DP-Protected LightGBM Framework for Smart Home Malicious Traffic Detection

With the maturation of IoT technology and the growing demand for improved quality of life, smart home systems are experiencing rapid development. The explosive growth in both the variety and quantity of smart home devices has brought increasingly severe cybersecurity challenges. However, most detection systems currently available on the market suffer from excessively high computational complexity, making them difficult to deploy in practical scenarios. Furthermore, many models lack sufficient security, allowing hackers to infer whether samples participated in model training by analyzing prediction probability distributions and confidence levels, leading to widespread issues such as information leakage. To address these challenges, this paper proposes a lightweight malicious information detection model integrating differential privacy training mechanisms. The model first performs data preprocessing, then applies differential privacy techniques to add noise to the training set, thereby resisting membership inference attacks. Finally, it employs the LightGBM model to detect malicious information within smart home devices. Experimental results demonstrate that even under low privacy budget constraints, the LightGBM model maintains low resource overhead while achieving high accuracy, precision, recall, F1 score, and AUC values.