Arms Verification with Information Barriers: Constraining Bayesian Networks For Confidence Building

Add to your list(s) Download to your calendar using vCal

• Michael Huth (Imperial College London)
• Thursday 30 March 2017, 11:00-12:00
• Computer Science, The Sloman Lounge (UG).

If you have a question about this talk, please contact Andreea Radu.

Arms control verification is a process that may be employed to build confidence between two parties that are subject to a nuclear arms control treaty. It builds confidence that both parties are undertaking agreed treaty obligations. We consider an arms inspection at a facility of a nuclear-weapons nation conducted by a nation or NGO . The aim of the inspection is to verify that a closed box within the facility contains a weapon (perhaps on its way to being destroyed) without compromising design or composition secrets of the weapon to the inspecting party. The tension caused by the inspecting party needing to take measurements of such a closed box and the inspected party not wishing to give away any weapon secrets can be mitigated by use of information barriers to measurements taken of the box. A key issue is whether both parties have sufficient confidence in this inspection process, including the use of information barriers, so that they can agree on whether or not a weapon is present in that box.

Bayesian Networks (BN) are probabilistic models that can express degrees of belief (subjective probabilities) and even trust (as bias in the processing of imperfect information). But using BNs to build confidence seems challenging: there is insufficient data available for learning a BN here, and decision support based on marginal probabilities may be sensitive to small changes in the BN.

We therefore developed a formalism called Constrained Bayesian Networks that allows us to refine Bayesian Networks to increase the confidence one can have in their probabilistic results. Technically, probability tables of nodes now contain symbolic arithmetic expressions and the network is enriched with a set of logical constraints. We develop algorithms for building confidence into such models, for example by computing optimal marginal probabilities, by determining worst-case sensitivity measures at critical nodes, and by understanding whether different constrained BNs render the same decision support.

Acknowledgments: This work was made possible with the generous support of AWE . The case study reported in this work is the result of a collaboration with Edward Day, Neil Evans, Sam Haworth, Tom Plant, and Catherine Roberts from AWE .

This talk is part of the Computer Security Seminars series.

This talk is included in these lists:

• Computer Science Departmental Series
• Computer Science Distinguished Seminars
• Computer Science, The Sloman Lounge (UG)
• Computer Security Seminars
• computer sience

Note that ex-directory lists are not shown.