Final Report: Sparsity-Based Design for Robust Deep Learning-Topic C. iii (3)
Abstract
Deep neural networks [DNNs] yield state of the art performance in many fields, but are known to be vulnerable to small adversarial perturbations. This, together with their lack of interpretability, is a major impediment to their use in many DoD applications, as well as in safety-critical commercial applications such as vehicular autonomy. The overarching goal of this project is to investigate techniques for understanding and robustifying DNNs. The original proposal focused on the specific approach of imposing sparsity constraints to attenuate and eliminate adversarial perturbations, with the goal of obtaining interpretable designs with guaranteed resilience. This is in contrast to the state of art defenses against adversarial perturbations, which are based on black box training with adversarially perturbed examples.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 18, 2022
- Accession Number
- AD1198333
Entities
People
- Upamanyu Madhow
Organizations
- University of California, Santa Barbara