A Study of a Frequency-Separation-Based and Access-Controlled Reversible Video De-identification System for Edge Environments

초록

This study proposes an edge-computing-based reversible video de-identification system for human-centered surveillance environments, focusing on facial regions as the primary PII target. The proposed system decomposes an input video into low-frequency and high-frequency components. The low-frequency component is used for real-time surveillance analysis, while the high-frequency component, which may contain face-related details, is converted into temporal residual metadata and selectively encrypted instead of being removed. The server generates access-controlled outputs according to user privileges, including low-frequency de-identified, partially restored, and fully restored videos. Experimental results show that the proposed system achieved an average end-to-end latency of 29.35 ms and 34.7 FPS, demonstrating its feasibility for real-time edge-based processing. The low-frequency video reduced Face Similarity to 0.2515 while maintaining a Detection Recall of 100.0%, indicating that face-based identifiability was reduced while object detection utility was preserved. In the access-controlled restoration evaluation, the facial ROI remained protected under the partial restoration condition, while the non-ROI area was restored with an SSIM of 1.0000. These results demonstrate that the proposed system can reduce facial PII exposure while supporting surveillance utility and privilege-based restoration in real-time edge surveillance environments. © 2026 IEEE.

키워드