Dissertation Defense Files

• Slides
  • Quicktime (recommended)
    • This version is a smaller download and contains the transitions as I intended them. You can control slide changes with mouse clicks or with the keyboard (space bar or arrow keys). For the keys to work, you have to click on the slides at least once (at least that's the case in my copy of Quicktime Player). If it wants to play in your browser, I recommend downloading it to your local disk instead (control-click on it on a Mac; right-click in Windows).
  • PDF
• Dissertation draft
• MASKS Annotations Software Downloads page
  • This is here in case you might be interested in playing around with VisageMap yourself. It might be useful to take a look at the guidelines page, which is partially a training manual.

Event Details

Date: Jan 12, 2009 (Monday)
Time: 1-3pm EST (10am-12pm PST)
Penn Location: Levine 307
Advisor: Jonathan Smith
Committee:

• Matt Blaze (chair), University of Pennsylvania
• Jianbo Shi, University of Pennsylvania
• Lawrence Saul, University of California, San Diego
• Paul Syverson (external), U.S. Naval Research Laboratory

Abstract

High-resolution digital cameras are becoming ever-larger parts of our daily lives, whether as part of closed-circuit surveillance systems or as part of portable digital devices that many of us carry around with us. Combining the broadening reach of these cameras with automatic face recognition technology creates a sensor network that is ripe for abuse: our every action could be recorded and tagged with our identities, the date, and our location as if we each had a investigator tasked only with keeping each of us under constant surveillance. Adding the continually falling cost of data storage to this mix, and we are left with a situation where the privacy abuses don't need to happen today: the stored imagery can be mined and re-mined forever, while the sophistication of automatic analysis continues to grow.

The MASKS project takes the first steps toward addressing this problem. If we would like to be able to de-identify faces before the images are shared with others, we cannot do so with ad hoc techniques applied identically to all faces. Since each face is unique, the method of disguising that face must be equally unique. In order to hide or reduce those critical identifying characteristics, we are delivering the following foundational contributions toward characterizing the nature of facial information:

• We have created a new pose-controlled, high-resolution database of facial images.
• The most prominent anatomical markers on each face have been marked for position and shape, establishing a new gold standard for facial segmentation.
• A parameterized model of the diversity of our subject population was built based on statistical analysis of the annotations. The model was validated by comparison with the performance of a standard set of artificial disguises.