Research map / Label-efficient and robust learning / OOD and anomaly detection
Dense Projection for Anomaly Detection
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 DPAD, which trains a bias-free neural projection to make normal training data locally dense via adaptively reweighted local distance minimization with norm regularization, then scores test points by kNN distance in the learned space.
Problem
Unsupervised anomaly detection where density-based methods struggle in high dimensions and one-class methods rely on restrictive decision-boundary assumptions (e.g., hyperspheres), especially when normal data is complex or multi-class.
Approach
Learn a low-dimensional projection that maximizes local density of normal data by minimizing exponentially reweighted pairwise distances (weights detached, adaptively set), regularized by weight-matrix Frobenius norms and bias-free layers to avoid collapse; then apply kNN in the embedding for detection.
Key results
- On CIFAR-10 one-class detection, DPAD reaches 74.5% average AUC, the best among compared methods (HRN 71.3%, IGD 74.3%).
- On Fashion-MNIST, DPAD averages 93.7% AUC, behind UMAP+kNN (95.2%) but the closest state-of-the-art competitor.
- In the 9-1 multi-class normal setting on CIFAR-10, DPAD achieves 66.1% AUC versus 54.3% for DROCC and 52.3% for DSVDD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DPAD instead of IGD | CIFAR-10 one-class anomaly detection · CIFAR-10 | AUC (%): 74.3 -> 74.5 | Helped |
| use DPAD instead of UMAP+kNN | Fashion-MNIST one-class anomaly detection · Fashion-MNIST | AUC (%): 95.2 -> 93.7 | Hurt |
| use DPAD instead of GOCC | Arrhythmia anomaly detection · Arrhythmia | F1-score (%): 61.8 ± 1.8 -> 66.7 ± 0.0 | Helped |
| use DPAD instead of GOCC | Abalone anomaly detection · Abalone | F1-score (%): 66.1 ± 4.3 -> 66.7 ± 1.5 | Helped |
| use DPAD instead of GOCC | Campaign anomaly detection · Campaign | F1-score (%): 74.1 ± 2.5 -> 73.4 ± 1.5 | Hurt |
| use DPAD instead of t-SNE+kNN | MAGIC-gamma anomaly detection · MAGIC-gamma | F1-score (%): 76.6± 0.0 -> 74.0 ± 0.5 | Hurt |
| use DPAD instead of DROCC | CIFAR-10 9-1 anomaly detection · CIFAR-10 | AUC (%): 54.3 -> 66.1 | Helped |
| use DPAD instead of t-SNE+kNN | Fashion-MNIST 9-1 anomaly detection · Fashion-MNIST | AUC (%): 78.7 -> 70.2 | Hurt |
| add adaptive weights eᵂᵢⱼ (ablation) | DPAD · CIFAR-10 | AUC (%): 68.5 -> 74.5 | Helped |
| add parameter constraint Cᵂ (ablation) | DPAD · Fashion-MNIST | AUC (%): 89.2 -> 93.7 | Helped |
| set γ to 100 (ablation) | DPAD · Fashion-MNIST | AUC (%): 93.3 -> 89.2 | Hurt |
| vary λ within the reported tested values (ablation) | DPAD · Fashion-MNIST | AUC (%): 89.2; 92.5; 92.5; 92.6; 91.9; 92.0 | No clear effect |
Surprising findings
- Dimensionality reduction plus kNN (UMAP, t-SNE) outperforms most deep anomaly detectors on Fashion-MNIST, and LOF with DR underperforms kNN.
- Removing bias terms is necessary to avoid model collapse, and excessively large gamma degrades learned dense representations.
Limitations
- Density estimators such as KDE and LOF are slower and less effective than kNN with DPAD; kNN scoring still scales with training set size.
- Theoretical guarantees (Lipschitz bound and initialization lower bound on weights) cover only initialization and do not fully characterize trained behavior.
- Performance on CIFAR-10 is far below supervised levels and tabular results rely on F1 with varying baselines reproduced by authors.
Experimental setup
Datasets: CIFAR-10 Fashion-MNIST Arrhythmia Abalone Campaign MAGIC Gamma
Benchmarks: CIFAR-10 Fashion-MNIST Arrhythmia Abalone Campaign MAGIC Gamma
Metrics: AUC F1 score
Baselines: kNN k-Means LOF OCSVM Isolation Forest KDE DAE DAGMM ADGAN DSVDD OCGAN TQM
Builds on: Deep SVDD kNN anomaly detection autoencoder UMAP t-SNE
Closest papers by meaning
- Unsupervised Anomaly Detection by Robust Density Estimation (AAAI 2022 · similarity 0.87)
- A General Framework For Detecting Anomalous Inputs to DNN Classifiers (ICML 2021 · similarity 0.86)
- Unsupervised Anomaly Detection for Tabular Data Using Deep Noise Evaluation (AAAI 2025 · similarity 0.85)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.84)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.84)
- A Dataset for Semantic Segmentation in the Presence of Unknowns (CVPR 2025 · similarity 0.83)
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.