Fairness-aware Anomaly Detection via Fair Projection

TL;DR

Fairness-aware Anomaly Detection via Fair Projection proposes mapping data from multiple protected groups onto a shared, simple target distribution to enable fair density-based anomaly detection in an unsupervised setting. It introduces two core assumptions that make group fairness feasible in UAD and presents end-to-end variants (Im-FairAD and Ex-FairAD) that optimize both detection and fairness without extra fairness regularizers. The authors also propose a threshold-free fairness metric, ADPD, to holistically evaluate performance across all decision thresholds. Empirical results on six real-world datasets show improved trade-offs between detection accuracy and fairness under both balanced and skewed group splits, including robustness on abnormal data. This approach offers a practical, scalable path to fair anomaly detection in high-stakes domains.

Abstract

Unsupervised anomaly detection is a critical task in many high-social-impact applications such as finance, healthcare, social media, and cybersecurity, where demographics involving age, gender, race, disease, etc, are used frequently. In these scenarios, possible bias from anomaly detection systems can lead to unfair treatment for different groups and even exacerbate social bias. In this work, first, we thoroughly analyze the feasibility and necessary assumptions for ensuring group fairness in unsupervised anomaly detection. Second, we propose a novel fairness-aware anomaly detection method FairAD. From the normal training data, FairAD learns a projection to map data of different demographic groups to a common target distribution that is simple and compact, and hence provides a reliable base to estimate the density of the data. The density can be directly used to identify anomalies while the common target distribution ensures fairness between different groups. Furthermore, we propose a threshold-free fairness metric that provides a global view for model's fairness, eliminating dependence on manual threshold selection. Experiments on real-world benchmarks demonstrate that our method achieves an improved trade-off between detection accuracy and fairness under both balanced and skewed data across different groups.

Figures (7)

• Figure 1: The illustration of Im-FairAD. For simplicity, we only visualize two different attribute values $s_i, s_j \in S$ for the protected variable $S$, but Im-FairAD does not impose such a restriction. The 'High' and 'Low' denote the relative density in the target distribution.
• Figure 2: Accuracy-fairness trade-off on balanced splitting. Note that the baselines, FairOD and CFAD are tailored to tabular data.
• Figure 3: Fairness of all baselines on abnormal data from the test set.
• Figure 4: Accuracy-fairness trade-off on skewed splitting.
• Figure 5: Fairness of all baselines on abnormal data from test set.

Theorems & Definitions (9)

• Definition 1: Demographic parity agarwal2018reductions
• Definition 2: Equal opportunity Gajane_2017
• Claim 1
• Definition 3: Predictive equality chouldechova2017fair
• Proposition 1
• Proposition 2
• proof
• proof
• proof