Abstract:The global Lipschitz constant of a neural network governs both adversarial robustness and generalization. Conventional approaches to ``certified training" typically follow a train-then-verify paradigm: they train a network and then attempt to bound its Lipschitz constant. Because the efficient ``trivial bound" (the product of the layerwise Lipschitz constants) is exponentially loose for arbitrary networks, these approaches must rely on computationally expensive techniques such as semidefinite programming, mixed-integer programming, or branch-and-bound. We propose a different paradigm: rather than designing complex verifiers for arbitrary networks, we design networks to be verifiable by the fast trivial bound. We show that directly penalizing the trivial bound during training forces it to become tight, thereby effectively regularizing the true Lipschitz constant. To achieve this, we identify three structural obstructions to a tight trivial bound (dead neurons, bias terms, and ill-conditioned weights) and introduce architectural mitigations, including a novel notion of norm-saturating polyactivations and bias-free sinusoidal layers. Our approach avoids the runtime complexity of advanced verification while achieving strong results: we train robust networks on MNIST with Lipschitz bounds that are small (orders of magnitude lower than comparable works) and tight (within 10% of the ground truth). The experimental results validate the theoretical guarantees, support the proposed mechanisms, and extend empirically to diverse activations and non-Euclidean norms.
Abstract:The Lipschitz constant plays a crucial role in certifying the robustness of neural networks to input perturbations and adversarial attacks, as well as the stability and safety of systems with neural network controllers. Therefore, estimation of tight bounds on the Lipschitz constant of neural networks is a well-studied topic. However, typical approaches involve solving a large matrix verification problem, the computational cost of which grows significantly for deeper networks. In this letter, we provide a compositional approach to estimate Lipschitz constants for deep feedforward neural networks by obtaining an exact decomposition of the large matrix verification problem into smaller sub-problems. We further obtain a closed-form solution that applies to most common neural network activation functions, which will enable rapid robustness and stability certificates for neural networks deployed in online control settings. Finally, we demonstrate through numerical experiments that our approach provides a steep reduction in computation time while yielding Lipschitz bounds that are very close to those achieved by state-of-the-art approaches.




Abstract:Consider an unknown nonlinear dynamical system that is known to be dissipative. The objective of this paper is to learn a neural dynamical model that approximates this system, while preserving the dissipativity property in the model. In general, imposing dissipativity constraints during neural network training is a hard problem for which no known techniques exist. In this work, we address the problem of learning a dissipative neural dynamical system model in two stages. First, we learn an unconstrained neural dynamical model that closely approximates the system dynamics. Next, we derive sufficient conditions to perturb the weights of the neural dynamical model to ensure dissipativity, followed by perturbation of the biases to retain the fit of the model to the trajectories of the nonlinear system. We show that these two perturbation problems can be solved independently to obtain a neural dynamical model that is guaranteed to be dissipative while closely approximating the nonlinear system.