PhishGuard Lite: A Hybrid Explainable Phishing Detection System Using Rule-Based Analysis and Machine Learning

Phishing attacks represent one of the most persistent and impactful threats in modern cybersecurity, exploiting human trust through deceptive communication, fraudulent links, and impersonation techniques. These attacks continue to evolve in sophistication, making detection a persistent challenge for both individuals and organizations [1]. This paper presents PhishGuard Lite, a lightweight and interpretable hybrid phishing detection system that integrates rule-based heuristic analysis with machine learning classification. The proposed system employs a structured pipeline consisting of text preprocessing, rule-based detection, risk scoring, and an explainability module. The detection engine leverages predefined heuristics, including suspicious keywords, urgency patterns, domain indicators, and URL analysis, to identify phishing characteristics. A weighted scoring mechanism computes a risk score mapped to categorical risk levels (Low, Medium, High) [2]. The inclusion of an explainability module enables the system to provide clear, human-readable justifications for each classification decision [3]. Experimental evaluation was conducted on a dataset comprising both publicly sourced and synthetically generated phishing and legitimate samples. Results indicate that PhishGuard Lite achieves competitive performance while maintaining full interpretability and low computational cost. The final implementation incorporates a lightweight Logistic Regression classifier alongside the rule-based engine, forming a hybrid phishing detection framework that balances explainability, computational efficiency, and predictive performance [4]. The findings suggest that hybrid systems combining rule-based reasoning with machine learning offer a viable solution for phishing detection, particularly in environments where explainability and resource constraints are critical [5]. Future work will explore adaptive rule generation and integration of transformer-based models to enhance detection performance.