Proactive Cyber Defense through Multitask Deep Learning: Real-Time Detection and Traffic Prediction in Enterprise Networks

Abstract

Modern network infrastructures face escalating challenges in detecting subtle anomalies and forecasting traffic dynamics due to the high volume, velocity, and complexity of contemporary network data. This research addresses a critical gap which is a single deep learning model simultaneously and effectively perform both anomaly detection and traffic metric prediction? To answer this, we propose a novel multitask deep LSTM-based framework that unifies classification (anomaly identification) and regression (future traffic forecasting) within a shared architecture. Leveraging the real-world CESNET-TimeSeries24 dataset, this research approach exploits temporal dependencies in network flow data to jointly learn robust representations for both tasks. The key advantage of this multitask design lies in its ability to enhance generalization by sharing learned features improving anomaly detection through contextual forecasting and refining predictions by recognizing abnormal patterns. This research   model achieves 95.8% classification accuracy, 76.8% anomaly recall, and an exceptional R² score of 0.98 in regression, demonstrating high operational reliability. This work contributes a practical, end-to-end solution that advances intelligent network monitoring by combining security and performance analytics in one system offering network operators a proactive tool for cyber defense, capacity planning, and real-time decision-making. By validating the synergy between prediction and detection, this research sets a foundation for more adaptive, efficient, and holistic AI-driven network management.