Authors: Rohan C. Shrestha
Abstract: Phishing attacks have become a critical cybersecurity threat in the digital era, targeting individuals, businesses, and organizations to obtain sensitive information such as login credentials, financial data, and personal identification details. The sophistication of modern phishing attacks has evolved beyond simple spam emails, encompassing spear phishing, clone phishing, smishing, and vishing, making detection increasingly difficult. Traditional detection mechanisms, including rule-based systems, blacklists, and heuristic approaches, often fail to detect new or obfuscated attacks and are prone to high false-positive rates, which can compromise security operations. Deep learning (DL), a subset of artificial intelligence, offers promising solutions to these challenges through its ability to automatically extract complex features, learn non-linear relationships, and detect patterns that are imperceptible to human analysts or conventional machine learning algorithms. This review examines the application of deep learning in enhancing phishing detection mechanisms, focusing on architectures such as Artificial Neural Networks (ANNs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, and autoencoders. The discussion highlights how these models improve detection accuracy, adaptability, and resilience against evolving phishing strategies. Furthermore, the review explores the utilization of diverse datasets, challenges in computational requirements and adversarial robustness, and the role of hybrid and ensemble models in optimizing performance. Finally, future directions, including explainable AI, multi-modal detection systems, and adaptive reinforcement learning frameworks, are addressed. Overall, deep learning provides a transformative approach to phishing detection, offering enhanced efficiency, robustness, and proactive threat mitigation, while opening avenues for continued research into intelligent, adaptive cybersecurity solutions.
DOI: http://doi.org/10.5281/zenodo.17879281
