Online Photo Privacy - Phase 3:
Identifying Sensitive Content in Online Photos & Investigating Users' Sharing Preferences

Choosing which photos to share online is difficult. Although machine learning methods exist that can identify content in photos, we currently do not have a taxonomy that describes what content is considered sensitive or how sharing preferences for content differs across potential photo recipients. To fill this gap, we introduced a new sensitive content elicitation method which overcame the limitations of machine learning approaches, and collected photos that contain sensitive content from 116 participants. We also recorded their sharing preferences for these photos with 20 recipient groups. Next, we conducted a card sort study on the 181 unique pieces of sensitive content identified in this study to surface user-defined categories of sensitive content. Using data from these studies, we generated a taxonomy that identifies 28 categories of sensitive content. We also establish how sharing preferences for content differs across groups of potential photo recipients. This taxonomy can serve as a framework for understanding photo privacy, which can in turn inform new photo privacy protection mechanisms.

Primary contributions include:

• Introducing a new method to elicit sensitive content from participants. This novel method removes many of the barriers in collecting private content by providing participants with alternative ways to identify sensitive data that preserve their privacy.

• Organizing prior work from across disciplines, test it, and extend it. We collected a much larger data set (563 total including 181 unique pieces of sensitive content) with a larger sample size compared to any of the prior work.

• Providing a much more granular level of detail (e.g., a prior work states "environment" is sensitive, but our data allows us to see that the broader category of "environment" is composed of hospital, party/bar, bedroom, bathroom, messy room). Thus, our taxonomy is more practical and detailed for privacy researchers, ML researchers and practitioners.

​

I published and presented this paper at CHI 2020.

​

Research Method:
Controlled experiment; Online survey; Card sorting

Participants:

• 116 participants recruited on Amazon MTurk for photo elicitation and collecting sharing preference

• 14 participants recruited on campus for grouping the collected sensitive content via card sorting

Independent Variable:

• Sensitive content category

Dependent Variable:

• Likelihood to share with 20 recipient groups

Quantitative Analysis Method:

• Mixed-effect model