MULTI-VIEW TEMPORAL CONTRASTIVE LEARNING NETWORK FOR ADAPTIVE REAL-TIME PHISHING DETECTION IN DYNAMIC AI-DRIVEN CYBERSECURITY ENVIRONMENTS

Abstract

The increasing dependence on digital communication, cloud services, and online financial transactions has intensified the frequency of phishing attacks, which is creating significant challenges for an intelligent cybersecurity system. This study is proposing a Multi-View Temporal Contrastive Learning Network (MTCLN) for the phishing detection in dynamic AI-driven cybersecurity environments. The proposed framework is combining the temporal feature learning module with the contrastive representation learning strategy in capturing the evolving phishing characteristics from the heterogeneous data sources. This includes URL structures, webpage content, DNS attributes, and behavioral metadata. A lightweight gated temporal encoder is extracting the sequential attack patterns, while a contrastive feature alignment module is improving the discrimination between the legitimate and phishing instances under the changing attack scenarios. Further, an adaptive confidence calibration mechanism is adjusting the classification threshold in accordance with the recent detection feedback. This is allowing robust operation without requiring the complete model retraining. The proposed method is presenting improved phishing detection capability by achieving an accuracy of 88.4%, precision of 87.8%, recall of 86.9%, and F1-score of 87.3%, which is performing better than the CNN-Based Phishing Detection Model, LSTM-Based Phishing Detection Model, and Attention-Based Transformer Phishing Detection Model. The framework is reducing the detection latency to 31 ms, when it is compared with the 37 ms, 40 ms, and 42 ms obtained by the Conventional methods.