Research map / Adversarial robustness and security / Certified and empirical robustness
Causal Adversarial Perturbations for Individual Fairness and Robustness in Heterogeneous Data Spaces
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes causal adversarial perturbations (CAP) built from a semi-latent space fair metric over an SCM, defines CAPI fairness linking individual fairness, counterfactual fairness and adversarial robustness, and trains a classifier (CAPIFY) with a causal adversarial regularizer.
Problem
Individual fairness, adversarial robustness and causal structure have been studied separately or in pairs, but not jointly; existing fair metrics do not capture causal structure over heterogeneous (categorical and continuous) features with discrete sensitive attributes.
Approach
A fair metric is defined in a semi-latent space (sensitive features kept, other features mapped to exogenous noise) with the sensitive attribute treated as a pseudometric; the resulting causal adversarial perturbation set (union of continuous balls around counterfactual twins) is used for adversarial training, yielding a regularizer that jointly enforces individual fairness, counterfactual fairnes
Key results
- CAPIFY reduces CAPI unfairness U_0.05 substantially versus baselines on Adult (0.07 vs 0.22 for AL and ERM) and COMPAS (0.06 vs 0.21 for ERM).
- CAPIFY achieves accuracy close to ERM, e.g. 0.78 vs 0.80 on Adult and 0.63 vs 0.68 on COMPAS, a slight accuracy trade-off.
- Results are averaged over 100 seeds; CAPIFY's standard deviation averages 0.028 versus 0.038 for other methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAPIFY instead of ERM | classifier training · Adult | accuracy: 0.80 -> 0.78 | Hurt |
| Use CAPIFY instead of ERM | classifier training · COMPAS | accuracy: 0.68 -> 0.63 | Hurt |
| Use CAPIFY instead of ERM | classifier training · IMF | accuracy: 0.72 -> 0.72 | No clear effect |
| Use CAPIFY instead of ERM | classifier training · LIN | accuracy: 0.68 -> 0.62 | Hurt |
| Use CAPIFY instead of ERM | classifier training · Loan | accuracy: 0.85 -> 0.70 | Hurt |
| Use CAPIFY instead of ERM | classifier training · NLM | accuracy: 0.71 -> 0.45 | Hurt |
| Use CAPIFY instead of ERM | classifier training · Adult | UAI (Δ=0.05): 0.21 -> 0.07 | Helped |
| Use CAPIFY instead of ERM | classifier training · COMPAS | UAI (Δ=0.05): 0.21 -> 0.06 | Helped |
| Use CAPIFY instead of ERM | classifier training · IMF | UAI (Δ=0.05): 0.05 -> 0.04 | Helped |
| Use CAPIFY instead of ERM | classifier training · LIN | UAI (Δ=0.05): 0.44 -> 0.19 | Helped |
| Use CAPIFY instead of ERM | classifier training · Loan | UAI (Δ=0.05): 0.35 -> 0.22 | Helped |
| Use CAPIFY instead of ERM | classifier training · NLM | UAI (Δ=0.05): 0.57 -> 0.38 | Helped |
Surprising findings
- The CAL baseline shows unsatisfactory accuracy, which the authors attribute to known issues in prior work.
- Real-world data showed a larger reduction in unfairness than in accuracy loss.
Limitations
- No theoretical guarantee that the trained classifier satisfies individual fairness.
- Analysis restricted to discrete sensitive attributes; continuous sensitive attributes are only claimed to work without experiments.
- Theoretical unfair-area results derived for linear SCMs and linear classifiers only.
Experimental setup
Datasets: Adult COMPAS LIN NLM IMF Loan
Benchmarks: Adult COMPAS LIN NLM IMF Loan
Metrics: accuracy Matthews correlation coefficient unfair area indicator (CAPI unfairness) counterfactual unfair area non-robust percentage
Baselines: AL (adversarial learning) CAL ERM LLR ROSS
Builds on: locally linear regularizer adversarial learning for fairness additive noise models pseudometric protected features
Closest papers by meaning
- On the Alignment between Fairness and Accuracy: from the Perspective of Adversarial Robustness (ICML 2025 · similarity 0.90)
- Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 2024 · similarity 0.87)
- Adversarial Fairness Network (AAAI 2024 · similarity 0.87)
- On the Fairness ROAD: Robust Optimization for Adversarial Debiasing (ICLR 2024 · similarity 0.87)
- The Fragility of Fairness: Causal Sensitivity Analysis for Fair Machine Learning (NeurIPS 2024 · similarity 0.87)
- DAFA: Distance-Aware Fair Adversarial Training (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.