Preventing Accidental Privacy Leakage in Ubiquitous Visual Sensing

The digital cameras are ubiquitous in our daily life and became an essential part of everyday devices. For example, there are cameras in cellphones, tablets, and many other wearable devices. This facilitates capturing and storing information around us through taking images and videos. Despite of the easiness and usefulness of cameras, they also bring a lot of serious privacy problems. These privacy problems can be either malicious intent or accidental user own ignorance, for example, malicious secret filming or unwanted recognition from a photo. To illustrate, the camera cannot distinguish whether part of the data captured is private or sensitive. Once the picture is exposed, accidental privacy leakage may happen. In many cases, we cannot avoid the malicious picture capture. However, we can avoid the information leakage in proactive way. Our goal in this thesis work is to prevent or minimize accidental privacy leakage. To achieve it, the basic idea is to mark the sensitive areas or objects by QR code, in order to encode the privacy information in such targets. Thus those areas and objects can be filtered out before publishing. Usually, face is a common case as the sensitive information, which we used in this thesis. In our implementation, the recognized faces were blurred from the original picture before publishing. We discussed the implementation of proactive protection of information from three aspects, including real time none-object based approach, offline none-object based approach and object based approach. In the real-time approach, QR code is used to implement real time processing. By using this feasible method, user can have fine grained control on what is revealed and what is kept private. In the offline method, a range selector is implemented to eliminate the specific area, even though there are still some defects of this approach. For example, it is not real time, and the lists of vii sensitive area differs from picture to picture. To help our implementation, this research utilized face recognition for the object based method as well. Compared with none object based sensing, this approach has less feasibility, lower speed, and less accuracy. The thesis categorizes existed ways by two types (permission requirement and predefine sensitive area) and compares these ways from different perspectives through showing the pros and cons for each side.Master of ScienceComputer and Information Science, College of Engineering and Computer ScienceUniversity of Michigan-Dearbornhttps://deepblue.lib.umich.edu/bitstream/2027.42/136612/1/YingZou_thesis.pdfDescription of YingZou_thesis.pdf : Thesi