Monitoring Algorithmic Fairness

Abstract Machine-learned systems are in widespread use for making decisions about humans, and it is important that they fair , i.e., not biased against individuals based on sensitive attributes. We present runtime verification of algorithmic fairness whose models unknown, but assumed to have a Markov chain structure. introduce specification language can model many common properties, such as demographic parity, equal opportunity, social burden. build monitors observe long sequence events generated by given system, output, after each observation, quantitative estimate how or the system was run until point time. The proven be correct modulo variable error bound confidence level, where gets tighter observed longer. Our two types, use, respectively, frequentist Bayesian statistical inference techniques. While compute estimates objectively with respect ground truth, subject prior belief system’s model. Using prototype implementation, we show monitor if bank giving loans applicants from different backgrounds, college admitting students while maintaining reasonable financial burden society. Although exhibit theoretical complexities certain cases, our experiments, both took less than millisecond update their verdicts observation.