Rethinking Fair Anomaly Detection via Representation Learning

Anomaly detection (AD) is widely used in high-stakes domains but often exhibits systematic bias across sensitive groups, motivating fairness-aware AD. Existing methods typically treat fairness as a performance-fairness trade-off, especially under severe group imbalance. In this work, we show that this trade-off is not inherent to unsupervised AD, but arises from representation-level distortions induced by fairness constraints. We identify two fundamental failure modes of existing fair AD methods: vanishing protected-group influence under group imbalance and dimensional collapse caused by overly restrictive fairness regularization. We provide theoretical analysis showing how commonly-used fairness constraints can suppress minority-group signals and compress learned embeddings, degrading anomaly discrimination. To address these issues, we propose FADIG, a fairness-aware contrastive learning framework that preserves representation expressiveness while promoting group fairness. FADIG combines an adaptively re-balanced autoencoder with a fairness-aware contrastive objective, and we establish theoretical guarantees that it alleviates both FADIG achieves accurate and fair anomaly detection under severe imbalance.