A Stratified Deep Detection Pipeline For Ddos Threats Leveraging Recursive Dimensionality Reduction, Probabilistic Latent Encoding, And Temporal Contextualization Networks

Abstract

Intelligent and adaptable detection systems are required due to the increasing complexity of Distributed Denial of Service (DDoS) attacks. This paper discuss  an innovative blended approach for detecting DDoS threats which uses an Attention-Enriched Transfer Learning (AETL)[20] framework with Variational Autoencoders (VAEs)[19]. In this model, we aimed to enhance anomaly detection. The model utilizes Deep Packet Inspection (DPI), sophisticated Recursive Feature Elimination (RFE), and temporal pattern recognition by Long-Short-Term Memory (LSTM) networks. To optimize the model's efficiency by differentiating between malicious and benign flows, VAEs are used, hence the unsupervised learning of latent representations of network traffic is done. These new experimental findings utilized the most recent datasets, such as CIC-DDoS2020, CIC-DDoS2019, CIC-DDoS2017 and TON_IoT, show enhanced  detection accuracy, F1-score, and response time. This architecture is shown to scale efficiently while adapting to emerging attack vectors.