Attacking Counterfactually Fair Transformations: Increased Accuracy at the Cost of Individual Fairness
MCML Authors
Abstract
Abstract
Counterfactually fair transformations remove the causal effect of protected attributes on observed features to provide a fair representation for training prediction models. This introduces a fairness-accuracy tradeoff, since real-world targets remain influenced by protected attributes. We show that even in the absence of latent confounding and with correctly specified causal relations between all observed variables, intentional omission of variables in the causal graph can still undermine the transformation: While the individual-level causal effect becomes unidentifiable, the mean squared error to the factual data is reduced at the cost of an increased violation of individual fairness. The proposed attack cannot be detected by constraint-based causal discovery and offers a business advantage to malicious actors obliged to apply counterfactually fair transformations. For linear SCMs with interactions and absolutely continuous noise distributions, we prove that marginalizing a latent covariate or mediator strictly reduces the mean squared error to factual values, while the violation of individual fairness increases. Our findings expose a fundamental limitation of quantile-based fair transformations and motivate regulatory safeguards.
inproceedings RB26
CDM @UAI 2026
Workshop on Causality for Decision Making at the 42nd Conference on Uncertainty in Artificial Intelligence. Amsterdam, The Netherlands, Aug 17-21, 2026. To be published. Preprint available.Authors
S. Rittel • L. BothmannLinks
URLResearch Area
BibTeXKey: RB26