MAPS: Multi-Layer Adaptive and Proactive Strategic Cyber Defense
Abstract
In Year 1, we studied problems in strategic communication between two human agents and formulated the misalignment in the information available to them as a game. We also carried out research studying signaling and privacy games in cyber-physical systems in adversarial settings. During Year 2, we investigated attacks on in-vehicle networks, modeling of distributed denial-of-service attacks, and demonstrated vulnerabilities in existing machine learning algorithms and assessed their robustness. We also continued our work on strategic information transmission in game-theoretic settings that involved human agents and sensor scheduling problems in cyber-physical systems. A significant direction of research during this year was the study of robustness of deep neural networks to adversarial inputs. We demonstrated that an adversary could subtly manipulate a video in a way that a human observer would perceive the content of the original video, but the network would return the adversaryÕs desired outputs. We also carried out experiments on image classifiers, and showed that the accuracy of state-of-the-art models on adversarial color-shifted images is at the level of random classification. During Year 3, we built on and extended our efforts from previous years. We also validated the methods developed in our research on real-world data and compared their performance against state-of-the-art techniques and algorithms. We integrated techniques from control, game theory, optimization, and formal methods in order to provide probabilistic guarantees on the performance of cyber-physical systems in adversarial environments. We also studied the integration of human-feedback and potential-based methods in reinforcement learning order to accelerate the learning process. These projects have opened new directions for research in the development of adaptive strategies in order for cyber-physical systems to meet high-level objectives in the presence of an intelligent adversary. We also continued our existing research on the use of adversarial examples to train neural networks, so that they can be robust to adversarial inputs at test time. We also continued our work on automobile CAN bus security analysis, where we presented formal models for the study of cloaking attacks in automobiles. Finally, we studied a game theoretic formulation to detect advanced threats that can mimic benign system activities in a distributed and stochastic manner in order to evade detection.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 06, 2020
- Source ID
- W911NF1610485
Entities
People
- Radha Poovendran
Organizations
- Army Contracting Command
- United States Army
- University of Washington