内容共享环境中基于隐私保护的图像搜索方案

IMAGE SEARCH SCHEME IN CONTENT SHARING ENVIRONMENTS BASED ON PRIVACY PROTECTION

Flicker或YouTube等现代内容共享环境包含大量私人资源,这些资源可能非常敏感,甚至泄露用户大量隐私信息。为了支持用户在共享照片时作出隐私决策,提出一种隐私图像自动检测及隐私图像搜索技术。首先,基于经过人工评估的大量Flickr照片,对隐私分类器进行训练和学习,将图像文本元数据和多种视觉特征结合起来,然后,利用基于线性支持向量机的分类模型搜索隐私照片,并且使查询结果多样化,为用户更好地提供隐私和公共内容。最后,大规模分类实验验证了基于不同视觉和文本特征的隐私图像检测性能,另外,基于查询结果排名的用户评估结果也证明了该算法的可行性。

Modern content sharing environments such as Flicker or YouTube contain a large amount of privacy resources. These resources can be highly sensitive, even might disclose a great deal of users＇ privacy information. In order to support users in making privacy decisions in the context of image sharing, we propose a technique of private images automatic detection and search. First, we arrange the privacy classifi- ers to be trained and learnt based on a large set of manually assessed Flicker photos, thus combines the textual metadata of images with a vari- ety of visual features. Then, we employ the linear SVM-based classification models to search privacy photos, and diversify the query results so as to provide users with a better coverage of privacy and public contents. Finally, the large-scale classification experiments verify the detection performance of privacy image based on different visual and textual features. Besides, the result of user evaluation based on query result rank- ing also demonstrates the feasibility of our approach.