Abstract
Recently, an increasing number of ROI (regions of interest) encryption algorithms have been proposed to efficiently encrypt the sensitive regions of image. Due to the powerful feature extraction capabilities of deep learning (DP), many DP-based object detection models have been increasingly applied to ROI encryption. However, some models with a large number of parameters are inefficient and not suitable for real-time detection, and the detected ROI often include some redundant regions. Moreover, the following encryption operations are always in serial manner, leaving room for improvement. To address these issues, we present a semantically enhanced selective image encryption scheme with parallel computing. The deep salient object detection (SOD) model is first lightweighted to improve detection efficiency. Then, the sensitive region is cropped based on the boundary information from the output saliency map, resulting in an ROI that removes redundant regions without revealing sensitive object information. In encryption stage, the three color channels of each pixel are assigned to a group and encrypted in parallel to further improve the efficiency. Furthermore, to enhance the practicality, we embedded the side information of the ROI into the image, eliminating the need to separately distribute the image and the corresponding side information. Finally, we carry out security and efficiency analyses, and the results demonstrate that the proposed encryption scheme can enable efficient and secure detection of sensitive regions, along with corresponding encryption protection.
•Lightweight object detection model ensures efficient and accurate detection of sensitive regions.•The proposed parallel encryption algorithms further improve the encryption efficiency.•Reversible steganography enhances the practicality by embedding side information.•The experimental results demonstrate the proposed scheme is secure and efficient.