Zwha Abdulhmid Mohamed Albeerish

Saedah Mohammed Omar Albeerish

Abstract:

This study, titled “Intelligent Detection of Cybersecurity Threats Using Deep Learning Algorithms”, discusses the mechanisms to counter the escalating cyber threats resulting from rapid digital transformation, the expansion of Internet of Things (IoT) networks, and smart cities. The study shows the inadequacy of traditional signature-based systems in detecting novel or zero-day attacks, which necessitated the development of intelligent anomaly detection systems.  The study reviews three main challenges facing current systems: First, the severe imbalance of data (class/imbalance) where rare attacks account for a very small percentage (up to 0.001%), which causes model bias; second, the inability of single models to grasp the complex spatial and temporal characteristics of network flows; third, the lack of mathematical transparency and classification of deep learning systems as “black boxes.”  To solve these problems, the study proposes an application framework based on the CICIDS2017  reference database. The framework is based on a hybrid neural model consisting of: automatic encoders (AE) for noise isolation and static feature extraction, 1D-CNN for spatial derivation, and bi-directional repetitive neural networks (Bi-LSTM) for tracking the temporal pattern of threats. To address the data flaw, the research adopted a three-pronged methodology that includes the Adaptive, SMOTE algorithm for intelligent sampling of rare categories, and Tomek, Links filters, and the  Focal Loss Function (Focal\,Loss).  The statistical results showed that the proposed model was superior to achieving an overall accuracy of 99.85% and a very low false alarm rate (FAR = 0.03%). The study also revealed a semantic gap in the data payload (Semantic, Payload, Gap) that prevents the detection of application layer attacks (such as XSS andSQLi) based solely on statistical characteristics, necessitating the integration of natural language processing (NLP)  algorithms.Finally, the study employed Shapley’s Added Values (SHAP) within Interpretable Artificial Intelligence (XAI)  tools to provide clear visual justifications for cybersecurity analysts.