The rapid proliferation of Internet of Things (IoT) deployments in safety-critical environments has intensified the demand for intelligent emergency detection mechanisms capable of operating reliably under stringent resource and latency constraints. Conventional cloud-centric analytics often introduce communication delays, increased energy consumption, and potential security vulnerabilities, thereby limiting their suitability for time-sensitive emergency scenarios. To address these limitations, this study proposes a secure TinyML-driven edge intelligence framework designed to perform real-time emergency detection directly on resource-constrained embedded devices. The proposed architecture integrates lightweight convolutional neural network (CNN) and decision tree models optimized through quantization-aware training and structured pruning to ensure efficient inference on low-power microcontrollers. Experimental evaluation was conducted using a combination of publicly available sensor datasets, including environmental hazard and human activity recognition data, supplemented with locally collected multi-sensor readings from gas, temperature, and motion sensors deployed on an ESP32-based edge node. The system was implemented using TensorFlow Lite for Microcontrollers within an embedded C environment, with encrypted communication protocols ensuring secure data transmission and device authentication. Performance analysis demonstrates that the optimized TinyML models achieve detection accuracy exceeding 96% while maintaining inference latency below 120 milliseconds and reducing energy consumption by approximately 35% compared to conventional edge inference baselines. The findings confirm that secure TinyML-enabled edge intelligence can significantly enhance the responsiveness, reliability, and operational efficiency of emergency detection systems in constrained IoT settings. This work contributes a practical and scalable framework that bridges the gap between lightweight machine learning, embedded security mechanisms, and real-time emergency response in next-generation edge computing infrastructures.

Keywords: TinyML, Edge Intelligence, Emergency Detection, IoT Security, Lightweight Machine Learning, Real-Time Systems, Resource-Constrained Devices