Privacy and Discrimination Threats in Social Network Ad Targeting and Delivery

Abstract The enormous financial success of online advertising platforms is partially due to the precise targeting and delivery features they offer. Recently, such platforms have been criticized for allowing advertisers to discriminate against users belonging to sensitive groups, i.e., to exclude users belonging to a certain race or gender from receiving their ads.  In this talk, I discuss two threads of work where we develop measurement methodologies in order understand the extent of discrimination in ads on online platforms.

First, I examine ad targeting — the advertisers' choices about who they wish to bid on.  Most platforms now allow advertisers to target users directly by uploading their personally identifiable information (PII), but it remains unclear which sources of PII ad platforms use for targeting.  Focusing on Facebook, I first develop a methodology that uses Facebook’s aggregate statistics in order to determine whether a given piece of PII matches an active account.  I then demonstrate that despite all privacy settings being enabled, phone numbers and email addresses added for security purposes (e.g.,  two-factor authentication), those provided to the Facebook Messenger app, and those included in friends’ uploaded contact databases are all used by Facebook to allow advertisers to target individual users.

Second, I examine ad delivery — the platform's choices about who should see an ad.  I first develop a measurement methodology using Facebook’s advertiser interface that can measure the influence of Facebook’s choices about how to deliver ads.  I then demonstrate that ad delivery can be significantly skewed on Facebook, due to the platform's own predictions about the "relevance" of ads to different groups of users. I show significant skew in delivery along gender and racial lines for "real" ads for employment and housing opportunities despite neutral targeting parameters.  These findings demonstrate previously unknown mechanisms that can lead to potentially discriminatory ad delivery, even when advertisers set their targeting parameters to be highly inclusive.