th propagation, H(l) denotes the hidden representation of each layer, Ablockdenotes the block diagonal sparse matrix. Finally, all sub-graphs of each layer are pooled at the sub￾graph level, and the features of each layer are concatenated to obtain the final features of these sub-graphs. Hpooled (i) = P … view at source ↗
Figure 3
Figure 3. Performance of our model against other models on the ROC curve. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The ablation experiments of our model against the voting mechanism were done on the MTLP dataset (uniformly sampled 2000, 7000, and 10000 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The generalisation performance of BiLSTM, TextCNN, TransURL, URLNet, and the proposed model in this paper in cross-dataset scenarios is visually [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Compare the robustness test of our model and the basic GNN model in the random edge deletion scenario. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Breaking Obfuscation: Cluster-Aware Graph with LLM-Aided Recovery for Malicious JavaScript Detection

    cs.CR 2025-07 conditional novelty 4.0 of 10

    DeCoda combines LLM-based JavaScript deobfuscation with a cluster-aware graph transformer, reporting F1 scores of 94.6% and 97.7% on two malicious-code datasets.

Reference graph

Works this paper leans on

39 extracted references · 36 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ali Aljofey, Saifullahi Aminu Bello, Jian Lu, and Chen Xu. Com- prehensive phishing detection: A multi-channel approach with variants tcn fusion leveraging url and html features.Journal of Network and Computer Applications, 238:104170, 2025

    work page 2025

  2. [2]

    Adaptpud: An accurate url-based detection approach against tailored deceptive phishing websites.Computer Networks, page 111303, 2025

    Zilaing Zhang, Jinmin Wu, Ning Lu, Wenbo Shi, and Zhiquan Liu. Adaptpud: An accurate url-based detection approach against tailored deceptive phishing websites.Computer Networks, page 111303, 2025

    work page 2025

  3. [3]

    Anti-phishing based on automated individual white-list

    Ye Cao, Weili Han, and Yueran Le. Anti-phishing based on automated individual white-list. InProceedings of the 4th ACM Workshop on Digital Identity Management, DIM ’08, page 51–60, New York, NY , USA, 2008. Association for Computing Machinery

    work page 2008

  4. [4]

    Adopting au- tomated whitelist approach for detecting phishing attacks.Computers & Security, 108:102328, 2021

    Nureni Ayofe Azeez, Sanjay Misra, Ihotu Agbo Margaret, Luis Fernandez-Sanz, and Shafi’i Muhammad Abdulhamid. Adopting au- tomated whitelist approach for detecting phishing attacks.Computers & Security, 108:102328, 2021

    work page 2021

  5. [5]

    New rule-based phishing detection method.Expert Systems with Applications, 53:231–242, 2016

    Mahmood Moghimi and Ali Yazdian Varjani. New rule-based phishing detection method.Expert Systems with Applications, 53:231–242, 2016

    work page 2016

  6. [6]

    Sustaining accurate detection of phishing urls using sdn and feature selection approaches.Computer Networks, 201:108591, 2021

    Raniyah Wazirali, Rami Ahmad, and Ashraf Abdel-Karim Abu-Ein. Sustaining accurate detection of phishing urls using sdn and feature selection approaches.Computer Networks, 201:108591, 2021

    work page 2021

  7. [7]

    Doyen Sahoo, Chenghao Liu, and Steven C. H. Hoi. Malicious url detection using machine learning: A survey, 2019

    work page 2019

  8. [8]

    Phishdef: Url names say it all

    Anh Le, Athina Markopoulou, and Michalis Faloutsos. Phishdef: Url names say it all. In2011 Proceedings IEEE INFOCOM, pages 191–195. IEEE, 2011

    work page 2011

Show all 39 references
  1. [9]

    Phishzoo: Detecting phishing websites by looking at them

    Sadia Afroz and Rachel Greenstadt. Phishzoo: Detecting phishing websites by looking at them. In2011 IEEE fifth international conference on semantic computing, pages 368–375. IEEE, 2011

    2011

  2. [10]

    Machine learning based phishing detection from urls.Expert Systems with Applications, 117:345–357, 2019

    Ozgur Koray Sahingoz, Ebubekir Buber, Onder Demir, and Banu Diri. Machine learning based phishing detection from urls.Expert Systems with Applications, 117:345–357, 2019

    2019

  3. [11]

    A new method for detection of phishing websites: Url detection

    Shraddha Parekh, Dhwanil Parikh, Srushti Kotak, and Smita Sankhe. A new method for detection of phishing websites: Url detection. In 2018 Second International Conference on Inventive Communication and Computational Technologies (ICICCT), pages 949–952, 2018

    2018

  4. [12]

    Toward more generalized malicious url detection models, 2024

    YunDa Tsai, Cayon Liow, Yin Sheng Siang, and Shou-De Lin. Toward more generalized malicious url detection models, 2024

    2024

  5. [13]

    Urlnet: Learning a url representation with deep learning for malicious url detection.arXiv preprint arXiv:1802.03162, 2018

    Hung Le, Quang Pham, Doyen Sahoo, and Steven CH Hoi. Urlnet: Learning a url representation with deep learning for malicious url detection.arXiv preprint arXiv:1802.03162, 2018

    2018 arXiv

  6. [14]

    Phishing url detection using deep learning with cnn models

    Alsadig Hadi Alsadig and Md Oqail Ahmad. Phishing url detection using deep learning with cnn models. In2024 Second International Conference on Intelligent Cyber Physical Systems and Internet of Things (ICoICI), pages 768–775. IEEE, 2024

    2024

  7. [15]

    Transurl: Improving malicious url detection with multi-layer transformer encoding and multi-scale pyramid features

    Ruitong Liu, Yanbin Wang, Zhenhao Guo, Haitao Xu, Zhan Qin, Wenrui Ma, and Fan Zhang. Transurl: Improving malicious url detection with multi-layer transformer encoding and multi-scale pyramid features. Computer Networks, 253:110707, 2024

    2024

  8. [16]

    Phishguard: A convolutional neural network-based model for detecting phishing urls with explain- ability analysis

    Md Robiul Islam, Md Mahamodul Islam, Mst Suraiya Afrin, Anika Antara, Nujhat Tabassum, and Al Amin. Phishguard: A convolutional neural network-based model for detecting phishing urls with explain- ability analysis. In2024 3rd International Conference on Artificial Intelligence...

    2024

  9. [17]

    Fed-urlbert: Client-side lightweight federated transformers for url threat analysis, 2023

    Yujie Li, Yanbin Wang, Haitao Xu, Zhenhao Guo, Fan Zhang, Ruitong Liu, and Wenrui Ma. Fed-urlbert: Client-side lightweight federated transformers for url threat analysis, 2023

    2023

  10. [18]

    A large-scale pretrained deep model for phishing url detection

    Yanbin Wang, Weifan Zhu, Haitao Xu, Zhan Qin, Kui Ren, and Wenrui Ma. A large-scale pretrained deep model for phishing url detection. InICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, 2023

    2023

  11. [19]

    KnowPhish: Large language mod- els meet multimodal knowledge graphs for enhancing Reference-Based phishing detection

    Yuexin Li, Chengyu Huang, Shumin Deng, Mei Lin Lock, Tri Cao, Nay Oo, Hoon Wei Lim, and Bryan Hooi. KnowPhish: Large language mod- els meet multimodal knowledge graphs for enhancing Reference-Based phishing detection. In33rd USENIX Security Symposium (USENIX Security 24), page...

    2024

  12. [20]

    Multiphish: Multi-modal features fusion networks for phishing detection

    Lei Zhang, Peng Zhang, Luchen Liu, and Jianlong Tan. Multiphish: Multi-modal features fusion networks for phishing detection. InICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3520–3524, 2021

    2021

  13. [21]

    Kavya and D

    S. Kavya and D. Sumathi. Multimodal and temporal graph fusion framework for advanced phishing website detection.IEEE Access, 13:74128–74146, 2025

    2025

  14. [22]

    Less defined knowledge and more true alarms: Reference-based phishing detection without a pre-defined reference list

    Ruofan Liu, Yun Lin, Xiwen Teoh, Gongshen Liu, Zhiyong Huang, and Jin Song Dong. Less defined knowledge and more true alarms: Reference-based phishing detection without a pre-defined reference list. In33rd USENIX Security Symposium (USENIX Security 24), pages 523– 540, Philade...

    2024

  15. [23]

    Haijun Zhang, Gang Liu, Tommy W. S. Chow, and Wenyin Liu. Textual and visual content-based anti-phishing: A bayesian approach.IEEE Transactions on Neural Networks, 22(10):1532–1546, 2011

    2011

  16. [24]

    Optimal personalized filtering against spear-phishing attacks

    Aron Laszka, Yevgeniy V orobeychik, and Xenofon Koutsoukos. Optimal personalized filtering against spear-phishing attacks. InProceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, AAAI’15, page 958–964. AAAI Press, 2015

    2015

  17. [25]

    Robust detection of malicious urls with self-paced wide & deep learning.IEEE Transactions on Dependable and Secure Computing, 19(2):717–730, 2022

    Yunji Liang, Qiushi Wang, Kang Xiong, Xiaolong Zheng, Zhiwen Yu, and Daniel Zeng. Robust detection of malicious urls with self-paced wide & deep learning.IEEE Transactions on Dependable and Secure Computing, 19(2):717–730, 2022

    2022

  18. [26]

    Beyond phish: Toward detecting fraudulent e-commerce websites at scale

    Marzieh Bitaab, Haehyun Cho, Adam Oest, Zhuoer Lyu, Wei Wang, Jorij Abraham, Ruoyu Wang, Tiffany Bao, Yan Shoshitaishvili, and Adam Doupé. Beyond phish: Toward detecting fraudulent e-commerce websites at scale. In2023 IEEE Symposium on Security and Privacy (SP), pages 2566–2583, 2023

    2023

  19. [27]

    Zhen Guo, Jin-Hee Cho, Ing-Ray Chen, Srijan Sengupta, Michin Hong, and Tanushree Mitra. Safer: Social capital-based friend recommendation to defend against phishing attacks.Proceedings of the International AAAI Conference on Web and Social Media, 16(1):241–252, May 2022

    2022

  20. [28]

    Phishing webpage de- tection via multi-modal integration of html dom graphs and url features based on graph convolutional and transformer networks.Electronics, 13(16):3344, 2024

    Jun-Ho Yoon, Seok-Jun Buu, and Hae-Jung Kim. Phishing webpage de- tection via multi-modal integration of html dom graphs and url features based on graph convolutional and transformer networks.Electronics, 13(16):3344, 2024

    2024

  21. [29]

    Multimodal large language models for phishing webpage detection and identification.arXiv preprint arXiv:2408.05941, 2024

    Jehyun Lee, Peiyuan Lim, Bryan Hooi, and Dinil Mon Divakaran. Multimodal large language models for phishing webpage detection and identification.arXiv preprint arXiv:2408.05941, 2024

    2024 arXiv

  22. [30]

    Phishagent: a robust multimodal agent for phishing webpage detection

    Tri Cao, Chengyu Huang, Yuexin Li, Wang Huilin, Amy He, Nay Oo, and Bryan Hooi. Phishagent: a robust multimodal agent for phishing webpage detection. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 27869–27877, 2025

    2025

  23. [31]

    Phishing detection model integrating url characters and html word semantic deep features

    Lihui Meng, Zhujuan Ma, and Erzhou Zhu. Phishing detection model integrating url characters and html word semantic deep features. In 2024 4th International Conference on Communication Technology and Information Technology (ICCTIT), pages 468–473, 2024

    2024

  24. [32]

    Ali Babar, Raj Gaire, and Alsharif Abuadbba

    Bushra Sabir, M. Ali Babar, Raj Gaire, and Alsharif Abuadbba. Reli- ability and robustness analysis of machine learning based phishing url detectors.IEEE Transactions on Dependable and Secure Computing, pages 1–18, 2022

    2022

  25. [33]

    Span: Spatial pyramid attention network for image manipulation localization

    Xuefeng Hu, Zhihan Zhang, Zhenye Jiang, Syomantak Chaudhuri, Zhenheng Yang, and Ram Nevatia. Span: Spatial pyramid attention network for image manipulation localization. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors,Computer Vi- sion – ECCV 2020...

    2020

  26. [34]

    Grounding dino: Marrying dino with grounded pre- training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding dino: Marrying dino with grounded pre- training for open-set object detection. In Aleš Leonardis, Elisa Ricci, Stefan Roth, O...

    2024

  27. [35]

    Transfer learning for phishing detection: Screenshot-based website classification

    Furkan Çolhak, Mert ˙Ilhan Ecevit, and Hasan Da ˘g. Transfer learning for phishing detection: Screenshot-based website classification. In2024 9th International Conference on Computer Science and Engineering (UBMK), pages 1–6, 2024

    2024

  28. [36]

    Pmanet: Malicious url detection via post- trained language model guided multi-level feature attention network

    Ruitong Liu, Yanbin Wang, Haitao Xu, Zhan Qin, Fan Zhang, Yiwei Liu, and Zheng Cao. Pmanet: Malicious url detection via post- trained language model guided multi-level feature attention network. Information Fusion, 113:102638, 2025

    2025

  29. [37]

    Urlbert:a contrastive and adversarial pre-trained model for url classification, 2024

    Yujie Li, Yanbin Wang, Haitao Xu, Zhenhao Guo, Zheng Cao, and Lun Zhang. Urlbert:a contrastive and adversarial pre-trained model for url classification, 2024

    2024

  30. [38]

    Semi- supervised conditional gan for simultaneous generation and detec- tion of phishing urls: A game theoretic perspective.arXiv preprint arXiv:2108.01852, 2021

    Sharif Amit Kamran, Shamik Sengupta, and Alireza Tavakkoli. Semi- supervised conditional gan for simultaneous generation and detec- tion of phishing urls: A game theoretic perspective.arXiv preprint arXiv:2108.01852, 2021

    2021 arXiv

  31. [39]

    Dephides: Deep learning based phishing detection system.IEEE Access, 12:8052– 8070, 2024

    Ozgur Koray Sahingoz, Ebubekir BUBEr, and Emin Kugu. Dephides: Deep learning based phishing detection system.IEEE Access, 12:8052– 8070, 2024

    2024

Pith tools

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