View Proposal

---

Proposer

Zi Hau Chin

Title

Explainable Detection of Phishing Websites Using URL and Page-Level Features

Goal

Classify websites as phishing or legitimate and explain the URL and webpage characteristics contributing to the decision.

Description

This project involves developing an explainable machine-learning system that detects potentially fraudulent websites using both URL-level and webpage-level features. URL features may include URL length, unusual characters, number of subdomains, use of an IP address, encoded text, suspicious top-level domains, and similarity to known domain names. Page-level features may include login forms, external form destinations, hidden elements, embedded frames, suspicious scripts, mismatches between displayed and actual links, and inconsistencies between the page title and domain. The student will prepare a phishing and legitimate website dataset, extract suitable features, and compare several classification models, such as logistic regression, decision trees, random forests, or gradient-boosted trees. The system should generate a phishing probability and explain the factors that influenced each prediction using techniques such as decision paths, feature importance, or SHAP. For example, it may warn that a page uses an unusually long URL, contains several misleading subdomains, and sends login information to an external domain. The final prototype may be implemented as a web-based URL checker or browser extension. It should display a clear risk level, confidence score, and understandable explanation rather than only showing a technical classification. The project should also investigate whether explanations help users distinguish phishing websites from legitimate ones and whether the detector remains reliable when attackers modify URLs or webpage elements to avoid detection. Evaluation should include precision, recall, F1 score, false-positive and false-negative rates, PR-AUC, processing time, explanation fidelity, robustness to evasion attempts, and user understanding where ethical approval permits a user study. -Does combining URL and page-level features outperform URL-only detection? -Do explanations improve users’ phishing decisions? -How well does the model generalise to newer domains? -How robust is it to common phishing evasion techniques?

Resources

PhiUSIIL UCI Dataset, PhiUSIIL on Mendeley Data, PhiUSIIL Data-Mining Project, Practical PhiUSIIL Detection Experiments

Background

Python, machine learning, web technologies, basic cybersecurity and phishing concepts, explainable AI, feature extraction and classification

Url

Difficulty Level

Moderate

Ethical Approval

Full

Number Of Students

1

Supervisor

Zi Hau Chin

Keywords

phishing detection, malicious url, explainable ai, url lexical features, html features, domain impersonation, browser security, adversarial evasion

Degrees

Bachelor of Science in Computing Science