Faces are Protected as Privacy: An Automatic Tagging Framework Against Unpermitted Photo Sharing in Social Media

Abstract

On social platforms like Facebook, it is popular and pleasurable to share photos among friends, but it also puts other participants in the same picture in jeopardy when the photos are released online without the permission from them. To solve this problem, recently, the researchers have designed some fine-grained access control mechanisms for photos shared on the social platform. The uploader will tag each participant in the photo, then they will receive internal messages and configure their own privacy control strategies. These methods protect their privacy in photos by blurring out the faces of participants. However, there is still some defect in these strategies due to the unpredictable tagging behaviors of the uploader. Malicious users can easily manipulate unauthorized tagging processes and then publish the photos, which the participants want them to be confidential in social media. To address this critical problem, we propose a participant-free tagging system for photos on social platforms. This system excludes potential adversaries through automatic tagging processes over two cascading stages: 1) an initialization stage will be applied to every new user to collect his/her own portrait samples for future internal searching and tagging, and; 2) the remaining unidentified participants will be tagged in cooperative tagging stage by the users who have been identified in the first stage. For the system evaluation of efficiency and effectiveness, we conducted a series of experiments. The results demonstrated the tagging efficiency (96% tagging rate and 0.77s/photo tagging speed on average), photo masking and unmasking efficiency (0.13s/face on average), and the privacy preserving performance (over 90% identities in both group and individual photo are preserved).INDEX TERMS Social media, face tagging, privacy protection, system security.

Real image improvement study based on pivotal tuning inversion

“…From the paper, table 1 shows the overall loss: From any perspective of comparing the similarity of the two photos, the PTI is a better choice. However, the visualization of Figure 11 shows that it is unfeasible to reduce the loss to zero [10].…”

Sparse‐Dyn: Sparse dynamic graph multirepresentation learning via event‐based sparse temporal attention network

“…Thus, it is crucial to balance the total number of patches, N , and the quality of structural embedding of each patch along the time dimension. We utilize a cost function of ADE to adaptively determine N and the embedding of patches in Equations ( 1) and (2).…”

“…Dynamic graph neural networks (DGNNs) have seen a notable surge of interest with the encouraging technique for learning complicated systems of relations or interactions over time. Since DGNNs append an additional temporal dimension to accumulate the variation of embedding or representations, they are powerful tools to employ in diverse fields, such as social media, 1,2 dynamic social networks, 3 bioinformatics, 4 knowledge bases, 5 brain neuroscience, 6 protein–protein interaction networks, 7 recommendation system, 8,9 image processing, 10–12 remote sensing, 13–15 information safety, 16–18 reinforcement Learning, 19,20 and so forth.…”

Foreground–background decoupling matting

“…Glance and Focus Matting (GFM) 11 proposes a parallel framework inspired by the comprehensive empirical analysis of the component pipelines in image matting, and it designs a composition route including resolution discrepancy, semantic ambiguity, sharpness discrepancy, and noise discrepancy to reduce the domain gap due to the differences in resolution, sharpness, noise, and so forth. More and more works [38][39][40][41][42] focus on privacy-preserving, privacy-preserving portrait matting (P3M) 14 further expands the GFM to the privacy benchmark by adding the information exchange between parallel branches. However, those methods model trimap-free matting as global segmentation and detail matting as shown in Figure 1.…”