Sample Selection for Fair and Robust Training

Fairness and robustness must be addressed together: The paper demonstrates that focusing on only one aspect—either eliminating bias or handling corrupted data—can actually harm performance on the other dimension, highlighting that trustworthy AI requires a balanced approach to both concerns simultaneously.

Sample selection is a practical solution: Rather than redesigning entire training algorithms, the proposed approach achieves fair and robust models by simply modifying which data samples are selected during training batches, making it practical for organizations to implement without overhauling existing systems.

The method works without requiring clean data: Unlike some competing approaches, this algorithm doesn't require access to additional high-quality datasets to achieve fairness and robustness, making it more feasible for real-world applications where obtaining perfectly clean data is often impossible.

Performance improvements are measurable: Experiments on both synthetic and real benchmark datasets show the algorithm achieves fairness and robustness outcomes that are "better than or comparable to state-of-the-art" methods, suggesting meaningful practical benefits for deployed AI systems.

The underlying problem is computationally hard but solvable: While the core optimization challenge is mathematically proven to be extremely difficult (NP-hard), the paper's greedy algorithm provides an efficient, practical solution that works well in practice despite the theoretical complexity.