Privacy-Preserving Face Trajectory Retrieval in Surveillance Video with an Authorization-Gated Reversible Reconstruction Mechanism

Abstract

Surveillance systems require balancing public privacy with law enforcement's need for forensic evidence recovery. Existing approaches either store raw face data (a privacy violation) or embed residuals into videos (no cross-camera search capability). VectorChain introduces a novel hybrid architecture that combines vector databases for real-time face-trajectory retrieval with warrant-gated reversible reconstruction. Face ROIs extract privacy-preserving embeddings (ArcFace+adversarial noise) indexed in Milvus for sub-second approximate nearest neighbor (ANN) search across CCTV networks. Matching trajectories link to IPFS-stored residuals via blockchain hashes, enabling pixel-perfect reconstruction only after warrant verification.