Privacy-Performance Trade-Offs in Sequential Decision-Making
Abstract
As the frontiers of autonomy are pushed into new areas and applications, the increased coupling between the cyber and physical domains offers fundamentally new capabilities. For example, unmanned vehicles may have the capability to track adversaries and enforce keepout zones with information harvested from remote sensors. However, challenges emerge when decisions must be made using information that is sensitive. For example, a drone that always avoids some area may be inadvertently revealing that it has detected adversaries in that area. This and other scenarios have the unusual property that sensitive information in the cyber domain is revealed by decision-making in the physical world. Of course, the disclosure of sensitive data must be avoided, but, without some form of data privacy, this may be at odds with taking effective actions based on sensitive data.The objective of this proposal is to develop theory and algorithms for sequential decision-making in Markov decision processes (MDPs) with both privacy and system performance in mind. In particular, the resulting algorithms will account for the privacy of the information used in decision-making and the performance of decision-making when privacy-preserving strategies are implemented. In analyzing "performance," we will quantify the effectiveness of decision-making in accomplishing its goals and the satisfaction of high-level specifications while doing so.The privacy framework we use is differential privacy, and we will develop new privacy mechanisms for autonomous assets to avoid compromising their objectives (what assets are trying to do), tactics (how they are trying to do it), and knowledge (what information assets use to drive decision-making). Mathematically, each of these can be encoded as a property of an MDP. We will (i) develop new differential privacy mechanisms to protect all of these objects while respecting the structure of data, e.g., a probability distribution over actions will remain a probability distribution after it is privatized, (ii) quantify privacy/performance trade-offs through quantifying the changes in a value function that result from the uncertainty introduced by privacy, and (iii) quantify trade-offs between privacy and specification satisfaction through bounding the probability of satisfaction of temporal logic specifications as a function of privacyprotections.The proposed effort has the unique opportunity to establish theoretical and algorithmic connections between the conventionally disparate disciplines of data privacy and autonomy. The need for these developments is amplified when the main characteristics of the operations of the Navy are factored in, e.g, the presence of adversaries, the need to act on sensitive information while concealing one s knowledge of it, and the criticality of performance and constraint satisfaction. This effort offers new algorithmic approaches to several problems that are known to be challenging and poses novel problems in which it will enable new capabilities for autonomous Naval assets to act while formally safeguarding the information that drives those actions.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2021
- Source ID
- N000142112502
Entities
People
- Matthew Hale
Organizations
- Office of Naval Research
- United States Navy
- University of Florida