Evaluating Deployment of Deep Learning Model for Early Cyberthreat Detection in On-Premise Scenario Using Machine Learning Operations Framework
Andrej Ralbovský, Ivan Kotuliak, Dennis Sobolev
Modern on-premises threat detection increasingly relies on deep learning over network and system logs, yet organizations must balance infrastructure and resource constraints with maintainability and performance. We investigate how adopting MLOps influences deployment and runtime behavior of a recurrent-neural-network–based detector for malicious event sequences. Our investigation includes surveying modern open-source platforms to select a suitable candidate, its implementation over a two-node setup with a CPU-centric control server and a GPU worker and performance evaluation for a containerized MLOps-integrated setup vs. bare metal. For evaluation, we use four scenarios that cross the deployment model (bare metal vs. containerized) with two different versions of software stack, using a sizable training corpus and a held-out inference subset representative of operational traffic. For training and inference, we measured execution time, CPU and RAM utilization, and peak GPU memory to find notable patterns or correlations providing insights for organizations adopting the on-premises-first approach. Our findings prove that MLOps can be adopted even in resource-constrained environments without inherent performance penalties; thus, platform choice should be guided by operational concerns (reproducibility, scheduling, tracking), while performance tuning should prioritize pinning and validating the software stack, which has surprisingly large impact on resource utilization and execution process. Our study offers a reproducible blueprint for on-premises cyber-analytics and clarifies where optimization yields the greatest return.