A Federated Learning Framework for Privacy Preserving Threat Detection in Zero Trust Network Access (ZTNA)

Zero Trust Network Access (ZTNA) has become one of the core cybersecurity strategies through the implementation of a continuous verification process and the least-privilege access management in disperse systems.Current ZTNA threat detection methodologies rely primarily on centralized machine learning models, which face issues of scalability, increased latency and major privacy issues with the centralized collection of sensitive endpoint in-formation.To overcome these issues, this paper presents a federated learning (FL) improved ZTNA frame-work for privacy preserving threat detection.In the proposed system, collaborative training is being performed by endpoint devices in which they train a shared machine learning training model by performing local training using endpoint telemetry and only sending privacy-protected model updates to a central aggregator (instead of raw data).A hybrid model of threat detection that uses a combination of Long Short-Term Memory network, auto encoder and Random Forest classifiers are used to extract temporal behavior, anomaly behavior and contextual threat behavior.Experimental evaluations performed on public benchmark data sets and large-scale enterprise telemetry show that proposed FL-based approach achieves the similar detection accuracy as centralized models (within 1.5-3.0%),while reducing measured privacy leakage by approx.35-40% and keeping communication overhead at acceptable level.These results show that federated learning offers a scalable and privacy-aware method for real-time threat detection in ZTNA environments in order to enable improved security without data locality or compliance with regulations.

Hali tarjima qilinmagan