@article{ART003348208},
author={Han, Seungwan and Kang Seung Ho},
title={Privacy Leakage Evaluation and the Effect of Access-Controlled Abstraction for VLM-based CCTV Surveillance Captions},
journal={ Journal of Software Forensics},
issn={3092-541X},
year={2026},
volume={22},
number={2},
pages={57-65},
doi={10.29056/jsf.2026.06.06}
TY - JOUR
AU - Han, Seungwan
AU - Kang Seung Ho
TI - Privacy Leakage Evaluation and the Effect of Access-Controlled Abstraction for VLM-based CCTV Surveillance Captions
JO - Journal of Software Forensics
PY - 2026
VL - 22
IS - 2
PB - Korea Software Assessment and Valuation Society
SP - 57
EP - 65
SN - 3092-541X
AB - As Vision-Language Models (VLMs) are adopted in CCTV surveillance systems, video content can be generated and retrieved as natural-language captions. However, such captions may also become a new text-based privacy leakage channel. This study proposes the Semantic Leakage Score (SLS), an eight-attribute sensitive semantic taxonomy, and a four-level access-controlled abstraction policy from A0 to A3 to quantify and mitigate semantic privacy leakage in VLM-based CCTV captions. Experiments on 300 English UCA(UCF-Crime Annotation) captions show that A2 reduces SLS from 0.118 to 0.013, an 89.0% reduction, while maintaining a weak keyword-matching utility baseline comparable to A0. In contrast, A3 reduces SLS to zero but lowers classification accuracy to the 12-class random-guessing level. An auxiliary AI Hub case study further shows that residual exposure may remain in information-dense Korean captions even after A2 abstraction. These results indicate that access-controlled abstraction is a practical method for balancing surveillance utility and privacy in VLM-based CCTV systems.
KW - CCTV surveillance;Vision-Language Model;semantic privacy leakage;textual de-identification;access-controlled abstraction
DO - 10.29056/jsf.2026.06.06
ER -
Han, Seungwan and Kang Seung Ho. (2026). Privacy Leakage Evaluation and the Effect of Access-Controlled Abstraction for VLM-based CCTV Surveillance Captions. Journal of Software Forensics, 22(2), 57-65.
Han, Seungwan and Kang Seung Ho. 2026, "Privacy Leakage Evaluation and the Effect of Access-Controlled Abstraction for VLM-based CCTV Surveillance Captions", Journal of Software Forensics, vol.22, no.2 pp.57-65. Available from: doi:10.29056/jsf.2026.06.06
Han, Seungwan, Kang Seung Ho "Privacy Leakage Evaluation and the Effect of Access-Controlled Abstraction for VLM-based CCTV Surveillance Captions" Journal of Software Forensics 22.2 pp.57-65 (2026) : 57.
Han, Seungwan, Kang Seung Ho. Privacy Leakage Evaluation and the Effect of Access-Controlled Abstraction for VLM-based CCTV Surveillance Captions. 2026; 22(2), 57-65. Available from: doi:10.29056/jsf.2026.06.06
Han, Seungwan and Kang Seung Ho. "Privacy Leakage Evaluation and the Effect of Access-Controlled Abstraction for VLM-based CCTV Surveillance Captions" Journal of Software Forensics 22, no.2 (2026) : 57-65.doi: 10.29056/jsf.2026.06.06
Han, Seungwan; Kang Seung Ho. Privacy Leakage Evaluation and the Effect of Access-Controlled Abstraction for VLM-based CCTV Surveillance Captions. Journal of Software Forensics, 22(2), 57-65. doi: 10.29056/jsf.2026.06.06
Han, Seungwan; Kang Seung Ho. Privacy Leakage Evaluation and the Effect of Access-Controlled Abstraction for VLM-based CCTV Surveillance Captions. Journal of Software Forensics. 2026; 22(2) 57-65. doi: 10.29056/jsf.2026.06.06
Han, Seungwan, Kang Seung Ho. Privacy Leakage Evaluation and the Effect of Access-Controlled Abstraction for VLM-based CCTV Surveillance Captions. 2026; 22(2), 57-65. Available from: doi:10.29056/jsf.2026.06.06
Han, Seungwan and Kang Seung Ho. "Privacy Leakage Evaluation and the Effect of Access-Controlled Abstraction for VLM-based CCTV Surveillance Captions" Journal of Software Forensics 22, no.2 (2026) : 57-65.doi: 10.29056/jsf.2026.06.06
