SFA-IoT: segment-to-flow aligned adversarial defense for interpretable IoT device identification

Existing ML-based IoT device identification models achieve strong performance on clean data but degrade substantially when even small adversarial perturbations are introduced. Moreover, these models offer limited insight into the reasons behind such failures. To address these limitations, we propose SFA-IoT, an adversarial purification framework for IoT device identification. Specifically, SFA-IoT leverages an empirical property observed in IoT traffic, in which adversarial perturbations tend to concentrate in short protocol-event regions rather than being uniformly distributed across the entire flow. Based on this observation, SFA-IoT enforces segment-to-flow alignment using a frozen expert identifier and dual identity constraints at both the feature and embedding levels. We evaluate four representative classifiers, namely DNN, SVM, XGBoost, and Decision Tree, under white-box attacks, including FGSM and PGD, as well as black-box transfer attacks. Experiments on the UNSW IoT dataset across four temporal resolutions show that FGSM and PGD attacks reduce accuracy by up to 92%, while SFA-IoT restores 9-40% accuracy in a single forward pass, with a peak improvement of 58.6% on decision trees. Finally, SHAP and LIME diagnostics are used not only for explanation but also to guide the defense architecture, showing that interpretability and adversarial robustness can serve as complementary objectives.