Theory of Mind as Test-Time Mitigation against Adversarial Communication

• Post by: Bahareh Arghavani
• November 29, 2022
• Comments off

Presenter: Nancirose Piazza (PhD Student – SAIL Lab)
Time: Friday 03/03/2023, 2pm – 3pm ET
Recording: TBA

Abstract:

In Cooperative Multi-Agent Reinforcement Learning, there are training time defenses against adversarial communication. However, these types of defenses are expensive and require retraining if the underlying agent policies are retrained. We propose a test-time mitigation against adversarial communication using a mechanism designed from theory of mind concepts and provide empirical results of its performance preservation after readaption training in comparison to another message filtering defense against adversarial communication, namely the Variational Auto-Encoding Bayes.

Bio:

Nancirose Piazza is a Ph.D candidate at the University of New Haven, focusing on deception and deception mitigation in Multi-Agent Systems. She holds a Bachelor of Science in Mathematics from Sacred Heart University and a Masters in Data Science from the University of New Haven. Her research investigates the conditions of emergent deception in various systems and proposes mitigation and defenses through Machine Learning models and decision-making frameworks.