Common Randomness Principles of Secrecy

This dissertation concerns the secure processing of distributed data by multi- ple terminals, using interactive public communication among themselves, in order to accomplish a given computational task. In the setting of a probabilistic multitermi- nal source model in which several terminals observe correlated random signals, we analyze secure distributed data processing protocols that harness the correlation in the data. The specific tasks considered are: computing functions of the data under secrecy requirements; generating secretly shared bits with minimal rate of public communication; and securely sharing bits in presence of a querying eavesdropper. In studying these various secure distributed processing tasks, we adopt a unified approach that entails examining the form of underlying common randomness (CR) that is generated at the terminals during distributed processing. We make the case that the exact form of established CR is linked inherently to the data processing task at hand, and its characterization can lead to a structural understanding of the associated algorithms. An identification of the underlying CR and its decomposi- tion into independent components, each with a different operational significance, is a recurring fundamental theme at the heart of all the proofs in this dissertation. In addition to leading to new theoretical insights, it brings out equivalences between seemingly unrelated problems. Another distinguishing feature of this work is that it considers interactive communication protocols. In fact, understanding the structure of such interactive communication is a key step in proving our results. We make the following contributions. First, we propose a new information theoretic formulation to study secure distributed computing using public communi- cation. The parties observing distributed data are trusted but an eavesdropper has access to the public communication network. We examine distributed communica- tion protocols that allow the trusted parties to accomplish their required computa- tion tasks while giving away negligible information about a specified portion of the data to an eavesdropper with access to the communication. Our theoretical results provide necessary and sufficient conditions that characterize the feasibility of vari- ous secure computing tasks; in many cases of practical importance, these conditions take a simple form and can be verified easily. When secure computing is feasible, we propose new algorithms in special cases. Next, we revisit the problem of generating shared secret keys (SKs). We investigate minimum communication requirements for generating information theo- retically secure SKs of maximum rates from correlated observations using interactive public communication. In particular, our approach allows us to examine the role of interaction in such communication. On the one hand, we find that interaction is not needed when the observed correlated bits are symmetrically correlated and therefore, in this case, simple noninteractive protocols are the most efficient means of generating optimum rate SKs. On the other hand, we illustrate that interactive pro- tocols can require a strictly lower rate of overall communication than noninteractive protocols. Finally, we consider the task of ensuring security against an eavesdropper who makes queries about a portion of the distributed data that the terminals share by communicating over a public network. We introduce an alternative notion of secrecy which requires rendering the task of a querying eavesdropper as onerous as possible. Our main contribution in this part is the development of a new technique for proving converse results for secrecy problems involving CR with interactive communication, which is employed then to obtain an upper bound for the maximum number of queries that can be inflicted on the eavesdropper for any CR and corresponding communication. Surprisingly, there is an equivalence between this notion of secrecy and that of information theoretic security, which leads to new theoretical results for SK generation; for instance, we prove a strong converse for the SK capacity. We conclude by hypothesizing the basic principles of secrecy generation that emerge from the results developed in this dissertation