In-Network Defense: Safeguarding the Network Against Evolving DDoS Attacks
Muhammad Saqib, Halima Elbiaze, Roch Glitho · 2024
Emerging technologies that encompass a multitude of tiny wearable devices are vulnerable to cyberattacks that can turn them into bots for launching Distributed Denial of Service (DDoS) attacks. In-network Machine Learning (ML) has emerged as a prominent solution for detecting and responding to such attacks in the shortest possible time to avoid disrupting user experience. However, the dynamic nature of attack traffic patterns necessitates continuous adaptation of conventional one-size-fits-all ML models. The manual process of identifying novel malicious traffic patterns and updating the ML model from the control plane to the network data plane is time-consuming and labor-intensive. This study aims to automate the identification of unseen malicious traffic patterns and update the ML model in programmable networks using a data-driven approach. Specifically, we determine drift detection thresholds from the baseline performance of historical (i.e., training) data and consider any deviation as anomalies in unseen (i.e., testing) data. These thresholds are continuously updated by considering changes in the data distribution and in-network inference results. We utilize an intrusion detection dataset (CIC-IDS2017) to illustrate the impact of emerging attacks on model performance degradation and the efficacy of our proposed data-driven method in mitigating these attacks. Our approach has proven effective in safeguarding against evolving DDoS attacks.