Relative Perfect Secrecy: Universally Optimal Strategies and Channel Design

Perfect secrecy describes cases where an adversary cannot learn anything about the secret beyond its prior distribution. A classical result by Shannon shows that a necessary condition for perfect secrecy is that the adversary should not be able to eliminate any of the possible secrets. In this paper we answer the following fundamental question: What is the lowest leakage of information that can be achieved when some of the secrets have to be eliminated? We address this question by deriving the minimum leakage in closed-form, and explicitly providing "universally optimal" randomized strategies, in the sense that they guarantee the minimum leakage irrespective of the measure of entropy used to quantify the leakage. We then introduce a generalization of Rényi family of asymmetric measures of leakage which generalizes the g-leakage and show that a slight modification of our strategies are optimal with respect to an important class of such measures. Subsequently, we show that our schemes constitute the Nash Equilibria of closely related two-person zero sum games. This game perspective provides implicit solutions for a wider set of structural constraints and asymmetric entropies. Finally we demonstrate how this work can also be seen as designing a universally optimal channel given a specified prior.