A Privacy-Risk-Driven Adaptive ECC Key Management Framework for Surveillance Video ROIs

Surveillance video streams continuously expose highly sensitive biometric data primarily human faces and vehicle license plates posing acute risks of identity disclosure and regulatory non-compliance under frameworks such as the GDPR.Existing selective encryption schemes apply a uniform cryptographic strength to all detected regions of interest, failing to account for the wide variation in per-region privacy sensitivity.This study presents a privacy-riskdriven adaptive elliptic curve cryptography (ECC) key management framework that dynamically calibrates the cryptographic strength according to a formally defined multi-factor privacy risk score.The risk score integrates the object class weight, spatial salience, YOLOv8 detection uncertainty, and temporal persistence, with weights systematically derived using the Analytic Hierarchy Process.Based on the computed privacy tier, the framework deterministically selects an appropriate elliptic curve (Curve25519 to NIST P-521), performs an ephemeral ECDH key exchange, and derives cryptographically independent per-ROI session keys via HKDF by embedding the risk score in the info field and the ByteTrack tracking identifier in the salt.The framework was quantitatively evaluated on approximately 32,000 regions of interest from the WIDER FACE validation set (with license plate ROIs extracted using OpenALPR) and qualitatively demonstrated on real-world surveillance sequences.It achieves a mean keygeneration latency of 0.90 ms per ROI on a workstation (Intel Core i9-13900K, NVIDIA RTX 4090) and 4.2 ms per ROI on NVIDIA Jetson Orin NX edge hardware while sustaining ≥32 fps on the edge device.Formal security analysis under the Computational Diffie-Hellman assumption proves resistance to key recovery, downgrade, and per-ROI inference attacks.Compared with uniform strong-curve baselines, the proposed framework delivers a 2.38× reduction in the total key-generation overhead and reduces under-encryption at the key-assignment level to 0.0% in our tier mapping, supplying cryptographically robust, tier-matched session keys precisely where the privacy risk is the highest.