Privacy Preservation for Context Sensing on Smartphone

The proliferation of sensor-equipped smartphones has enabled an increasing number of context-aware applications that provide personalized services based on users' contexts. However, most of these applications aggressively collect users' sensing data without providing clear statements on the usage and disclosure strategies of such sensitive information, which raises severe privacy concerns and leads to some initial investigation on privacy preservation mechanisms design. While most prior studies have assumed static adversary models, we investigate the context dynamics and call attention to the existence of intelligent adversaries. In this paper, we identify the context privacy problem with consideration of the context dynamics and malicious adversaries with capabilities of adjusting their attacking strategies. Then, we formulate the interactive competition between users and adversaries as a competitive Markov decision process (MDP), in which the users attempt to preserve the context-based service quality and their context privacy in the long-term defense against the strategic adversaries with the opposite interests. In addition, we propose an efficient minimax learning algorithm to obtain the optimal policy of the users and prove that the algorithm quickly converges to the unique Nash equilibrium point. Our evaluations on real smartphone context traces of 94 users demonstrate that the proposed algorithm largely improves the convergence speed by three orders of magnitude compared with traditional algorithm and the optimal policy obtained by our minimax learning algorithm outperforms the baseline algorithms.