Collaborative Research: SaTC: CORE: Medium: Novel Algorithms and Tools for Empowering People Who Are Blind to Safeguard Private Visual Content

盲人经常使用个人设备拍摄照片和视频，并与他人分享。主要原因是在日常视觉任务中寻求帮助(例如，识别物体、阅读邮件)和在线社交。不管是什么原因，他们都没有简单或独立的手段来评估他们即将分享的图像或视频是否无意中包含私人信息。对于这个项目，一个跨学科的团队将设计和评估一个自动助手，该助手将提醒用户他们的视觉媒体中的隐私泄露，并编辑他们的图片和视频，以混淆任何潜在的私人内容。如果成功，这些贡献将首次使盲人能够独立避免意外的视觉隐私泄露，从而使他们能够过上更加独立和互联的生活。这个项目对更广泛的人群也有帮助，因为他们有时会在可视内容中忽略私人信息。项目组预计，这项工作将在稍作适应的情况下推广，使其他人群受益，如视力低下的人、认知障碍的人、老年人和儿童。该项目将涉及设计新颖的计算机视觉算法和最终用户机制，使盲人能够独立保护他们照片和视频中的私人信息。具体地说，它将由三个关键任务组成：(1)创建后端计算机视觉算法，其通过仅观察图像和视频中的几个例子来学习定位图像和视频中的私人内容(即，少镜头语义分割)；(2)创建学习在图像和视频中定位前景对象的后端计算机视觉算法(即，前景对象分割)，以便支持仅保留该内容；以及(3)为可访问的前端接口建立有效的设计选择，该设计选择产生给定算法性能的适当的信任级别。该团队将通过共享生成的文物来促进这项工作的未来扩展，包括隐私保护算法的代码、隐私保护技术设计建议的使用特征和工作原型。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

People who are blind regularly use personal devices to take pictures and videos and share them with others. Primary reasons are to seek assistance with everyday visual tasks (e.g., recognizing objects, reading mail) and to socialize online. Regardless of the reason, they have no easy or independent means of assessing whether an image or video they are about to share inadvertently contains private information. For this project, an interdisciplinary team will design and evaluate an automated assistant that alerts users of privacy disclosures in their visual media and edits their pictures and videos to obfuscate any potential private content. If successful, these contributions will enable blind individuals for the first time to independently avoid accidental visual privacy leaks, thereby empowering them to live more independent and connected lives. This project can also be helpful to the broader population as they sometimes overlook private information in their visual content. The project team anticipates this work would generalize, with minor adaptations, to benefit other populations such as people with low vision, people with cognitive impairments, aging adults, and children. This project will involve designing novel computer vision algorithms and end-user mechanisms that empower people who are blind to independently safeguard private information in their pictures and videos. Specifically, it will consist of three key tasks: (1) creating back-end computer vision algorithms that learn to locate private content in images and videos by observing only a few examples of it (i.e., few shot semantic segmentation), (2) creating back-end computer vision algorithms that learn to locate the foreground object in images and videos (i.e., foreground object segmentation) in order to support retaining only that content, and (3) establishing effective design choices for an accessible front-end interface that engenders a level of trust that is appropriate given the algorithms' performance. The team will facilitate future extensions of this work by sharing the generated artifacts including the code for privacy-preserving algorithms, characterization of use with design recommendations for privacy-preserving technology, and a working prototype.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.