Distributed and Privacy-Preserving Learning in Multi-Agent Systems
Abstract
Motivated by recent e -orts in the utilization of large data sets in the coordination of multiple intelligent robots, this proposaleeks to develop new control theoretic tools and algorithms that can exploit them in a principled manner. Existing algorithms generally leave individual systems unprotected against powerful adversaries who can have access to data, the output of algorithms,and internal network structure. While there has been a great e -ort to make multi-agent and control systems resilient against different attaks of sub-classes of adversaries, no much work exists to make individual these assets oblivious to a data breach in the system. In addition, safe data-handling has to be juxtaposed with the apparent contradiction of designing both useful and robust learningalgorithms that can be effectively used in fast decision-making scenarios for multiple agents.The proposed research will establish fundamental theory and algorithms that will help tune a desired balance between the preservation of data integrity and the effectiveness of high-confidence, data-based decision-making algorithms for autonomous systems. To do this, we will formulate estimation and decision-making problems for multiple agents in a distributional setting. On theone hand, this will allow us to ?nd novel conditions to naturally tune the differential privacy of assets or agents participating in an algorithmic process. On the other hand, the same framework will serve to develop new robust algorithms that can be used for the assimilation of data in time-varyingscenarios with performance guarantees. As a byproduct, a new class of Monte-Carlo estimation and distributed learning algorithms will be designed.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 15, 2019
- Source ID
- N000141912471
Entities
People
- Sonia Martinez Diaz
Organizations
- Office of Naval Research
- United States Navy
- University of California, San Diego