AMPDF: A Hybrid Deep Learning Framework for Multi-Modal Phishing Detection in Cybersecurity

Abstract

Phishing is a type of cyberattack where attackers deceive users into revealing sensitive information via fake emails or fake websites. With the rise of online banking and social networking, they have been taking advantage of user vulnerabilities instead of network flaws, and phishing attacks have become more advanced. Such attacks generally involve emails containing harmful links that lead victims to fake sites that can swipe personal data. Classic anti-phishing tools heavily depend on blocklists and allowlists which make them ineffective against novel attacks leading towards high false positive rates. To overcome this, we introduce the Adaptive Multi-Modal Phishing Detection Framework (AMPDF). This data-driven hybrid model identifies phishing via analyzing three datasets: URL data, page content as well as traffic data using AI techniques. For these features, AMPDM Uses a CNN to extract the URL pattern then feeds it into a Dense layer with a LeakyReLU activation to analyze the page and finally another Dense layer for traffic behavior. These extractors integrate into a Fusion Layer followed by a Dense Layer with Dropout to avoid overfitting and a final classification layer splits phishing from legitimate instances. The model was evaluated using precision, recall, and accuracy metrics, with 98% accuracy and a test loss of 0.0708 on the test dataset. The experiments justify AMPDF as a significant and non-intrusive detection model against the traditional methods that are popular.