Design of an AI-Based Anomalous Behavior Detection Model Using Access Records in Public Office Buildings and Exploration of Implementation Requirements

This study aims to enhance the crime prevention capability of public office buildings by designing an artificial intelligence (AI)-based anomalous behavior detection model using access-log big data and by exploring its implementation requirements and practical feasibility from a qualitative research perspective. Existing physical security systems in public office buildings are largely limited to checking individual access terminal records and relying on passive, post-incident monitoring by control rooms, making it difficult to proactively prevent abnormal stays or deviations from normal movement paths. To address this limitation, the study proposes a foundational spatio-temporal AI model architecture that integrates temporal continuity, modeled with Long Short-Term Memory(LSTM), and spatial interactions, modeled with Graph Neural Networks(GNN), within a converged security framework that combines access-log data in cyberspace with physical security resources. As a research method, in-depth interviews were conducted with 15 relevant experts, and qualitative content analysis was performed. The results show, first, that five core indicators for risk scoring in the AI model—time appropriateness, spatial path anomaly, duration anomaly, frequency anomaly, and user-type appropriateness—were derived. The weights(coefficients) assigned to these indicators are exploratory initial values that reflect expert opinions obtained through qualitative in-depth interviews, rather than the outcome of quantitative statistical estimation. Second, a four-stage response process was designed that links the calculated risk levels to automatic CCTV activation in control rooms and to dynamic patrol deployment by security personnel. Third, the study identifies, as key prerequisites for practical implementation, the need for log standardization across heterogeneous access control systems, de-identification of access logs and access-control governance for personal data protection, and an AI operation architecture compatible with network-segregated environments. By presenting a foundational design and policy implications for shifting the security system of public office buildings from experience-based, post-incident responses to a proactive, prediction-oriented approach grounded in access-log data, this study proposes a data-driven, dynamic Crime Prevention Through Environmental Design(CPTED) model that can contribute to crime prevention and security enhancement across public administrative facilities.