An Intelligent Deep Learning Framework For Early Detection Of Distributed Denial-Of-Service Attacks

Distributed Denial-of-Service (DDoS) attacks remain a critical threat to modern network infrastructures, particularly in cloud and software-defined networking environments. This study proposes a temporal-aware deep learning framework for early detection of DDoS attacks using flow-based traffic features from the CIC-DDoS2019 dataset. A subset of 500,000 traffic flows was utilized, including BENIGN, DrDoS_NetBIOS, and DrDoS_SSDP classes. To ensure realistic evaluation, a strict chronological data-splitting strategy was applied to prevent temporal leakage, and class-weighted optimization was incorporated to address severe imbalance. A feed-forward deep neural network was implemented and evaluated against Logistic Regression and Random Forest baselines. Experimental results demonstrate strong detection performance in the binary configuration, with stable macro F1-scores and high attack recall under imbalance-aware training. Progressive timestamp-based evaluation further confirms that malicious traffic can be reliably detected during early stages of observation, with minimal performance degradation across increasing portions of the traffic timeline. A feature ablation study indicates that static flow-level statistical attributes provide strong discriminative signals. The findings highlight the importance of temporal-aware validation and imbalance-sensitive optimization for realistic and deployment-oriented DDoS detection systems.