Classification and phishing websites detection by fuzzy rules and modified inclined planes optimization

چکیده مقاله

This paper presents a novel intelligent approach for detecting phishing websites by combining a fuzzy rule-based feature selection mechanism with an enhanced meta-heuristic optimization algorithm. The primary goal is to simultaneously address three critical challenges in phishing detection: achieving flexible and dynamic selection of impactful features, adapting the classification algorithm's behavior to different website characteristics, and efficiently processing large volumes of websites. The proposed method first introduces an adaptive threshold mechanism, where fuzzy rules are used to dynamically select a relevant subset of features based on their phishing threat level. This allows the system to focus on the most discriminative attributes for each set of websites, reducing computational complexity while maintaining detection accuracy.

The core of the classification system is an improved version of the Inclined Planes Optimization (IPO) algorithm, referred to as MIPO (Modified Inclined Planes Optimization). Key enhancements include endowing the algorithm with memory—by retaining past positions and heights of search agents—and incorporating a soft decay function to control the influence of historical data. Furthermore, to intelligently adjust the exploration and exploitation parameters of the algorithm, a Mamdani-type fuzzy inference system with 12 rules is designed. This system dynamically tunes the acceleration and velocity coefficients based on the algorithm’s convergence state, leading to more efficient and balanced global-local search behavior during optimization.

The methodology was implemented and tested on a standard dataset sourced from PhishTank, comprising 1000 websites categorized as legitimate, suspicious, or phishing. The system was evaluated using three different feature-set sizes (15, 8, and 3 features) selected through the proposed fuzzy rule mechanism. Experimental results demonstrate that the MIPO-based classifier consistently outperforms the original IPO across all feature configurations, achieving lower classification error rates and faster convergence. For instance, with 15 features, MIPO reduced the error rate to 2.7% compared to IPO's 3.2%, while also significantly cutting down detection time.

A notable achievement of the study is its competitive performance compared to other state-of-the-art meta-heuristic methods such as Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), Bacterial Foraging Optimization Algorithm (BFOA), and Modified Bat Algorithm (MBAT). Despite using far fewer features (15, 8, or 3 versus 27 in some benchmarks), the MIPO approach achieved comparable or superior accuracy with substantially lower computational time. This highlights the effectiveness of the adaptive feature selection mechanism in reducing problem dimensionality without sacrificing detection quality. Moreover, the integration of fuzzy logic for parameter control and feature selection introduces a high degree of adaptability, making the system robust against evolving phishing tactics and suitable for large-scale, real-time website analysis.

In conclusion, this research successfully develops a flexible, accurate, and efficient hybrid framework for phishing website classification. By intelligently selecting features and dynamically optimizing the classifier’s parameters, the proposed system addresses key limitations of existing methods. The positive results confirm its potential as a reliable tool for enhancing cybersecurity, capable of handling diverse website populations with high precision and speed, thereby offering a practical solution for real-world phishing detection challenges.