Presentation by Jun Hao Xu The Price of Free Privacy leakage in personalized mobile in-app ads Authors: Wei Meng, Ren Ding, Simon P. Chung, Steven Han, and Wenke Lee Presentation by Jun Hao Xu
Motivation Mobile advertisements are crucial to the mobile app ecosystem Win-win situation for both users and publishers. But what are the hidden costs of mobile ads? Mobile devices have become very intimate devices for us. The risk of privacy leakage is a serious issue. http://www.coganlawoffice.com/wp-content/uploads/2014/06/cel-privacy-e1405699296949-1.jpg
Background Ecosystem of Mobile Advertising Publishers, Advertisers and Ad Networks Targeting in Mobile Advertising Topic Targeting Interest Targeting Demographic Targeting Vulnerabilities Mobile vs Web https://1.bp.blogspot.com/--iy0JM49ZDc/VyCydJX6_YI/AAAAAAAAETc/QyKdN_5HnPcjFls1jiFiIBZKPE7mi3A3wCLcB/s1600/Best%2BAd%2BNetworks%2BFor%2BPublishers%2B%2526%2BAdvertisers%2B%255BEARN%2BMONEY%255D.png
Goals How much information do ad networks know about users and how much of this is used? Could personalized mobile in-app ads be served as a channel of private user information leakage?
Challenges Synthesized user profiles Removing external factors Using synthesized user profiles leads to validity problems. Real users were used to solve this. Removing external factors Geo-location is a important factor Used VPN to hide location Different apps may have different ad control Developed own app http://www.performics.com/wp-content/uploads/2015/04/mobile-apps.jpg
Methodology Focus on Android phone and Google’s Admob for data collection. Over 200 users were used for this study. Install blank app dedicated for collecting mobile ads. Blank app sends 100 ad requests. Landing URL’s were used to represent the ads. Ad’s were categorised into 1 of 24 categories.
Results Interests Mobile Ads are highly personalized based on user’s interests. For 41% of users, more than 57% of their ad impressions match their real interests. Google actively personalizing large fraction of its ads.
Results Demographic Gender Parental Status Income Religion Google’s personalization algorithm Religion Result of other categories. Age, Education, Ethnicity Least targeted.
Privacy Leakage The ability of observing personalized ads opens up a new avenue of attack on private personal information. Demographics Learning from Personalized Mobile Ads Machine learning algorithms to build user profiles based on ads seen. Evaluated accuracy of user profiles Dummy algorithm, which randomly guesses a user’s profile, used as a baseline classifier for comparison.
Results – Privacy Leakage Classification algorithms performed better than dummy algorithms in all categories. Potential risks involve selling information or exploiting users.
Countermeasures Ad networks changing their policies Add noise or randomness into the personalized results. Providing coarser grained targeting options Can’t expect all ad networks to follow this. Find a way to isolate app and ad process.
Criticism - Positives Addresses problems of other research. Introduces problem well and expands on it.
Criticism – Other Ad Networks AdMob only used. Other Ad Networks? InMobi – Advertisements by basing the targeting on the users’ existing and previous applications instead of traditional metrics such as demographics or geography.
Criticism - Android Android only used in this experiment. IOS or Windows phones. http://swiftmobilespy.co.za/wp-content/uploads/2015/03/Apple-logo-icon-Aluminum-300x300-300x300.png
Criticism – Geo Location Geo Location isn’t considered. Geo Location considered a significant factor. Recruit more users. http://cdn.whatismyipaddress.com/images-v4/geolocation.png
Criticism – Small Sample Size Only 217 Users recruited. Introduces demographic bias Larger data set means more reliable data.
Criticism – Methodology Assumes that 100 requests will not alter the user profile significantly. More Ad requests to confirm the validity of the results.
Questions?