Dynamic Information Disclosure for Deception

We analyze in this paper how a deceptive information provider can shape the shared information in order to control a decision maker's decisions. Data-driven engineering applications, e.g., machine learning and artificial intelligence, build on information. However, this implies that information (and correspondingly information providers) can have influential impact on the decisions made. Notably, the information providers can be deceptive such that they can benefit, while the decision makers suffer, from the strategically shaped information. We formulate (and provide an algorithm to compute) the optimal deceptive shaping policies in the multi-stage disclosure of, general, multi-dimensional Gauss-Markov information. To be able to deceive the decision maker, the information provider should anticipate the decision maker's reaction while facing a trade-off between deceiving at the current stage and the ability to deceive in the future stages. We show that optimal shaping policies are linear within the general class of Borel-measurable policies even though the information provider and the decision maker could be seeking to minimize quite different quadratic cost functions.