EdgeDiffusion: Latency-Optimized Diffusion Models for Real-Time On-Device Visual Anonymization

Zihao Han, Xueyan Gong, Fengyi Wu, Yifei Dong, Jingqi Xu, Jingdong Sun, Feng Liu, Siyu Huang, Zhi-Qi Cheng

Keywords: visual privacy, anonymization, diffusion models, edge computing, real-time vision, on-device generation

Paper Track: Long Paper (archival)

TL;DR: A latency-optimized diffusion framework enabling real-time visual anonymization of faces, license plates, and text directly on edge devices.

Abstract: Protecting visual privacy in surveillance systems requires anonymization methods that are both effective and computationally efficient, especially when deployed on resource-constrained edge devices. Existing approaches often rely on server-side processing, exposing sensitive visual data during transmission, or degrade perceptual fidelity to meet strict latency requirements. We present EdgeDiffusion, a diffusion-based framework for real-time, on-device anonymization of privacy-sensitive regions, including faces, license plates, and scene text. The framework combines three components: (i) a lightweight multi-task detector optimized for high-recall identification of privacy-sensitive regions, (ii) a distilled diffusion backbone that generates anonymized content in as few as 4–8 denoising steps, and (iii) category-aware anonymization modules with an adaptive fidelity controller that balances privacy protection and downstream visual utility. Anonymization effectiveness is evaluated using a Privacy Accuracy (PA) metric, which measures the reduction in successful re-identification by pretrained face, text, and license-plate recognition models after anonymization. Experiments on COCO-Privacy, Cityscapes-Privacy, and VIRAT-Privacy show that EdgeDiffusion achieves real-time inference (<50 ms per frame) on commercial edge hardware while improving PA by up to 12.4% over prior real-time baselines. EdgeDiffusion produces visually consistent anonymized outputs with limited artifacts while largely preserving scene context, suggesting its potential for privacy-preserving perception in applications such as intelligent transportation, healthcare monitoring, and smart-city analytics.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 26