PROBLEMS OF OPTIMIZING MACHINE LEARNING MODELS FOR CONTINUOUS PROTECTION OF CORPORATE INFORMATION SYSTEMS

Abstract

The work investigates the mathematical and algorithmic problems of optimizing machine learning (ML) models used for continuous monitoring and protection of modern corporate information systems (CIS). The growth of network traffic volumes and the transition to the Zero Trust architecture require security systems to analyze data in real-time, leading to a critical increase in computational load. The key barriers to ML implementation are considered, in particular, the high dimensionality of the feature space, the problem of class imbalance, and latency during inference. The necessity of applying dimensionality reduction methods and ensemble approaches to increase the accuracy of detecting multi-vector cyberattacks while reducing the false positive rate is justified. Directions of mathematical modeling for loss function minimization under dynamic changes in CIS are proposed.