Bayesian Leakage Analysis

We introduce a framework based on Bayesian statistical inference for analyzing leakage in cryptography and its vulnerability to inference attacks. Our framework naturally integrates auxiliary information, defines a notion of adversarial advantage, and provides information-theoretic measures that capture the security of leakage patterns against both full and functional recovery attacks. We present two main theorems that bound the advantage of powerful inference techniques: the maximum a posteriori (MAP), the maximum likelihood estimate (MLE) and the MAP test. Specifically, we show that the advantage of these methods is exponentially bounded by new entropy measures that capture the susceptibility of leakage patterns to inference. To demonstrate the applicability of our framework, we design and implement an automated leakage attack engine, Bayle, which leverages a novel inference algorithm that efficiently computes MAP estimates for a large class of i.i.d. leakage models. These models include query equality leakage, the combination of query equality and volume leakage, and leakage patterns arising from naive conjunctions.