Towards Robust Moving Target Defense: A Game Theoretic and Learning Approach

Personnel

• Principal Investigator: Zizhan Zheng
• Postdoctoral Researcher: Wen Shen
• Graduate Students: Henger Li, Xiaolin Sun
• Undergraduate Students: Dung Ngo (Fall 2021), Emma LeBouef (2020-2021), Tom Roginsky (2019-2020), Harrison Pratt (2019-2020), Reid Backman (2018-2019)

Goals

A fundamental obstacle to achieving effective defense is information asymmetry, where, under the traditional static and passive defense schemes, the attacker has essentially limitless time to observe and learn about the defender, while the defender knows very little about the attacker. A promising approach to reverse the information asymmetry is Moving Target Defense (MTD), where the defender dynamically updates system configurations to impede the attacker’s learning process. While MTD has been successfully applied to various domains, existing solutions typically assume an attacker with fixed capabilities and behavioral patterns that are known to the defender. The overarching goal of this project is to develop the foundations for the design and analysis of robust MTD mechanisms that can provide a guaranteed level of protection in the face of unknown and adaptive attacks. The proposed research contributes to the emerging field of the science of security via a cross-disciplinary approach that combines techniques from cybersecurity, game theory, and machine learning.

Tasks

Developing robust MTD faces three major challenges induced by (1) the coupling of system dynamics and incentives; (2) the hidden behavior of stealthy attacks; (3) the necessity of coordinating multiple defenders in large systems. To tackle these challenges, the investigator will focus on three interrelated thrust areas. In the first thrust, a dynamic two-timescale MTD game that captures a variety of attack patterns and feedback structures is designed and techniques for handling games with large state spaces are investigated. In the second thrust, reinforcement learning-based MTD policies for thwarting unknown attacks are studied. The focus is on developing approximately optimal solutions with low complexity that can effectively exploit the delayed and noisy feedback during the game. In the third thrust, the MTD game and learning framework are extended to incorporate multiple attackers and defenders, and information sharing and mediation schemes for enabling coordinated MTD are investigated.

Publications

• Learning to Attack Federated Learning: A Model-based Reinforcement Learning Attack Framework  [code]
  Henger Li*, Xiaolin Sun*, and Zizhan Zheng (*Co-primary authors)
  Conference on Neural Information Processing Systems (NeurIPS), Dec. 2022.
• Robust Moving Target Defense against Unknown Attacks: A Meta-Reinforcement Learning Approach  [code]
  Henger Li and Zizhan Zheng
  Conference on Decision and Game Theory for Security (GameSec), Oct. 2022.
• Coordinated Attacks Against Federated Learning: A Multi-Agent Reinforcement Learning Approach
  Wen Shen, Henger Li, and Zizhan Zheng
  ICLR 2021 Workshop on Security and Safety in Machine Learning Systems (SecML), selected for travel award, May 2021.
• Learning to Attack Distributionally Robust Federated Learning
  Wen Shen, Henger Li, and Zizhan Zheng
  NeurIPS-20 Workshop on Scalability, Privacy, and Security in Federated Learning (SpicyFL), selected for oral presentation, Dec 2020.
• Spatial-Temporal Moving Target Defense: A Markov Stackelberg Game Model  [code]
  Henger Li, Wen Shen, and Zizhan Zheng
  International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), May 2020.
• Defending Against Stealthy Attacks on Multiple Nodes with Limited Resources: A Game-Theoretic Analysis
  Ming Zhang, Zizhan Zheng, and Ness B. Shroff; IEEE Transactions on Control of Network Systems (TCNS), 2020.
• Optimal Timing of Moving Target Defense: A Stackelberg Game Model.
  Henger Li and Zizhan Zheng;  International Conference for Military Communications (MILCOM), Nov. 2019.

MTD Testbed

We are building an MTD testbed as a Virtual Private Cloud (VPC) on Amazon Web Services (AWS). A tutorial for setting up a basic version of the testbed can found here (contributed by Harrison Pratt).

Support

The project is funded by National Science Foundation (NSF) grant award CNS-1816495.

Disclaimer: Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.