Using Privacy Loss to Guide Decisions During Distributed CSP Search

In cooperative problem solving, the communication necessary for solution search can also lead to privacy loss on the part of the agents involved. Such loss can be assessed either by directly tallying the number and importance of specific items of information revealed or by tracking reductions in the set of possible values associated with a particular item of information. In both cases information loss can occur either because of direct communication or by inferences that other agents make from one’s communications. The results of these inferences are stored in the “views” that agents have of other agents. In the present work on distributed constraint solving, such views are organized as extensions of normal CSP representations that model information about possible values in unknown CSPs of other agents. Here we show how this approach can be extended so that agents also maintain views of other agents’ views of themselves; the latter are called “mirror views”. Mirror views can in turn be used to monitor one’s own privacy loss, and can support strategies designed to reduce the loss of particular kinds of private information. Experiments with a simulated meeting scheduling system show that it is possible to reduce privacy loss with strategies based