Privacy-Preserving Federated Learning for Critical Infrastructure: A Systematic Review of Security Threat Models, Deployment Patterns, And Governance

Abstract

This study presented a quantitative systematic synthesis of privacy-preserving federated learning research in critical infrastructure contexts by examining how security threat models, privacy mechanisms, deployment architectures, and governance controls were jointly represented and evaluated across the empirical literature. A total of 120 peer-reviewed and high-quality technical studies were coded and analyzed using structured extraction protocols that transformed heterogeneous experimental reports into comparable variables. Descriptive analysis showed that malicious-client and colluding-client threat models dominated the evidence base (approximately 55% and 32%, respectively), while availability-focused threats were reported in fewer than 20% of studies. Passive inference attacks were evaluated more frequently (67.5%) than active attacks (40.8%), with confidentiality emerging as the most frequently targeted security property (65.8%). In terms of privacy mechanisms, secure aggregation (50.8%) and differential privacy (45.0%) were the most prevalent, whereas cryptographic computation approaches (21.7%) and trusted execution techniques (15.8%) appeared less often and were associated with higher reported overhead. Deployment analysis indicated strong dominance of centralized aggregation (70.8%) and cross-silo regimes (60.0%), reflecting institutional coordination patterns typical of critical infrastructure systems. Governance reporting was uneven, with role definitions (55.8%) and access rules (59.2%) appearing more frequently than consent artifacts (27.5%) and accountability procedures (17.5%). Regression analysis demonstrated that active threat exposure and cryptographic or hybrid privacy mechanisms were positively associated with standardized overhead burden, while cross-silo deployment and higher governance maturity scores were negatively associated with both overhead and utility loss. Governance maturity and auditability indices achieved acceptable-to-strong internal consistency (Cronbach’s α ranging from 0.72 to 0.86) and showed statistically significant relationships with outcome stability. Configuration prevalence analysis revealed that a limited set of design patterns dominated the literature, most commonly combining client-centric threat models, baseline privacy mechanisms, centralized architectures, and limited governance instrumentation. Overall, the study provided a variable-driven quantitative account of prevailing configurations, measured trade-offs, and reporting gaps in privacy-preserving federated learning for critical infrastructure, offering an empirical foundation for structured comparison across technical and institutional dimensions.