Foundations of a CPS Resilience - January 2022

PI: Xenofon Koutsoukos

HARD PROBLEM(S) ADDRESSED

The goals of this project are to develop the principles and methods for designing and analyzing resilient CPS architectures that deliver required service in the face of compromised components. A fundamental challenge is to understand the basic tenets of CPS resilience and how they can be used in developing resilient architectures. The primary hard problem addressed is resilient architectures. In addition, the work addresses scalability and composability as well as metrics and evaluation. 

PUBLICATIONS

[1]    Bradley Potteiger, Feiyang Cai, Zhenkai Zhang, and Xenofon Koutsoukos, “Data Space Randomization for Securing Cyber-Physical Systems”, International Journal on Information Security. 2021.

[2]    Jiani Li, Feiyang Cai, and Xenofon Koutsoukos. “Byzantine Resilient Aggregation in Distributed Reinforcement Learning”, 18th International Conference on Distributed Computing and Artificial Intelligence (DCAI'21). Lecture Notes in Networks and Systems, vol 327, pp. 55-66, Springer, Cham.

[3]    Yi Li, Xenofon Koutsoukos, and Yevgeniy Vorobeychik. “Adversarial Gaussian Process Regression in Sensor Networks”, Game Theory and Machine Learning for Cyber Security, 149-159, 2021.

[4]    Waseem Abbas, Mudassir Shabbir, Yasin Yazıcıoğlu, and Xenofon Koutsoukos. "Edge Augmentation with Controllability Constraints in Directed Laplacian Networks," 60th IEEE Conference on Decision and Control, 2021.

KEY HIGHLIGHTS

This quarterly report presents two key highlight that demonstrate: (1) an algorithm for Byzantine resilient aggregation in distributed reinforcement learning and (2) and an approach for adversarial regression in sensor networks.

Highlight 1: Byzantine Resilient Aggregation in Distributed Reinforcement Learning

Recent distributed reinforcement learning techniques utilize networked agents to accelerate exploration and speed up learning. However, such techniques are not resilient in the presence of Byzantine agents which can disturb convergence. In this work, we present a Byzantine resilient aggregation rule for distributed reinforcement learning with networked agents that incorporates the idea of optimizing the objective function in designing the aggregation rules. We evaluate our approach using multiple reinforcement learning environments for both value-based and policy-based methods with homogeneous and heterogeneous agents. The results show that cooperation using the proposed approach exhibits better learning performance than the non-cooperative case and is resilient in the presence of an arbitrary number of Byzantine agents. Our results are presented in [1]. 

[1]    Jiani Li, Feiyang Cai, and Xenofon Koutsoukos. “Byzantine Resilient Aggregation in Distributed Reinforcement Learning”, 18th International Conference on Distributed Computing and Artificial Intelligence (DCAI'21). Lecture Notes in Networks and Systems, vol 327, pp. 55-66, Springer, Cham.

Highlight 2: Adversarial Gaussian Process Regression in Sensor Networks

Cyber-physical systems are fundamental to operations of many safety-critical systems, from power plants to autonomous cars. Such systems feature a control loop that maps sensor measurements to control decisions. In many applications, these decisions involve maintaining system state features, such as temperature and pressure, in a safe range, with anomaly detection employed to ensure that anomalous or malicious sensor measurements do not subvert system operation. Although anomaly detection has been studied in the literature, many existing approaches focus on the cases with passive adversaries. Our first contribution is a novel stealthy attack on systems featuring Gaussian Process regression (GPR) for anomaly detection—a popular choice for this task. Next, we pose the problem of robust GPR for anomaly detection as a Stackelberg game and present a novel algorithmic approach for solving it. Our experimental evaluation demonstrates both the vulnerability of baseline systems to attack, as well as the increased robustness offered by our approach. Our results are presented in [2]. 

[2]    Yi Li, Xenofon Koutsoukos, and Yevgeniy Vorobeychik. “Adversarial Gaussian Process Regression in Sensor Networks”, Game Theory and Machine Learning for Cyber Security, 149-159, 2021.

COMMUNITY ENGAGEMENTS

• Our research was presented in the following conferences: 60th IEEE Conference on Decision and Control (CDC 2021), 18th International Conference on Distributed Computing and Artificial Intelligence (DCAI'21).
• PI Xenofon Koutsoukos gave an invited talk on “Resilient Distributed Consensus, Optimization, and Learning in Networked Cyber-Physical Systems” at the CCI-MHI Cyber-Physical Systems Seminar, University of Southern California, October 27, 2021.