RobustFair: Adversarial Evaluation through Fairness Confusion Directed Gradient Search

In one sentence. Deep neural networks are vulnerable to two kinds of adversarial perturbation: false ones that reduce accuracy and biased ones that make predictions unfair for similar inputs. RobustFair uses the fairness confusion matrix to find the input features most useful for each kind, and evaluates the accurate fairness of a network under adversarial perturbation.

Idea

The fairness confusion matrix from the AAAI-23 work classifies every prediction as true fair, true biased, false fair or false biased. RobustFair turns that classification into a search direction: starting from a seed instance and its similar counterparts, a gradient search is steered towards the perturbations that most increase a chosen off-diagonal cell, so the evaluation deliberately looks for accurate-but-biased and fair-but-inaccurate behaviour rather than only misclassification.

• Fairness-confusion-directed perturbations. The loss of the target cell guides the gradient, so a single search surfaces both robustness and individual-fairness failures.
• Evaluation on tabular benchmarks. Adult, Bank, COMPAS and German Credit, with baseline models retrained on the generated instances to measure the defence.
• Companion of RIFair. The KDD '26 work extends the same quadrant view to natural-language inputs with a black-box, semantics-preserving attack.

Code

The implementation, including seed preparation, the directed search, loss-trend analysis, retraining and result export, is at Xuran-LI/RobustFairEvaluation.

Citation

@article{li2023robustfair,
  title   = {RobustFair: Adversarial Evaluation through Fairness Confusion Directed Gradient Search},
  author  = {Li, Xuran and Wu, Peng and Dong, Kaixiang and Zhang, Zhen and Chen, Yanting},
  journal = {arXiv preprint arXiv:2305.10906},
  year    = {2023}
}