Presenter: Dr. Robert Bassett

Description: Adversarial manipulations show that many state-of-the-art neural network models suffer from a catastrophic lack of robustness. In this talk, we investigate how a capable adversary can exploit this weakness to disrupt autonomous acoustic sensors. We take the perspective of the adversary, formulating a partial differential equation (PDE) constrained optimization problem that constructs an acoustic signal which optimally disrupts a defender’s classifier. The primary novelty in our formulation is the PDE constraint, which connects the acoustic signal emitted by the adversary to the signal received by the classifier. By incorporating this PDE constraint, our formulation models an adversary which can make changes that impact the ambient acoustic environment but cannot directly manipulate inputs into the classifier. After discussing the problem formulation and our proposed methods for solving it, we conclude by summarizing the implications of our results on national security applications.

Bio: Robert L Bassett is an Assistant Professor of Operations Research and Applied
Mathematics at the Naval Postgraduate School. His research focuses on mathematical
programming applied to problems in statistics and signal processing, with an emphasis on
defense applications. Before joining NPS, Dr. Bassett worked as a research mathematician with the National Security Agency and the Institute for Defense Analyses, and as a system analyst with Sandia National Laboratories.

Hosted by: Applied Mathematics Department