Comprehensive Evaluation of XBNet for Multi-Class IoT Attack Detection

IoT environments face growing security threats due to their heterogeneity, resource limits, and scale. This paper evaluates an IoT-optimized Xtremely Boosted Network (XBNet) for multi-class attack detection, incorporating protocol-aware normalization, advanced neural architectures, and ensemble strategies. Using the UNB CIC IoT 2023 dataset (33 attacks, 105 devices), the authors conducted hyperparameter, complexity, and error analyses. XBNet achieved 99.5% binary, 94.5% 8-class, and 96.7% 34-class accuracy—outperforming traditional methods with efficient computation. SHAP analysis highlighted protocol-specific features: flow duration (DoS) and packet variance (DDoS). Error analysis showed 68% of DoS/DDoS misclassifications were due to temporal pattern issues. Runtime tests showed feasible deployment from edge to servers, with 42% memory savings via quantization at 98.8% accuracy. The results offer practical insights for real-world IoT security and guide future intrusion detection advances.