A privacy aware lightweight edge intelligent framework for real-time lethal threat interception and risk assessment

Real-time intelligent surveillance systems require a careful balance between predictive reliability, computational efficiency, energy consumption, inference latency, and privacy preservation. However, existing vision-based approaches often achieve high detection accuracy at the expense of computational complexity, limiting their deployment on resource-constrained edge devices. To address this challenge, we propose a privacy-preserving, lightweight edge intelligence framework guided by a multi-objective architectural design methodology that jointly balances accuracy, computational efficiency, energy efficiency, real-time responsiveness, and on-device privacy.​ The proposed framework combines hybrid local-global representation learning with a lightweight multi-scale spatial attention mechanism to enhance the recognition of small and partially occluded lethal threats. A lightweight Temporal Threat Validation Module suppresses transient false alarms through temporal consistency analysis, while a context-aware Threat Assessment Network transforms validated detections into quantitative threat intelligence for priority-based decision-making. The complete pipeline executes locally on a Raspberry Pi 5 integrated with a Hailo-8L neural accelerator, transmitting only encrypted threat metadata to ensure low-latency, privacy-preserving operation.Experimental results demonstrate a Precision of 95.36%, Recall of 93.74%, an F1-score of 94.54%, and an [email protected] of 96.58%. The proposed framework reduces localization error by 34.1%, decreases false alarms by 60.9%, and increases actionable alerts to 92.8% while sustaining 31.8 FPS, 31.4 ms inference latency, and an average power consumption of 9.1 W on the edge platform. These results demonstrate that the proposed framework effectively unifies lightweight representation learning, temporal reasoning, contextual threat intelligence, and privacy-preserving edge computing, providing an energy-efficient and practically deployable solution for next-generation real-time intelligent public safety systems.