Research map / Label-efficient and robust learning / OOD and anomaly detection
Fairness-aware Anomaly Detection via Fair Projection
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 FairAD, an unsupervised anomaly detector that learns a projection mapping each demographic group's normal data to a shared truncated Gaussian target (Sinkhorn loss plus reconstruction), with a threshold-free fairness metric ADPD.
Problem
Group fairness in unsupervised anomaly detection, where only normal data are available, so fairness on anomalies cannot be directly enforced, and existing fairness metrics depend on thresholds.
Approach
Learn an encoder that maps every protected group's normal data onto a common truncated isotropic Gaussian target via Sinkhorn distance, with a reconstruction decoder preserving information; density in the shared latent space gives the anomaly score, so fairness follows without extra fairness regularizers. Also introduces ADPD, a threshold-free demographic parity difference averaged over thresholds
Key results
- On COMPAS, Im-FairAD reaches 63.89% AUC with 4.76% ADPD (all) under balanced splitting, versus 59.23% AUC / 10.59% ADPD for LOF.
- On Adult under skewed splitting, Ex-FairAD achieves 69.56% AUC with 1.79% ADPD (all), compared with 58.30% AUC and 14.35% ADPD for FairOD.
- On Titanic, Ex-FairAD reaches 64.87% AUC versus 55.63% for CFAD, with ADPD (normal) of 6.35% versus 26.77% for CFAD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Im-FairAD instead of the strongest reported baseline | Unsupervised anomaly detection · COMPAS, Adult, Credit, CelebA, Titanic, and SP; balanced and skewed splits where reported | AUC | Mixed |
| Use Im-FairAD instead of the strongest reported baseline | Unsupervised anomaly detection · COMPAS, Adult, Credit, CelebA, Titanic, and SP; balanced and skewed splits where reported | ADPD | Mixed |
| Transition from balanced to skewed demographic-group splitting | Fairness-aware anomaly detection methods · COMPAS, Adult, Credit, and CelebA | ADPD | Mixed |
| Use Im-FairAD instead of Deep SVDD | Anomaly detection · COMPAS, Adult, Credit, CelebA, Titanic, and SP; reported splits | F1 | Mixed |
Surprising findings
- Fairness-aware baselines (e.g., Deep Fair SVDD, CFAD) degrade markedly in ADPD under skewed group splits, while the proposed methods stay stable.
- Fairness on abnormal test data appears to transfer from normal-data fairness, since abnormal samples are assumed to be a mapped version of normal data.
Limitations
- Fairness guarantees on anomalies rely on assumptions (learnable abnormality, transferable and generalizable parity) that are only empirically supported.
- Results on small datasets (SP, Titanic) show high variance across runs.
- Im-FairAD does not clearly dominate Ex-FairAD in all settings, and some baselines (e.g., FarconVAE+LOF) achieve the best ADPD in specific cases.
Experimental setup
Datasets: Adult COMPAS Credit CelebA Titanic SP
Benchmarks: Adult COMPAS Credit CelebA Titanic SP
Metrics: AUC F1 ADPD fairness ratio
Baselines: LOF Deep SVDD FarconVAE+LOF FarconVAE+Deep SVDD FairOD Deep Fair SVDD CFAD
Builds on: Deep SVDD Sinkhorn distance FarconVAE Deep Fair SVDD
Closest papers by meaning
- Fair Deepfake Detectors Can Generalize (NeurIPS 2025 · similarity 0.85)
- Fairness without Demographics through Shared Latent Space-Based Debiasing (AAAI 2024 · similarity 0.85)
- On the Maximal Local Disparity of Fairness-Aware Classifiers (ICML 2024 · similarity 0.85)
- FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.85)
- Searching for Unfairness in Algorithms’ Outputs: Novel Tests and Insights (AAAI 2025 · similarity 0.85)
- Intersectional Fairness via Mixed-Integer Optimization (NeurIPS 2026 · similarity 0.84)
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.