Extra-Lightweight AI-Based Privacy Preserving Framework for Egocentric Wearable Cameras

Wearable egocentric cameras are gaining increasing popularity for monitoring eating behaviors, as they enable users to track eating habits and analyze dietary intake, thereby offering valuable insights into dietary patterns and supporting healthier lifestyle. However, they create privacy concerns, as cameras can inadvertently capture sensitive visual information, such as faces of bystanders. Despite the growing prevalence of these devices, no existing research addresses privacy preservation for wearable cameras with extremely constrained resources. Existing privacy protection require intensive computational resources, making them unsuitable for wearables. In this work, we present a novel, extra-lightweight AI-based privacy preserving framework (Egocentric Privacy-preserving Intelligent Camera: EPIC), tailored to resource-constrained wearable cameras. Our approach employs a styololc_v1 model, deployed via the CubeAI platform, for face detection in captured images and obfuscates detailed information to make them unrecognizable before storing or sending the images, effectively safeguarding user privacy. We implemented EPIC framework in an eating behavior monitoring application running on an STM32L4 microcontroller unit (MCU) with only 640 KB of RAM. Experimental result show that EPIC framework achieves a 41.38% mean average precision (mAP) for face detection and successfully eliminated sensitive information to preserve privacy. This work highlights the feasibility of advanced privacy-preserving solutions for low-power wearable devices with severely limited resources, addressing a critical gap in the field. By bridging this gap, our framework paves the way for more secure and privacy-conscious wearable technologies in the future.