Privatizing Constraint Optimization, Thesis Abstract

My thesis proposes novel schemes for achieving privacy in constraint optimization and new ways of analyzing the privacy properties of constraint optimization algorithms. Thus, it aims to contribute to the fields of constraint processing and multiagent systems. Motivation: Many problem domains within constraint processing require algorithms that better accomodate the need for privacy. Two salient examples are described in the following paragraphs. Coordinating the schedules of muliple agents is a problem domain that naturally motivates the need for privacy preserving constraint optimization algorithms. In this domain, a number of agents wish to schedule events subject to individual constraints, which they wish to keep private. Their calendars may contain the intersection of multiple professional endeavors as well as events relevant to their personal lives. The agents involved want a solution that will optimize global social welfare, but are not willing for everyone to know their preferences. For example, scheduling constraints may expose attendance at support group meetings. Resource allocation problems provide another motivating example. Imagine several government, corporate and nonprofit groups that come together to help with some disaster relief effort. They all wish to contribute resources that will lead to the optimal end, but they are mutually distrustful in most of their other endeavors and do not want the other entities to know the details of their individual holdings and constraints. Background: Initially, all constraint processing algorithms were centralized; all constraints were communicated to and optimized by a central server. The central server thus knew everything about everyone. Concerns about privacy and efficiency led to the development of distributed constraint satisfaction and optimization algorithms (Yokoo et al. 1998; Modi et al. 2005). In these algorithms, each agent owns its own constraints and passes messages as necessary to other agents to collaboratively compute the solution. Since information is distributed among many agents, these algorithms were thought to provide more privacy. However, recent work has shown that in some cases, distributed algorithms lead to greater privacy loss than the traditional centralized solutions; everybody learning everything about