A Stochastic Sampling Approach to Privacy

Chuanghong Weng,Ehsan Nekouei

2024-05-22

Abstract:This paper proposes an optimal stochastic sampling approach to privacy, in which a sensor observes a process which is correlated to private information, and a sampler decides to keep or discard the sensor's observations. The kept samples are shared with an adversary who might attempt to infer the private process. The privacy leakages are captured with the mutual information between the private process and sampler's output. We cast the optimal sampling design as an optimization problem that (i) minimizes the reconstruction error of the observed process using the sampler's output, (ii) and reduces privacy leakages. We first show the optimal reconstruction is obtained by solving a one-step optimization problem at each time step. We derive the optimality equations of the sampler for a general processes via the dynamic decomposition method, and show the sampler controls adversary's belief about the private input. Also, we propose a simplified design for linear Gaussian processes by restricting the sampling policy to a special collection. We show that the optimal reconstruction of states, the belief state and the optimization objective can be analytically expressed based on a conditional mean and covariance matrix. We develop an numerical algorithm to optimize the sampling and reconstruction policies based on the implicit function theorem. Finally, we verify our design and show its capabilities in state reconstruction, privacy protection and data size reduction via simulations.

Systems and Control