The Effect of Location Granularity on Semantic Location Inferences

Abstract

Privacy issues related to location-based services have been extensively studied in recent years, but the focus has mainly been on providing what is commonly referred to as k-anonymity. The underlying assumptions in this approach, however, are often flawed, since the user's identity is often already known to the service provider. This paper studies the effect of location obfuscation on semantic location attacks, i.e. attacks inferring which places the user has visited, against an adversary with a prior distribution over the possible semantic locations. The prior is estimated using Foursquare venue data from five U. S. cities of varying sizes. Our results indicate that location data must be very low resolution to guarantee that an adversary can never infer a user's location with significant confidence, but the resolution can be increased by an order of magnitude if a low expected probability of a successful inference suffices.