Yunhao Yao, Zhiqiang Wang, Ruiqi Li, Haoran Cheng, Puhan Luo, xiangyang Li · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840584
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
As the Internet of Things (IoT) becomes deeply embedded in daily life, users are increasingly concerned about privacy leakage, especially from video, which holds rich, sensitive information. While many techniques can protect privacy, frame-by-frame protection in large-scale video analytics (e.g., smart communities) introduces significant latency. A more efficient solution is to first detect frames containing privacy objects (e.g., faces) and then apply selective protection. However, pixel-domain detection requires fully decoding video frames, incurring decoding overhead. Compressed-domain approaches directly analyze compressed video streams to accelerate inference, but often sacrifice accuracy, and the per-frame processing remains inefficient. Therefore, we propose ComPrivDet, an efficient framework for detecting privacy objects in compressed video by reusing I-frame inference results. ComPrivDet provides a pixel-domain detector on I-frames and a lightweight compressed-domain detector on P/B-frames. Leveraging compressed-domain cues to determine the presence of new objects, ComPrivDet either reuses I-frame results to skip P/B-frame detections or refines them with a lightweight detector. We prototype ComPrivDet on an edge server with an NVIDIA GeForce RTX 3090 GPU. ComPrivDet maintains 99.75% accuracy for private face detection and 96.83% for private license plate detection while skipping over 80% of inferences. It averages 1.36% higher accuracy with 57.23% lower inference latency than pixel-domain baselines, and 10.93% higher accuracy with 44.65% lower latency than compressed-domain baselines.
Last synced
This work has 1 recorded citation, but citing papers have not been linked locally yet.
No comments yet — start the discussion below.