Explainability-Aware Evaluation of Transfer Learning Models for IoT DDoS Detection Under Resource Constraints

Researchers evaluated seven pre-trained CNN architectures for IoT DDoS attack detection, finding that DenseNet and MobileNet models provide the best balance of accuracy, reliability, and interpretability under resource constraints. The study emphasizes the importance of combining performance metrics with explainability when deploying AI security models in IoT environments.

• →DenseNet169 offers the strongest reliability and interpretability alignment for IoT DDoS detection.
• →MobileNetV3 provides an effective latency-accuracy trade-off suitable for fog-level deployment.
• →Transfer learning models show promising detection accuracy but require comprehensive evaluation beyond performance metrics.
• →The research integrates explainability tools like Grad-CAM and SHAP to assess model interpretability in security applications.
• →Resource-constrained IoT environments need AI models that balance computational efficiency with detection reliability.