Abstract Hidden Markov Models: A Monadic Account of Quantitative Information Flow

Hidden Markov Models, HMM's, are mathematical models of Markov processes whose state is hidden but from which information can leak via channels. They are typically represented as 3-way joint probability distributions. We use HMM's as denotations of probabilistic hidden-state sequential programs, after recasting them as “abstract” HMM's, i.e. computations in the Giry monad D, and equipping them with a partial order of increasing security. However to encode the monadic type with hiding over state X we use DX→D
²
X rather than the conventional X→DX. We illustrate this construction with a very small Haskell prototype. We then present uncertainty measures as a generalisation of the extant diversity of probabilistic entropies, and we propose characteristic analytic properties for them. Based on that, we give a “backwards”, uncertainty-transformer semantics for HMM's, dual to the “forwards” abstract HMM's. Finally, we discuss the Dalenius desideratum for statistical databases as an issue in semantic compositionality, and propose a means for taking it into account.