An Energy-Efficient Cooperative DDoS Attack Mitigation using Federated Deep Learning Framework in SDN

Abstract

A Distributed Denial of Service (DDoS) attack can be launched using the vast number of edge connected devices in Software Defined AIoT systems. The shortage of modern labeled training data makes centralized defenses ineffective, while privacy concerns restrict sharing sensitive traffic information. To address these challenges, this work proposes a decentralized defense framework for AIoT based on a trustworthy Federated Learning (FL) architecture. The framework integrates three key components. First, a secure multiparty computation based aggregation scheme enables privacy preserving exchange of local model updates while remaining robust to outliers. Second, a horizontal cooperation based mitigation mechanism reduces packet loss and limits attack impact on legitimate traffic. Third, a semi supervised FL based anomaly detection strategy leverages a small number of labeled outliers to accurately identify attack classes while reducing false alarms. Experiments conducted on real world datasets, including CICIDS2017 and InSDN, show that the proposed approach significantly outperforms state of the art SDN based protection systems in accuracy and efficiency. The study further examines FL enabled by Software Defined Networking as a paradigm that improves security and reduces training time without degrading learning performance. However, SDN assisted FL introduces new vulnerabilities, particularly DDoS attacks targeting the SDN control plane and disrupting model parameter exchanges. To mitigate this risk, a contrastive learning based detector using a two layer deep neural architecture is introduced. The LSTM SVM achieves 99.75 percent accuracy and 99.80 percent recall, while the CNN BiGRU reaches 99.36 percent accuracy and 99.55 percent recall in practical deployments scenarios.