Effective Certification of Monotone Deep Equilibrium ModelsDownload PDF

Sep 29, 2021 (edited Nov 15, 2021)ICLR 2022 Conference Withdrawn SubmissionReaders: Everyone
  • Keywords: Deep Equilibrium Models, Certified Robustness, Convex Relaxations
  • Abstract: Monotone Operator Equilibrium Models (monDEQs) represent a class of models that combine the powerful deep equilibrium paradigm with convergence guarantees. As monDEQs are inherently robust to adversarial perturbations, investigating new methods to certify their robustness is a promising research direction. Unfortunately, existing certification approaches are either imprecise or severely limited in their scalability. In this work, we propose the first scalable \emph{and} precise monDEQ verifier, based on two key ideas: (i) a novel convex relaxation which enables efficient inclusion checks, and (ii) non-trivial mathematical insights characterizing the fixpoint operations at the heart of monDEQs on sets rather than concrete inputs. An extensive evaluation of our verifier demonstrates that on the challenging $\ell_\infty$ perturbations it exceeds state-of-the-art performance in terms of speed (two orders of magnitude) and scalability (an order of magnitude) while yielding 25\% higher certified accuracies on the same networks.
  • One-sentence Summary: We obtain state-of-the-art results in robustness verification of monDEQs by proposing a novel convex relaxation which admits efficient fixed point computation.
10 Replies

Loading