Abstract

As Internet of Things (IoT) devices increasingly integrate into contemporary existence, they have simultaneously emerged as attractive targets for extensive cyber assaults, particularly the Mirai botnet. Existing methods for identifying this virus typically require excessive computational resources for IoT devices with limited capabilities, while the quicker “lightweight” options frequently function as “black boxes”, obscuring the decision-making process from users. This research presents a Lightweight Artificial Neural Network (LANN) integrated with Explainable Artificial Intelligence (XAI) using the Deep SHapley Additive exPlanations (DeepSHAP) framework to develop a detection system that is efficient and comprehensible. We trained the model to discover unusual attack patterns in the IoT-23 dataset that normal AI misses using a cost-sensitive learning technique. Our model had 99.22% accuracy and a 0.96 F1-score for minority attacks. Our DeepSHAP analysis explains its decisions and identifies cat_proto_udp as the key indicator of an attack, which matches Mirai behavioral patterns. This framework demonstrates that transparency and efficiency can be simultaneously accomplished, which makes it a viable option for protecting Internet of Things networks in the real world. It is a more practical and trustworthy solution for real-world Internet of Things (IoT) security than existing approaches, which either lack lightweight design, explainability, or both. The proposed framework, on the other hand, achieves all three objectives in a unique way, achieving a compact model size of 0.04 MB and an inference time of 0.096 s.