Protection of image ROI using chaos-based encryption and DCNN-based object detection

Images always contain sensitive information, e.g., a clear face on a photo, which needs to be protected. The simple way is to encrypt the whole image for hiding “everything” securely, but it brings huge amounts of unnecessary encryption operations. Considering the most sensitive regions of an image, this paper focuses on protecting the important regions, thus reducing the redundant encryption operations. This paper employs the latest DCNN-based object detection model (YOLOv4) for choosing regions (i.e., multiple objects) and chaos-based encryption for fast encryption. We analyze object detection algorithm from a security perspective and modify YOLOv4 to guarantee that all areas of the detected objects are contained in the output regions of interest (ROI). Later, we propose a multi-object-oriented encryption algorithm to protect all the detected ROI at one go. We also encrypt the ROI coordinates and embed them into the whole image, relieving the burden of distributing ROI coordinates separately. Experimental results and security analyses show that all the detected objects are well protected.