Optimization of LightGBM Model with Bayesian Optimization for Malware Detection

Abstract

Cyberattacks through malware on Android devices continue to rise, making accurate detection crucial. This research optimizes the LightGBM model using Bayesian Optimization to enhance accuracy and efficiency in detecting Android malware. A feature selection mechanism based on Attention Mechanism is applied to select the most relevant features for classification. The dataset used comes from the Canadian Institute for Cybersecurity (CIC) and consists of 17,804 Android applications, with a balanced distribution between malware and normal applications. The dataset is split into ratios of 80%:20%, 75%:25%, and 70%:30%. Feature selection reduces the number of features from 9503 to 300, 500, and 1000. The LightGBM model is then optimized with Bayesian Optimization to fine-tune parameters such as learning rate, number of iterations, and maximum number of leaves. The model's performance is evaluated using accuracy, precision, and recall metrics. Experimental results show that the model achieves 96,99% accuracy, 97,30% precision, and 96,70% recall with an 80%:20% dataset split and 1000 features. The combination of Attention Mechanism and Bayesian Optimization effectively improves processing efficiency without compromising performance.