EDGE-CENTRIC PRIVACY-PRESERVING VIDEO ANALYTICS FOR LATENCY-SENSITIVE MONITORING

Real-time video analytics at the network edge is essential for smart surveillance, autonomous monitoring, and safety-critical decision making. However, maintaining high detection accuracy under severe latency limits and demonstrable privacy protection is difficult. Raw video transmission on the cloud increases delay and privacy risks, while entirely on-device solutions are limited by computational and energy resources.This paper presents a hybrid edge-centric video analytics architecture that balances latency, accuracy, and privacy via split inference, lightweight on-device processing, and privacy-preserving feature preservation. A compact backbone's early layers do motion filtering, region-of-interest extraction, and feature computing on the device, while a nearby edge server performs deeper inference. Quantization and differential privacy safeguard intermediate characteristics, with selective homomorphic encryption applied to tiny crucial layers for better guarantees. Based on bandwidth and device demand, an adaptive runtime controller chooses the split point and compression level. On VIRAT, PEViD, and Sultani-derived datasets, the suggested technique achieves sub-100 ms median end-to-end latency and near-cloud detection accuracy. Results indicate that modest differential privacy (ε ≈ 1) significantly reduces membership-inference attack success with minimal accuracy loss, whereas full homomorphic encryption is too latency-intensive. Split inference and lightweight differential privacy create a realistic and deployable privacy-aware real-time edge video analytics solution.