Federated Learning Architecture for Self-Improving Biometric Access Control in Office Buildings

Biometric identification has been practised for centuries. Facial recognition, the most intuitive and non-invasive modality, is now ubiquitous in consumer devices. Extending this convenience to office buildings with multiple access tiers promises frictionless entry while tightening security. Integrating continuously learning computer-vision models into an access-control system increases architectural complexity and raises strict privacy and regulatory concerns. Efficiency, robustness, and legal compliance must be balanced at design time. Drawing on recent literature, we derive functional and non-functional requirements for a self-improving, camera-based access system. The system is modeled with arc42 and SysML, emphasizing information flows and using sequence diagrams to depict the runtime model, including registration, the happy path with RFID fallback, and federated update and rollback. Federated Learning enables edge devices to update locally and exchange only encrypted parameters and aggregated metrics, while embeddings are extracted and and never leave the access nodes. The paper provides an architectural blueprint for a continuously learning biometric access system for office environments, a concise requirements catalogue spanning privacy, security, and lifecycle quality management, and sequence diagrams and deployment views that operationalise continuous learning through edge inference, verified annotations, audit logging, and rollback without transferring raw biometrics. The concept targets long-term performance improvements without eroding trust or compliance.