Detection of Personal Privacy Information in Social Data Taking Into Account Data Augmentation and Mask R‐ CNN
Year: 2026
Type: article
Abstract: ABSTRACT Visual privacy‐sensitive objects in social images are often small, partially occluded, and unevenly represented in available training data, which limits the reliability of automatic privacy detection. This study develops a privacy‐sensitive object detection framework combining multi‐strategy image augmentation with SE‐Mask R‐CNN. Conventional photometric and geometric transformations are complemented by a DCGAN‐style generator to increase sample diversity, while SE modules are introduced into the ResNet50 backbone to recalibrate channel responses. A pseudo‐mask‐supervised branch is further incorporated to improve spatial localization of facial privacy regions. Experiments on the Human face subset of Open Images V6 show that GAN‐based augmentation reduces FID to 24.3 and increases SSIM to 0.92. The downstream detector trained with GAN‐augmented samples achieves an mAP of 78.90% and a recall of 82.50%. The complete SE‐Mask R‐CNN obtains an mAP of 0.88 and a small‐object AP of 0.79, compared with 0.72 and 0.58 for Faster R‐CNN and 0.81 and 0.68 for Mask R‐CNN, respectively. These results indicate that sample‐diversity enhancement, channel recalibration, and mask supervision provide complementary improvements for detecting small facial privacy targets in complex social images.
Source: Security and Privacy
DOI: 10.1002/spy2.70255
Author: Shidong Li
Institution: North University of China
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7215834648