In the rapidly evolving landscape of cybersecurity, researchers Tharindu Lakshan Yasarathna and Nhien-An Le-Khac from the School of Computer Science at University College Dublin are shedding light on critical vulnerabilities in deep learning (DL) based autonomous anomaly detection (AAD) systems used in Software-Defined Networking-Internet of Things (SDN-IoT) networks. Their recent Systematic Literature Review (SoK) study offers a comprehensive analysis of adversarial threats, providing a structured approach to understanding and mitigating these risks.
The integration of SDN and IoT technologies has revolutionized network control and flexibility, enabling real-time threat detection through DL-based AAD systems. However, these advanced systems are not impervious to attacks. Yasarathna and Le-Khac highlight that adversarial attacks can manipulate input data or exploit model weaknesses, significantly degrading detection accuracy. Their study introduces a structured adversarial threat model and a detailed taxonomy of attacks, categorizing them into data, model, and hybrid-level threats. This systematic approach sets their work apart from previous research, offering a more nuanced understanding of the vulnerabilities specific to DL-based AAD systems in SDN-IoT environments.
The researchers evaluated various attack strategies, including white, black, and grey-box attacks, across popular benchmark datasets. Their findings are alarming: adversarial attacks can reduce detection accuracy by up to 48.4%. Among the attacks studied, Membership Inference caused the most significant drop in accuracy. Techniques like Carlini & Wagner (C&W) and DeepFool achieved high evasion success rates, underscoring the severity of the threat. Despite these challenges, the study also reveals that adversarial training can enhance the robustness of DL models. However, the high computational overhead of this method limits its real-time deployment in SDN-IoT applications.
To address these vulnerabilities, Yasarathna and Le-Khac propose adaptive countermeasures, including real-time adversarial mitigation, enhanced retraining mechanisms, and explainable AI-driven security frameworks. By integrating structured threat models, their study offers a more comprehensive approach to attack categorization, impact assessment, and defence evaluation. The practical applications of their research are significant for the defence and security sector. As SDN-IoT networks become increasingly prevalent, understanding and mitigating adversarial threats is crucial for maintaining network integrity and security.
This foundational study serves as a vital reference for researchers and practitioners aiming to enhance the security of DL-based AAD systems in SDN-IoT networks. By providing a systematic adversarial threat model and conceptual defence evaluation based on prior empirical studies, Yasarathna and Le-Khac offer practical recommendations for improving resilience, interpretability, and computational efficiency. Their work underscores the importance of proactive measures in safeguarding advanced network systems against evolving cyber threats.
This article is based on research available at arXiv.