VERA: A Privacy-Preserving Framework for Deep Learning Data Collection and Object Detection in Private Settings

This paper introduces VERA (Vision Expert Real Analysis), a privacy-supporting cyber-physical framework designed for real-time data collection and visual analysis in healthcare environments. VERA limits exposure to identifiable RGB content by ensuring that annotators interact only with non-identifiable edge-based representations, while original images remain encrypted at rest using AES-CFB, with integrity verification performed before in-memory decryption. The system integrates edge-based obfuscation, secure annotation, in-memory decryption, and dynamic data augmentation to train YOLO-based person detection models without compromising patient privacy. Experimental results on a curated COCO subset show that VERA enables effective person detection, improving mean Average Precision (mAP) from an intentionally minimal baseline of 0.61 percent to 99.94 percent after full training and augmentation. This baseline is used solely to illustrate the contribution of the secure data preparation pipeline and is not intended to represent a fully optimized YOLO configuration. The results demonstrate that privacy-supportive workflows can maintain strong model performance while aligning with data protection practices common in regulated environments. Although this work focuses on person detection as a foundational stage, the VERA architecture is designed to support future extensions toward privacy-preserving Human Activity Recognition (HAR) tasks in clinical and assisted-living settings.