A Research Proposal: Machine Learning Approaches for the Early Detection of Phishing Websites

Written by an AHC subject expert in Computer Science. This is a sample research proposal provided for study and reference. Distinction / First-class standard.

Abstract

Phishing remains one of the most persistent and damaging forms of online fraud, exploiting human trust and the visual familiarity of legitimate web pages to harvest credentials, financial data and personal information. Traditional defences, particularly blacklists, react only after a fraudulent site has been reported, leaving a critical window during which newly registered phishing domains operate undetected. This proposal sets out a plan to investigate supervised machine learning as a means of detecting phishing websites earlier than reputation-based systems allow, using features that can be extracted from a URL and its associated page without waiting for a site to be reported. The study will compare four families of classifier, logistic regression, random forest, support vector machines and a feedforward neural network, on a publicly available, labelled phishing dataset. Models will be evaluated using precision, recall, the F1 score and the area under the ROC curve, with stratified cross-validation to guard against optimistic estimates. The intended contribution is a rigorous, reproducible comparison of these approaches under a consistent experimental protocol, together with an analysis of which feature groups contribute most to accurate and timely detection. The work is scoped for completion within a single Master’s dissertation cycle and raises no significant ethical risks, as it relies only on aggregated, anonymised, publicly released data.

Background and Rationale

Phishing attacks impersonate trusted organisations, typically banks, payment providers, government services or well-known technology brands, to deceive users into disclosing sensitive information. The attack surface is broad and the economic cost is considerable: industry monitoring bodies such as the Anti-Phishing Working Group have consistently recorded hundreds of thousands of unique phishing sites per quarter in recent years (APWG, 2023). Because attackers register new domains continuously and abandon them quickly once detected, the operational lifetime of a phishing site is often measured in hours rather than days.

The dominant industrial defence is the blacklist, a curated list of known malicious URLs distributed to browsers and email gateways. Blacklists are effective against previously reported sites but are inherently reactive; a freshly created phishing page will not appear on any list until it has been observed, submitted and verified. This latency is precisely the vulnerability that attackers exploit. Sheng et al. (2009) demonstrated that at the moment a phishing campaign launches, blacklist coverage is limited, with a substantial proportion of fresh URLs unrecognised in the first hours after appearance. The problem is compounded by modern attacker tooling: phishing kits automate the mass registration of look-alike domains, and services that generate short-lived subdomains or abuse legitimate hosting platforms allow a single campaign to rotate through many distinct URLs. Each rotation resets the reporting clock, so a defence that depends on prior observation is structurally disadvantaged.

Machine learning offers a complementary, proactive approach. Rather than checking whether a URL has been seen before, a trained classifier inspects measurable properties of the URL and page, for example the length and structure of the address, the use of an IP address in place of a domain name, the presence of suspicious characters, or the mismatch between a link’s anchor text and its destination, and predicts whether the site is likely to be phishing. Because these features are available at the moment a page is encountered, a classifier can, in principle, flag a threat that no blacklist yet knows about. This proactive quality is the central motivation for the study. It is worth stressing that machine learning is not proposed here as a replacement for blacklists but as a first line of defence that operates in the interval blacklists cannot cover; in a deployed system the two would be layered, with reputation lookups handling known threats cheaply and a classifier assessing the residual unknown traffic.

The wider survey literature reinforces this framing. Reviews of the field distinguish list-based, heuristic and learning-based countermeasures, and repeatedly identify the zero-hour or zero-day window, the period immediately after a site goes live, as the point of greatest exposure and the natural target for predictive methods (Aleroud and Zhou, 2017; Chiew, Yong and Tan, 2018). Situating the present work against that taxonomy, it belongs squarely in the learning-based, feature-driven tradition, and its purpose is to quantify, under controlled conditions, how far that tradition can be pushed with standard classifiers before more elaborate machinery becomes necessary.

Aim, Objectives and Research Questions

The aim of this research is to evaluate and compare supervised machine learning approaches for the early detection of phishing websites, and to identify which feature groups and algorithms deliver the most reliable detection under a consistent experimental protocol.

The objectives are as follows:

1. To review feature-based machine learning methods for phishing detection and consolidate the categories of features most frequently reported as discriminative. 2. To prepare and describe a publicly available, labelled phishing dataset suitable for supervised classification. 3. To implement and train four classifier families, logistic regression, random forest, support vector machine and a feedforward neural network, under identical preprocessing and validation conditions. 4. To evaluate the models using precision, recall, F1 and ROC-AUC, and to analyse the relative contribution of different feature groups. 5. To discuss the practical implications, limitations and deployment considerations of the findings.

The research questions are:

• RQ1: How do the four candidate algorithms compare in detecting phishing websites when evaluated under identical preprocessing and validation conditions?
• RQ2: Which categories of features (URL-based, HTML/content-based, and domain/host-based) contribute most to detection performance?
• RQ3: What trade-offs between precision and recall are observed, and what do these imply for a system intended to warn users without generating excessive false alarms?

Literature Context

Feature-based machine learning for phishing detection has been an active field for over a decade, and a consistent theme is that carefully engineered features drawn from the URL and page structure allow standard classifiers to perform well. An early and influential content-based approach was CANTINA, which treated a page’s most informative terms as a signature and cross-checked them against a search engine to judge legitimacy (Zhang, Hong and Cranor, 2007). CANTINA established the principle that a page’s own content, not merely its reputation, carries signal, but it also exposed the cost of retrieving and parsing full pages in real time, motivating the later shift towards lighter-weight features.

Mohammad, Thabtah and McCluskey (2014) proposed an influential set of website features, spanning address-bar properties, abnormal HTML elements and domain characteristics, that has been widely reused as a benchmark feature schema. Their work established that many phishing indicators can be encoded as simple rule-derived attributes, making the problem tractable for conventional classifiers, and the resulting attribute set became the basis for several public datasets. A parallel line of work by Whittaker, Ryner and Nazif (2010) showed that classification could be scaled to production volumes, maintaining a large blacklist through an automated classifier rather than manual review, which demonstrated the operational viability of learned detectors at scale.

Subsequent studies extended this foundation in two directions. The first pursued ever-cheaper features. Sahingoz et al. (2019) focused on URL-only detection, arguing that lexical features of the address alone can achieve strong performance without needing to retrieve page content, an appealing property for real-time filtering where fetching a page may be undesirable. Verma and Das (2017) made a similar case, showing that fast lexical extraction from the URL string supports high-throughput screening, and Marchal et al. (2014) demonstrated with PhishStorm that features derived purely from the URL and lightweight external signals can be computed in a streaming setting suitable for live traffic. Jain and Gupta (2018) complemented this by developing a client-side detector that runs in the browser, emphasising that latency and privacy constraints at the point of use shape which features are practical. The second direction concerned the choice of learning algorithm. Comparative studies have repeatedly reported that ensemble methods, particularly random forests, tend to outperform single linear models on this task, a finding attributed to their capacity to model non-linear interactions among heterogeneous features (Abdelhamid, Ayesh and Thabtah, 2014). This is consistent with the theoretical properties of the algorithm, which reduces variance by averaging many decorrelated trees (Breiman, 2001), and with the general observation that margin-based and ensemble methods behave differently on tabular data with mixed feature types (Hastie, Tibshirani and Friedman, 2009).

More recently, the field has moved towards deep learning models that operate directly on raw URL character sequences, reducing the need for manual feature engineering (Le et al., 2018). While these approaches are promising, they typically require larger datasets and greater computational resources, and their gains over well-tuned classical models on tabular feature sets are not always decisive; recent surveys note that reported comparisons are frequently confounded by differences in data, preprocessing and evaluation, so headline accuracy figures are not directly comparable across papers (Zieni, Massari and Calzarossa, 2023). This proposal therefore positions a feedforward neural network alongside three classical models to test whether added model complexity is justified on a standard tabular feature set, rather than assuming it.

The gap this study addresses is not the invention of new features, but a controlled, reproducible comparison in which every model sees the same data, the same preprocessing and the same validation regime, so that performance differences can be attributed to the algorithms rather than to inconsistent experimental conditions. The survey literature explicitly identifies this lack of standardisation as a barrier to cumulative progress in the field (Chiew, Yong and Tan, 2018; Zieni, Massari and Calzarossa, 2023), which is the methodological need the present work is designed to meet.

Proposed Methodology

The research follows a quantitative, experimental design in which the same labelled data is passed through a fixed pipeline for every candidate model. The overall workflow, from raw data through to evaluation, is summarised in Figure 1; each stage is described in the subsections that follow.

Dataset

The study will use a well-known, publicly available phishing dataset, such as those distributed through the UCI Machine Learning Repository or the PhishTank and Mendeley Data collections commonly cited in the literature (Dua and Graff, 2019). These datasets typically provide several thousand labelled instances, each represented as a vector of pre-extracted features with a binary label indicating phishing or legitimate. Using an established public dataset supports reproducibility and allows the results to be compared against prior published work. The dataset will be described fully in the dissertation in terms of size, class balance and feature definitions; no attempt will be made to characterise it precisely here, as the exact instance count and feature list will be confirmed at the point of selection. Where more than one candidate dataset is suitable, the primary dataset will be chosen for its documentation quality and citation history, and a second dataset may be reserved for an external-validity check, testing whether a model tuned on one source degrades when applied to another, since dataset-specific overfitting is a known weakness of published phishing detectors.

Feature Engineering

Features will be grouped into three categories to support RQ2, and the grouping is itself part of the experimental design because it lets the analysis attribute performance to a family of signals rather than to individual, dataset-specific columns:

• URL-based (lexical) features: address length, number of dots and special characters, use of an IP address instead of a domain, presence of an “@” symbol, use of URL shortening, and the presence of a hyphen in the domain. These are the cheapest to compute and can be extracted from the address string alone.
• HTML and content-based features: proportion of external links, presence of forms submitting to foreign domains, use of pop-ups or iframes, and mismatches between anchor text and link destination. These require the page body and therefore a retrieval step, trading latency and safety for richer signal.
• Domain and host-based features: domain registration length, presence and validity of an SSL certificate, and DNS record availability. These draw on external lookups and reflect the observation that phishing infrastructure is often newly created and thinly provisioned.

Where the chosen dataset already provides these as engineered attributes, they will be used directly; where raw URLs are provided, a documented extraction pipeline will compute them. Categorical features will be encoded appropriately, using one-hot or ordinal encoding as suited to the feature, and numeric features will be standardised where required by the model, for example for the support vector machine, whose distance-based decision boundary is sensitive to feature scale. To answer RQ2 rigorously, models will be trained not only on the full feature set but also on each feature group in isolation and on cumulative combinations, so that the marginal contribution of each group can be measured; this ablation is more informative than reading a single importance ranking, because it captures redundancy between groups.

Candidate Algorithms

Four supervised classifiers will be trained and compared. They are chosen to span a deliberate range of model complexity and inductive bias, from a transparent linear model to a learned non-linear representation, so that the study can test whether complexity pays for itself on tabular phishing features. Table 1 summarises their characteristics.

The four models are, specifically:

1. Logistic regression, as an interpretable linear baseline against which more complex models are judged; its coefficients also offer a transparent read on the direction of each feature’s influence. 2. Random forest, an ensemble of decision trees expected to capture non-linear feature interactions and provide feature-importance estimates, and known to be robust to unscaled, mixed-type inputs (Breiman, 2001). 3. Support vector machine (SVM) with a radial basis function kernel, to test margin-based classification on the feature space; the RBF kernel allows a non-linear boundary while retaining the margin-maximising principle of the linear SVM (Cortes and Vapnik, 1995). 4. A feedforward neural network, a compact multilayer perceptron with one or two hidden layers, to assess whether a learned non-linear representation improves on the classical models without the data volume that sequence models demand.

All models will be implemented in Python using scikit-learn for the classical algorithms and a standard deep learning library for the neural network (Pedregosa et al., 2011). Hyperparameters, for example the regularisation strength of the logistic model, the number and depth of trees in the forest, the kernel width and penalty of the SVM, and the layer sizes and learning rate of the network, will be tuned via grid or randomised search on the training folds only, never on the held-out test data, to avoid leakage. To keep the comparison fair, every model will receive the same feature representation and the same class-imbalance treatment, so that any difference in measured performance reflects the algorithm rather than an incidental advantage in preprocessing.

Evaluation Metrics

Because phishing detection is a class-imbalanced problem in which the cost of a missed attack differs from the cost of a false alarm, accuracy alone is inadequate: a classifier that simply predicted the majority class could report a deceptively high accuracy while catching no phishing at all. The primary metrics will therefore be:

• Precision — the proportion of flagged sites that are genuinely phishing, reflecting the false-alarm burden on users.
• Recall — the proportion of phishing sites correctly caught, reflecting protective coverage.
• F1 score — the harmonic mean of precision and recall, giving a single balanced figure.
• ROC-AUC — the area under the receiver operating characteristic curve, summarising the trade-off across all decision thresholds.

Confusion matrices will be reported to make the error structure explicit, and precision-recall behaviour will be examined across thresholds rather than at a single default cut-off, since the operating point of a user-facing warning system is a design decision that RQ3 is intended to inform. Where models are compared, differences will be reported with their variability across folds so that a small numerical lead is not over-interpreted.

Validation

To obtain reliable estimates and avoid over-optimistic results, stratified k-fold cross-validation (k = 10) will be used, preserving class proportions in each fold (Hastie, Tibshirani and Friedman, 2009). A separate held-out test set will provide a final, untouched evaluation. All preprocessing steps, including scaling, encoding and any imbalance handling, will be fitted within cross-validation folds rather than on the whole dataset, to prevent information from the validation data leaking into training, a subtle but common source of inflated results in the phishing literature. Results will be reported as means with standard deviations across folds so that differences between models can be assessed for stability rather than judged on a single run. It is important to stress that no target performance figures are asserted in advance; the metrics above define how success will be measured, not what results are expected.

Ethical and Legal Considerations

The proposed research carries a low ethical risk profile because it uses aggregated, anonymised, publicly released datasets rather than data collected from human participants. No personal data will be gathered, and no individual users will be observed, tracked or profiled. Nonetheless, several considerations warrant explicit attention.

First, the study will comply with the data licences of any dataset used, citing sources correctly and respecting redistribution terms. Second, phishing URLs are, by definition, links to potentially malicious content; live URLs will not be visited during the research, and only the pre-extracted features or safely handled offline data will be processed, avoiding any accidental interaction with active malicious infrastructure. Where a retrieval step is genuinely required to compute content-based features, it will be performed in an isolated, sandboxed environment with no user credentials present, and any such procedure will be documented and approved in advance. Third, the dual-use nature of detection research is acknowledged: understanding which features distinguish phishing from legitimate sites could in principle inform evasion. This risk is mitigated by relying on features already documented extensively in the public literature, so the work discloses nothing that is not already widely known. Fourth, the fairness and error costs of an automated classifier deserve consideration even in a purely offline study: a detector deployed to warn users could, through false positives, wrongly flag legitimate small-business or non-English sites whose structural features happen to resemble the phishing patterns learned from the training data, and the analysis will therefore treat false-positive behaviour as a first-class concern rather than an afterthought. The research will be conducted in accordance with the host institution’s research ethics policy and its guidance on responsible security research, and formal ethics sign-off will be sought before work begins where required.

Limitations and Scope

Several limitations are acknowledged at the outset so that the findings can be interpreted honestly. The study is bounded to four classifier families and to feature-based, tabular detection; it does not evaluate raw-sequence deep models, visual or screenshot-based detection, or the natural-language content of pages, all of which are legitimate but separate research directions. Findings will be conditioned on the chosen public dataset, and because such datasets are collected at a point in time, they may not reflect the current tactics of active campaigns; a model that performs well on historical data could degrade against newer evasion techniques, a phenomenon known as concept drift. The external-validity check across a second dataset partially addresses this but cannot eliminate it. The study measures offline classification quality, not real-world deployment performance, so factors such as inference latency, integration with browsers or gateways, and adversarial robustness to feature manipulation are discussed rather than experimentally evaluated. Finally, the comparison is intentionally narrow in order to be rigorous: by holding data, preprocessing and validation constant, the study trades breadth of models for confidence in the conclusions it does draw. These limitations define the boundary of the claims the dissertation will make and mark out the most natural extensions for future work.

Timeline

The project is scoped for a standard Master’s dissertation period of approximately six months, organised into six sequential but partly overlapping phases. Figure 2 presents the schedule as an illustrative Gantt-style chart; the accompanying table lists the same phases with their activities.

Milestones include a completed literature review at week 4, a working evaluation harness at week 11, and a full results set at week 20.

Expected Contribution

The principal contribution of this research is a controlled, reproducible comparison of four widely used classifier families for phishing website detection, conducted under a single consistent experimental protocol. Much of the existing literature reports results from studies that differ in dataset, preprocessing and validation method, which makes head-to-head comparison difficult (Zieni, Massari and Calzarossa, 2023). By holding these factors constant, the study aims to clarify which algorithms genuinely offer an advantage and by how much. A secondary contribution is an analysis of which feature groups, lexical, content-based or host-based, contribute most to timely detection, which has practical value for designing lightweight, real-time filters, since a group that delivers most of the signal at the lowest extraction cost is the natural foundation for a deployable system. Finally, by releasing a documented methodology and code, the project is intended to support reproducibility and to provide a usable baseline for future work, including later extensions towards deep sequence models, cross-dataset robustness testing, or real-time browser integration.

References

Abdelhamid, N., Ayesh, A. and Thabtah, F. (2014) ‘Phishing detection based associative classification data mining’, Expert Systems with Applications, 41(13), pp. 5948–5959.

Aleroud, A. and Zhou, L. (2017) ‘Phishing environments, techniques, and countermeasures: A survey’, Computers & Security, 68, pp. 160–196.

Anti-Phishing Working Group (APWG) (2023) Phishing Activity Trends Report. Available at: https://apwg.org/trendsreports/ (Accessed: 20 July 2026).

Breiman, L. (2001) ‘Random forests’, Machine Learning, 45(1), pp. 5–32.

Chiew, K.L., Yong, K.S.C. and Tan, C.L. (2018) ‘A survey of phishing attacks: Their types, vectors and technical approaches’, Expert Systems with Applications, 106, pp. 1–20.

Cortes, C. and Vapnik, V. (1995) ‘Support-vector networks’, Machine Learning, 20(3), pp. 273–297.

Dua, D. and Graff, C. (2019) UCI Machine Learning Repository. Irvine, CA: University of California, School of Information and Computer Science. Available at: https://archive.ics.uci.edu/ml (Accessed: 20 July 2026).

Hastie, T., Tibshirani, R. and Friedman, J. (2009) The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2nd edn. New York: Springer.

Jain, A.K. and Gupta, B.B. (2018) ‘Towards detection of phishing websites on client-side using machine learning based approach’, Telecommunication Systems, 68(4), pp. 687–700.

Le, H., Pham, Q., Sahoo, D. and Hoi, S.C.H. (2018) ‘URLNet: Learning a URL representation with deep learning for malicious URL detection’, arXiv preprint arXiv:1802.03162.

Marchal, S., François, J., State, R. and Engel, T. (2014) ‘PhishStorm: Detecting phishing with streaming analytics’, IEEE Transactions on Network and Service Management, 11(4), pp. 458–471.

Mohammad, R.M., Thabtah, F. and McCluskey, L. (2014) ‘Predicting phishing websites based on self-structuring neural network’, Neural Computing and Applications, 25(2), pp. 443–458.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M. and Duchesnay, É. (2011) ‘Scikit-learn: Machine learning in Python’, Journal of Machine Learning Research, 12, pp. 2825–2830.

Sahingoz, O.K., Buber, E., Demir, O. and Diri, B. (2019) ‘Machine learning based phishing detection from URLs’, Expert Systems with Applications, 117, pp. 345–357.

Sheng, S., Wardman, B., Warner, G., Cranor, L.F., Hong, J. and Zhang, C. (2009) ‘An empirical analysis of phishing blacklists’, in Proceedings of the 6th Conference on Email and Anti-Spam (CEAS). Mountain View, CA.

Verma, R. and Das, A. (2017) ‘What’s in a URL: Fast feature extraction and malicious URL detection’, in Proceedings of the 3rd ACM International Workshop on Security and Privacy Analytics (IWSPA). New York: ACM, pp. 55–63.

Whittaker, C., Ryner, B. and Nazif, M. (2010) ‘Large-scale automatic classification of phishing pages’, in Proceedings of the Network and Distributed System Security Symposium (NDSS). San Diego, CA.

Zhang, Y., Hong, J.I. and Cranor, L.F. (2007) ‘Cantina: A content-based approach to detecting phishing web sites’, in Proceedings of the 16th International Conference on World Wide Web (WWW). New York: ACM, pp. 639–648.

Zieni, R., Massari, L. and Calzarossa, M.C. (2023) ‘Phishing or not phishing? A survey on the detection of phishing websites’, IEEE Access, 11, pp. 18499–18519.