CertiFair: A Framework for Certified Global Fairness of Neural Networks

We consider the problem of whether a Neural Network (NN) model satisfies global individual fairness. Individual Fairness (defined in (Dwork et al. 2012)) suggests that similar individuals with respect to certain task are be treated similarly by decision model. In this work, we have two main objectives. The first is construct verifier which checks fairness property holds for given NN classification or provides counterexample if it violated, i.e., fair all classified same, and unfair pair differently. To end, sound complete verifies properties ReLU classifiers using distance-based similarity metrics. second objective paper provide method training provably from (biased) data. propose loss can used during enforce outcomes individuals. then provable bounds on resulting NN. run experiments commonly datasets publicly available show improved 96 % without significant drop test accuracy.