AI-Assisted Security Alert Data Analysis with Imbalanced Learning Methods
Samuel Ndichu, Tao Ban, Takeshi Takahashi, Daisuke Inoue
Abstract
Intrusion analysis is essential for cybersecurity, but oftentimes, the overwhelming number of false alerts issued by security appliances can prove to be a considerable hurdle. Machine learning algorithms can automate a task known as security alert data analysis to facilitate faster alert triage and incident response. This paper presents a bidirectional approach to address severe class imbalance in security alert data analysis. The proposed method utilizes an ensemble of three oversampling techniques to generate an augmented set of high-quality synthetic positive samples and employs a data subsampling algorithm to identify and remove noisy negative samples. Experimental results using an enterprise and a benchmark dataset confirm that this approach yields significantly improved recall and false positive rates compared with conventional oversampling techniques, suggesting its potential for more effective and efficient AI-assisted security operations.