Location questions are generated based on a user's recent location traces tracked by the application.
For naive adversarial users, battery and location questions had the highest confidence ratings.
For strong adversaries, location questions had the highest confidence rating (i.e., strong adversarial users were more confident when guessing location questions compared to other six question types) followed by activity questions.
Please note that this algorithm only runs whenever the application needs to generate location questions for a user.
For example, the longest time was taken for location questions with mean 28.56 s (median: 23 s) for legitimate users.
Moreover, naive adversarial users had higher confidence ratings for location questions, which may stem from the fact that they knew that the other participants were from the same campus/locality.
For example, for (n=3) and when modeled against strong adversary, the average accuracy rates are 96.0%% for call questions, 94.1 % for SMS questions, 96.3 % for location questions, 97.1 % for app questions, 96.8 % for music questions, 98.0 % for activity questions and 95.9 % for battery questions.
On the other hand, when modeled against strong adversary and (n=3), the average accuracy rates are slightly lower 93.1 % for call questions, 86.8 % for SMS questions, 85.8 % for location questions, 93.4 % for app questions, 85.9 % for music questions, 86.8 % for activity questions and 93.6 % for battery questions.
For location questions, as users may not place the marker on exactly the same location coordinates estimated and identified by the system, there is an error tolerance (e.g., 75 m great circle distance) in our system, which is calculated based on the Haversine distance formula [34] between the selected coordinates and the estimated location.
