Vikram Krishnamurthy, "Adversarial Sensing: Infer, Reveal, and Conceal"

Abstract: In sophisticated sensing and inference systems, such as cognitive radars or networks of interacting LLMs, observations can reveal far more than the physical state: actions and responses also expose hidden beliefs, strategies, and objectives. An adversary can exploit these observed actions and responses to infer the underlying belief. This talk presents a unifying perspective on sensing as an adversarial inference problem. We first consider how hidden beliefs can be reconstructed from observed actions using inverse filtering, and how information is fused across networks through word-of-mouth learning. Extending this viewpoint to networks of interacting large language models reveals a striking phenomenon: endogenous information exchange can create a persistent “glass ceiling” in influence. We then discuss how revealed preference theory and inverse reinforcement learning serve as principled tools for inferring utility from sensing actions. Finally, we reverse the perspective: how can a sensing system conceal its objectives from an adversarial observer while preserving operational effectiveness? We discuss utility masking and information-theoretic approaches that degrade identifiability while maintaining performance. Bio: Vikram is a professor in Electrical and Computer Engineering at Cornell University. His research interests are in statistical signal processing, inverse reinforcement learning, stochastic optimization and partially observed Markov decision processes. He is a Fellow of IEEE, served as distinguished lecturer for the IEEE Signal Processing Society and Editor in Chief of IEEE Journal Selected Topics in Signal. He was awarded an honorary doctorate from KTH, Sweden in 2013