Invisible identities: Privacy-aware object detection for autonomous vehicles and public surveillance

Abstract

Anonymising visual data dynamically presents challenges for achieving privacy and security by design, as currently addressed by Privacy-Enhancing Technologies (PETs) and Security-Enhancing Technologies (SETs). The need to decrypt data before applying anonymization or masking techniques introduces potential vulnerabilities, where data could be exposed during processing. This article provides a comprehensive solution to implement robust security measures throughout the entire data lifecycle, including training machine learning models on data without applying differential privacy. Our approach ensures end-to-end encryption to protect data at rest, during processing, and in transit. We encrypt the region of interest in visual imagery at the source, allowing machine learning or deep learning models to use it directly for training purposes. This ensures models focus on abstract visual information rather than private details such as facial features or clothing colour. Our work demonstrates that machine learning models trained on privacy-aware data can be as accurate as those trained on raw data. We employed widely used AES encryption and object detectors like YOLO, Single Shot MultiBox Detector (SSD-MobileNetv1-FPN-640x640, SSD-ResNet101-FPN-640x640), and the EfficientDet family (EfficientDet-D0 and EfficientDet-D1) for detection and classification of visual objects.