Explainable AI for Adaptive Security in Regulated Environments: A Unified Framework Integrating Federated Risk-Based Authentication and Privacy-Preserving On-Premises LLM Deployment
TL;DR
A unified Explainable AI (XAI) security framework that synthesizes federated learning-enhanced Dynamic Risk-Based Authentication with privacy-preserving on-premises Large Language Model (LLM) deployment into a cohesive, compliance-native paradigm is proposed.
Abstract
Healthcare and public-sector organizations face a compounding security imperative: protecting sensitive personal data against adaptive cyber threats while preserving the operational fluency that practitioners require in time-critical workflows. This paper proposes a unified Explainable AI (XAI) security framework that synthesizes federated learning-enhanced Dynamic Risk-Based Authentication (FL-RBA) with privacy-preserving on-premises Large Language Model (LLM) deployment into a cohesive, compliance-native paradigm. The framework introduces SHAP-based rationale generation to address the interpretability gap inherent in transformer-based authentication scoring, producing audit-ready decision trails that satisfy HIPAA and GDPR requirements. Pilot evidence from a live healthcare deployment demonstrates a 95% high-risk interception rate, 2.5% false-positive rate, and sub-1.2-second decision latency. On-premises LLM integration eliminates all external data transmission while delivering clinical intelligence capabilities comparable to cloud-based alternatives. The unified model provides a replicable blueprint for regulated organizations seeking to operationalize AI without compromising data sovereignty or regulatory alignment.