A Comparative Study of Imbalanced and Balanced Data in IoT DDoS Detection
Soe Kalayar Naing, Pyke Tin · 2025
The Internet of Things (IoT) is currently leading edge in institute, businesses, and daily actions. Cybercriminals have comprised the integrity and security of user evidence by adding the employments of IoT devices. Distributed denial of service (DDoS) attacks in special pose a serious risk to user information and Internet security. The flow in IoT devices has increased vulnerability to Distributed Denial of Service (DDoS) attacks, which can overfill network appliances and cause vital harm. Even though detecting these attacks in real time is vital, an imbalance dataset for training machine learning models can lead to contort predictions and higher false error rates. This study compared the analysis of imbalanced and balanced datasets to DDoS attacks of IoT networks with the UNSW-NB15 dataset. It investigates the effects SMOTE and Borderline-SMOTE of data-balancing techniques by measuring accuracy, precision, recall, and computing efficiency on the Light Gradient Boosting Model (LightGBM). The results revealed that models trained using Borderline-SMOTE of balancing technique and LightGBM on UNSW-NB15 datasets had an accuracy of 95.41%, 0.0015 as MAE of lower error rates and 0.0016 as RMSE of error rate, and an effective training time of 4.3180 seconds. These results demonstrate the important role of balancing techniques for enhancing DDoS attack detection in IoT systems to guarantee swift detection and reaction, especially under real-time scenarios.