Aprendizagem de máquina aplicada na predição de riscos cibernéticos

Abstract

Cyber risk stands as one of the prime threats to modern enterprises. Cyber-attacks can lead to substantial financial damages, loss of intellectual property, and even operational disruptions. This research aimed to explore the effectiveness of different Machine Learning models in predicting cyber risk. Using data obtained from IBGE (Brazilian Institute of Geography and Statistics), comprising 16 statistically significant covariates across 16,725 companies, a comparative analysis was conducted among neural networks, decision tree, GAMLSS, GBM, and random forest models. The findings revealed that neural networks achieved the highest agreement rate, with an 86% accuracy in identifying targeted companies. In contrast, GBM showed an agreement rate of 56%, marking the lowest among the mentioned methods. While GAMLSS didn't yield as promising results as neural networks, it demonstrated the capability to fit the data to a probability distribution, notably the Poisson Zero-Inflated (ZIP) distribution, with significant parameters for potential applications in cyber insurance pricing.