Provable Fairness for Neural Network Models Using Formal Verification

Machine learning models, extensively deployed in critical decision-making roles, necessitate verification for potential gender or racial biases from training data. Traditional fairness strategies focus on curating training data and subsequent statistical evaluation of model fairness. We, however, propose techniques that offer formal proof of fairness, leveraging recent advancements in neural network model verification. Our methods offer robust guarantees and uniquely, they eliminate the need for explicit training or evaluation data, often proprietary, to analyze a trained model. Experimental results from the well-known ADULTS dataset demonstrate that appropriate training can decrease unfairness by an average of 65.4% with less than 1% cost in AUC score.