Reduced Products of Abstract Domains for Fairness Certification of Neural Networks

Denis Mazzucato, Caterina Urban

Abstract

We present Libra, an open-source abstract interpretation-based static analyzer for certifying fairness of ReLU neural network classifiers for tabular data. Libra combines a sound forward pre-analysis with an exact backward analysis that leverages the polyhedra abstract domain to provide definite fairness guarantees when possible, and to otherwise quantify and describe the biased input space regions. The analysis is configurable in terms of scalability and precision. We equipped Libra with new abstract domains to use in the pre-analysis, including a generic reduced product domain construction, as well as search heuristics to find the best analysis configuration. We additionally set up the backward analysis to allow further parallelization. Our experimental evaluation demonstrates the effectiveness of the approach on neural networks trained on a popular dataset in the fairness literature.