Probabilistic Lipschitz Analysis of Neural Networks

Ravi Mangal, Kartik Sarangmath, Aditya V. Nori, Alessandro Orso

Abstract: We are interested in algorithmically proving the robustness of neural networks. Notions of robustness have been discussed in the literature; we are interested in probabilistic notions of robustness that assume it feasible to construct a statistical model of the process generating the inputs of a neural network. We find this a reasonable assumption given the rapid advances in algorithms for learning generative models of data. A neural network f is then defined to be probabilistically robust if, for a randomly generated pair of inputs, f is likely to demonstrate k-Lipschitzness, i.e., the distance between the outputs computed by f is upper-bounded by the $$k^{th}$$ multiple of the distance between the pair of inputs. We name this property, probabilistic Lipschitzness. We model generative models and neural networks, together, as programs in a simple, first-order, imperative, probabilistic programming language, $$pcat$$ . Inspired by a large body of existing literature, we define a denotational semantics for this language. Then we develop a sound local Lipschitzness analysis for $$cat$$ , a non-probabilistic sublanguage of $$pcat$$ . This analysis can compute an upper bound of the “Lipschitzness” of a neural network in a bounded region of the input set. We next present a provably correct algorithm, $$\mathtt{PROLIP}$$ , that analyzes the behavior of a neural network in a user-specified box-shaped input region and computes - (i) lower bounds on the probabilistic mass of such a region with respect to the generative model, (ii) upper bounds on the Lipschitz constant of the neural network in this region, with the help of the local Lipschitzness analysis. Finally, we present a sketch of a proof-search algorithm that uses $$\mathtt{PROLIP}$$ as a primitive for finding proofs of probabilistic Lipschitzness. We implement the $$\mathtt{PROLIP}$$ algorithm and empirically evaluate the computational complexity of $$\mathtt{PROLIP}$$ .