FR-Train: A Mutual Information-Based Approach to Fair and Robust Training

FR-Train addresses a critical gap in trustworthy AI by tackling fairness and robustness simultaneously. Most existing fairness techniques treat poisoned (deliberately corrupted) data as just another form of bias, which causes significant performance problems. This paper's approach prevents the severe accuracy and fairness degradation that occurs when these two challenges are handled separately.

The method uses a "discriminator" component to identify and reduce the influence of poisoned data, relying on a clean validation dataset. By distinguishing between legitimate bias in data and malicious poisoning, FR-Train maintains fairness protections while defending against attacks—showing "almost no decrease in fairness and accuracy" even when data is poisoned.

Clean validation sets are practical to obtain through crowdsourcing, according to the authors' demonstration. This finding is important because it suggests organizations don't need perfect data to implement this fairness-and-robustness approach; they can use crowdsourced validation data to help identify problematic training examples.

The paper introduces new benchmark datasets for testing fair and robust AI systems. These resources can help the broader AI community evaluate whether their models maintain both fairness and accuracy under realistic conditions where data may be biased or deliberately corrupted.