Abstract
As phishing assaults grow more sophisticated, conventional machine learning models frequently fail to detect phishing attacks because of their inability to comprehend intricate, non-linear feature relationships and their susceptibility to noisy data. This study introduces an unconventional Two-Stage Multi-Layer Architecture aimed at improving the identification of phishing URLs by integrating computational efficiency with profound analytical depth. The study addresses two primary limitations identified in the existing Hybrid LSD model: Inadequate feature selection and linear classification constraints. Recursive Feature Elimination (RFE) was used to remove unnecessary "noise" features in the first stage, and the Hybrid LSD Fast Filter then handles cases that are obvious. An essential advancement of this architecture is the Confidence-Based Triage Gate, which uses piecewise logic (thresholds of 0.1 and 0.9) to identify "ambiguous" samples instances when conventional models exhibit uncertainty. The experimental result shows that the proposed system outperforms the existing baselines. The classification accuracy increased from a baseline of 97.50% to 99.00% with the integration of the Stage 2 deep analysis layer. Additionally, the model's True Positive Rate (TPR) of 0.99 and False Positive Rate (FPR) of 0.02 demonstrate its remarkable capacity to identify malicious URLs with the least amount of interference with legitimate user traffic. By bridging the gap between the speed of conventional machine learning and the analytical depth of deep learning, this research offers a scalable, high-precision framework that offers a strong defense against contemporary cyber threats.

Key words: Phishing Detection, Machine Learning, Deep Learning, Cybersecurity, Multi-Layer Perceptron