Data privacy-utility trade-off based on the Mutual Information Neural Estimator MINE

To address the difficulty of accurately characterizing mutual information with small samples or unknown distributions, the study proposes a privacy funnel based on the Mutual Information Neural Estimator (MINE), which estimates mutual information by training rather than by conventional direct computation. The method uses estimated mutual information to measure privacy and utility, and trains a neural network to optimize data utility under the constraint that the privacy leakage given by the estimator is below a threshold. Simulation results show that with limited sample sizes the estimate can still approximate mutual information well and optimize the privacy-utility trade-off.