The rapid escalation of cyber threats in decentralized environments has necessitated the development of collaborative defense mechanisms that do not compromise data sovereignty. Traditional centralized machine learning requires the aggregation of sensitive telemetry data, creating significant privacy risks and regulatory hurdles. This review explores the paradigm of Federated Learning (FL) as a transformative solution for privacy-preserving security systems. By enabling the training of global threat detection models across distributed nodes—such as edge devices, corporate branches, or mobile endpoints—without transferring raw data to a central server, FL addresses the fundamental tension between collective intelligence and individual privacy. This article categorizes current FL architectures, including horizontal, vertical, and transfer-based federated systems, and examines their application in intrusion detection, malware analysis, and anomaly-based behavioral monitoring. We analyze the integration of Differential Privacy and Secure Multi-Party Computation within the FL pipeline to mitigate data leakage from model updates. Furthermore, the review addresses the challenges of communication overhead, non-independent and identically distributed (non-IID) data, and vulnerability to poisoning attacks. By synthesizing recent research and industrial implementations, this paper provides a strategic roadmap for the deployment of self-evolving, privacy-aware security frameworks. The findings suggest that Federated Learning not only complies with stringent data protection mandates like GDPR but also enhances model robustness by training on diverse, real-world datasets that were previously inaccessible due to privacy constraints.