Adversarial Attacks on Deep Learning Models in SDN-IoT Networks

Abstract

The integration of Software-Defined Networking (SDN) with large-scale Internet of Things (IoT) deployments enhances network programmability and visibility, yet it also introduces new security risks across emerging edge cloud computing environments. Deep Learning (DL) based autonomous anomaly detection (AAD) has emerged as a key defence mechanism for securing distributed and edge-enabled IoT infrastructures. Despite their effectiveness, the vulnerability of these DL models to adversarial attacks in SDN-IoT contexts remains underexplored, particularly in edge-cloud AI deployments, where models may operate across distributed infrastructures. This paper presents a comprehensive adversarial analysis of three representative DL architectures, CNN, LSTM, and AE-LSTM, evaluated across three diverse datasets: CICIDS2017, InSDN, and CICIoT2023. We investigate both data-level evasion attacks (FGSM, PGD) under white-box conditions and model-level threats via a noisebased Data Free Model Extraction (DFME) attack under realistic black-box conditions. Results reveal that minor perturbations can severely degrade detection performance, while DFME successfully reconstructs high-fidelity surrogate models without access to training data. These findings highlight significant integrity and confidentiality risks for AI-driven intrusion detection systems deployed across the edge-cloud continuum and emphasise the need for adversarially robust DL defences for distributed AAD systems in SDN-IoT networks.