Fair Anomaly Detection for Underrepresented Groups

This direction is split into other directions: fairness-aware anomaly detection. Its last version stays readable and citable here.

Version 1 of 124 papers33 conceptsfrozen 2026-10-04

The fundamental goal of anomaly detection is to identify data points that deviate significantly from the norm. However, in real-world datasets, "the norm" is often defined by the majority population. This creates a dangerous side effect: individuals from underrepresented demographic groups—who are statistically rare by definition—are frequently flagged as anomalies simply because they do not match the dominant statistical patterns.

Recent research has identified this as a "representation disparity," where the failure to capture the specific characteristics of minority groups leads to systematic algorithmic bias.

The Rarity-Anomaly Conflation

The core intuition behind most unsupervised anomaly detection (UAD) models is that "normal" data is dense and frequent, while "anomalous" data is sparse and rare. As Leman Akoglu points out in Anomaly Mining: Past, Present and Future, these systems often conflate statistical rarity with malicious or erroneous behavior.

When a model is trained primarily on a majority group, it learns to characterize that group's anatomical or behavioral variations as the universal standard for normality. Consequently, as Felix Meissen and colleagues demonstrate in (Predictable) Performance Bias in Unsupervised Anomaly Detection, there is a predictable, linear relationship between a group's representation in the training data and the model's performance on that group. Minority groups suffer from higher false-positive rates because the model treats their legitimate, healthy variations as outliers.

Why Standard Models Fail Minority Groups

The bias in these systems is not always a simple matter of "bad data." It often stems from a fundamental mismatch between the algorithm's design assumptions and the multi-modal nature of real-world populations. Leman Akoglu, Rui Xi, and Xueying Ding argue in Outlier Detection Bias Busted that even with balanced data, an algorithm might inadvertently penalize specific group characteristics if those characteristics are naturally more sparse or noisy.

Consider two common mechanisms for detection:

  1. Reconstruction-based models: These models (like autoencoders) try to compress and then reconstruct data. If the model is optimized for the majority, it will reconstruct majority samples well and minority samples poorly. This "reconstruction error" is then used as an anomaly score, unfairly flagging the minority.
  2. Density-based models: These models flag points in low-density regions of the feature space. Because underrepresented groups occupy smaller, sparser regions of the space, they are naturally more likely to be flagged as "atypical."

Mert Yuksekgonul and James Zou suggest in Beyond Confidence: Reliable Models Should Also Consider Atypicality that models must explicitly distinguish between confidence (how sure the model is about a prediction) and atypicality (how rare the sample is). Without this distinction, rarity is almost always interpreted as an error.

Strategies for Fair Detection

To solve this, researchers are moving toward "fairness-aware" detection mechanisms that incorporate parity constraints. The goal is to ensure that being part of a statistical minority does not result in being unfairly flagged.

  • Population Bias Filtering: In web-scale applications, Jamie Pool and the Microsoft team developed Lumos, which uses propensity score matching to filter out shifts in user demographics. By accounting for external factors like geographic distribution, they prevent demographic shifts from being misidentified as system failures.
  • Fair Projections: Feng Xiao and Jicong Fan propose FairAD, which projects data from different demographic groups into a single, compact "target distribution." This mathematically aligns different groups so the model learns a shared definition of normality.
  • Re-balanced Learning: The FAIRAD framework by Lecheng Zheng and Jingrui He uses a re-balanced autoencoder combined with contrastive learning. This forces the model to learn the specific features of minority groups rather than ignoring them in favor of the dominant statistical patterns.
  • Adversarial Blindness: Hanyu Song and Hongfu Liu introduced DCFOD, which uses adversarial training to ensure that the data representations used for outlier detection are "blind" to sensitive attributes like race or gender.

The High Stakes of Clinical and Social Applications

The need for these fair mechanisms is most acute in high-stakes domains. In medical AI, Karina Zadorozhny and Paul Elbers warn in Out-of-Distribution Detection for Medical Applications that OOD detectors may incorrectly flag atypical patients as anomalies, potentially denying them care or triggering unnecessary interventions. Similarly, in federated learning, Edith C. H. Ngai notes in ARFL that automated weighting systems designed to exclude "corrupted" data risk penalizing minority groups whose data is simply harder for a global model to learn.

By decoupling statistical rarity from anomalousness, this research direction aims to build systems that are robust not just to errors, but to the inherent diversity of the real world.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_fair_anomaly_detection_for_underrepresen_be4d5f26,
  title = {Fair Anomaly Detection for Underrepresented Groups},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fair-anomaly-detection-for-underrepresented-groups/rdr_be4d5f26bc3e419aab52cfdfb08016fd},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_fair_anomaly_detection_for_underrepresen_be4d5f26_v1,
  title = {Fair Anomaly Detection for Underrepresented Groups},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fair-anomaly-detection-for-underrepresented-groups/rdr_be4d5f26bc3e419aab52cfdfb08016fd/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}