Research map / Label-efficient and robust learning / OOD and anomaly detection

Contamination-Resilient Anomaly Detection via Adversarial Learning on Partially-Observed Normal and Anomalous Data

ICML 2024 · Label-efficient and robust learning › OOD and anomaly detection · approach: Unified multi-class anomaly detection · 0 citations (OpenAlex)

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 a flipping-regularized, bidirectional GAN (CR-GAN) that learns the normal-data distribution from a contaminated dataset plus small clean normal and anomaly sets, and scores anomalies by reconstruction error and latent norm.

Problem

Anomaly detection when the training set is contaminated with anomalies and only small clean normal and clean anomaly sets are available, with overfitting from the tiny auxiliary sets and incomplete coverage of anomaly types.

Approach

Reformulate f-divergence minimization between mixtures of contaminated data, clean normal data and a generator, so the generator converges to the normal distribution; add a label-flipping mechanism with adaptive flip probability to reduce overfitting to the small clean sets, and a BiGAN-based score combining reconstruction error and latent norm.

Key results

• On MNIST with 20% contamination and one normal type, the proposed method reaches 93.7 AUROC versus 90.8 for the best baseline (AA-BiGAN) in Table 1.
• Ablation on MNIST: adding the clean normal and anomaly datasets raises AUROC from 84.0 to 88.2, and flipping with adaptive gamma raises it further to 90.6.
• Achieves best AUROC on seven of nine classic anomaly detection datasets (e.g., Cardio 99.4 vs 97.6).

Measured effects reported in the paper

Surprising findings

• The method's performance under contamination drops more slowly than baselines, yet the flipping probability must be kept small (e.g., 0.05 or 0.1) to satisfy the theoretical condition.

Limitations

• Theoretical convergence requires disjoint supports of normal and anomaly distributions and a flip probability bounded by a condition on beta and lambda.
• With incomplete anomaly types, the generator converges to a mixture that still includes unobserved anomaly types.
• Experiments on toy datasets rely on simulated contamination by category splits.

Experimental setup

Datasets: MNIST Fashion-MNIST CIFAR10 20newsgroups UNSW-NB15 HAR Arrhythmia Cardio Satellite Satimage-2 Shuttle Thyroid

Benchmarks: MNIST Fashion-MNIST CIFAR10 20newsgroups HAR UNSW-NB15 Arrhythmia Cardio Satellite Satimage-2 Shuttle Thyroid

Metrics: AUROC

Baselines: Deep SVDD Deep Isolation Forest SLAD ICL Deep SAD FeaWAD RoSAS PReNet AA-BiGAN SOEL

Builds on: GAN LSGAN BiGAN Deep SAD AA-BiGAN

Closest papers by meaning

• An Evidence-Based Post-Hoc Adjustment Framework for Anomaly Detection Under Data Contamination (NeurIPS 2025 · similarity 0.87)
• Robust Conformal Outlier Detection under Contaminated Reference Data (ICML 2025 · similarity 0.86)
• RODEO: Robust Outlier Detection via Exposing Adaptive Out-of-Distribution Samples (ICML 2024 · similarity 0.86)
• Boosting Fine-Grained Visual Anomaly Detection with Coarse-Knowledge-Aware Adversarial Learning (AAAI 2025 · similarity 0.86)
• ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.86)
• Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders (ICLR 2025 · similarity 0.85)