Reduced Products of Abstract Domains for Fairness Certification of Neural Networks

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