The experimental nature of our system means that we neither work with any commercial advertising companies nor generate any revenue by displaying ads. Instead, we create mock-up ads using product listings from amazon.com
. As far as profiling and targeting goes, we are experimenting with two simple heuristics: search-based targeting and product-based targeting. First looks at user keyword searches and flags them as product-relates or not by looking up individual terms in a dictionary built from words that appear more than a 100 times in titles of ca. 80 million products randomly retrieved from amazon.com
. Second identifies the specific product the user is browsing by checking if the currently displayed page is a shopping website and by applying a scraper customized for this particular shopping website. These heuristics are only used with a number of whitelisted websites (which includes major search engines and most popular online retailers).
Privad client uses identified product titles and product-related keywords to request ads from the backend infrastructure in an anonymous, privacy-preserving manner (see http://adresearch.mpi-sws.org/overview.html
for details). Backend servers use these ad requests to search for matching products on amazon.com
and build textual ads from the resulting products. The client then stores these mock-up Privad ads on the user computer and displays them in Google adboxes instead of Adsense ads.
Another experimental feature in Privad addon is differentially private data aggregation (http://en.wikipedia.org/wiki/Differential_privacy)
. What it means in practical terms is that Privad client locally collects ads-related statistical data and allows for numerical queries over these data with some amount of noise added to the query results. This mechanism makes it possible to obtain deep insight into ads performance, which goes beyond simple click-through and conversion-rates, without forfeiting user privacy.