Research map / Label-efficient and robust learning / OOD and anomaly detection
Swift Hydra: Self-Reinforcing Generative Framework for Anomaly Detection with Multiple Mamba Models
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.
Swift Hydra trains a C-VAE generator guided by an RL policy in latent space to synthesize challenging anomalies, then trains a Mixture of Mamba Experts detector on the augmented data for fast inference, evaluated on ADBench.
Problem
Anomaly detection models generalize poorly to unseen anomalies due to scarce anomaly data, and generative augmentation (GANs, VAEs, oversampling) lacks a strategy to produce challenging yet diverse anomalies; a single overparameterized detector is also slow at inference.
Approach
An RL policy operating on C-VAE latent variables generates diverse anomalies with a reward combining entropy and detector evasion; generated samples augment the detector, which is then distilled into a sparsely gated mixture of Mamba experts whose clusters are assigned via a penalized probabilistic scheme to avoid winner-take-all.
Key results
- Swift Hydra (MoME) reaches 0.93 average AUC-ROC over 57 ADBench datasets at 40/60 train/test ratio, versus 0.86 for ADGym and 0.82 for DTE.
- At 10/90 split, Swift Hydra (MoME) achieves 0.87 AUC-ROC with TIF 5.84s, versus 18.52s for the single-detector variant.
- Swift Hydra (MoME) inference time (4.01s TIF at 40%) is comparable to DTE (4.02s) and Rejex (3.89s), which are faster.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Swift Hydra (MoME) instead of ADGym | ADBench anomaly detection · ADBench (40/60% train/test) | AUC-ROC: 0.86 -> 0.93 | Helped |
| use Swift Hydra (MoME) instead of ADGym | ADBench anomaly detection · ADBench (30/70% train/test) | AUC-ROC: 0.82 -> 0.91 | Helped |
| use Swift Hydra (MoME) instead of ADGym | ADBench anomaly detection · ADBench (20/80% train/test) | AUC-ROC: 0.79 -> 0.90 | Helped |
| use Swift Hydra (MoME) instead of ADGym | ADBench anomaly detection · ADBench (10/90% train/test) | AUC-ROC: 0.77 -> 0.87 | Helped |
| replace a single large detector with MoME (ablation) | Swift Hydra · ADBench (40/60% train/test) | AUC-ROC: 0.91 -> 0.93 | Helped |
| replace a single large detector with MoME (ablation) | Swift Hydra · ADBench (30/70% train/test) | AUC-ROC: 0.90 -> 0.91 | Helped |
| replace a single large detector with MoME (ablation) | Swift Hydra · ADBench (20/80% train/test) | AUC-ROC: 0.87 -> 0.90 | Helped |
| replace a single large detector with MoME (ablation) | Swift Hydra · ADBench (10/90% train/test) | AUC-ROC: 0.86 -> 0.87 | Helped |
| replace a single large detector with MoME (ablation) | Swift Hydra · ADBench (40/60% train/test) | TIF (seconds): 13.11 -> 4.01 | Helped |
| replace a single large detector with MoME (ablation) | Swift Hydra · ADBench (30/70% train/test) | TIF (seconds): 14.38 -> 4.79 | Helped |
| replace a single large detector with MoME (ablation) | Swift Hydra · ADBench (20/80% train/test) | TIF (seconds): 16.13 -> 5.22 | Helped |
| replace a single large detector with MoME (ablation) | Swift Hydra · ADBench (10/90% train/test) | TIF (seconds): 18.52 -> 5.84 | Helped |
Surprising findings
- Swift Hydra (MoME) with 10% training data outperforms DTE and Rejex, which the authors frame as evidence of generalization from RL-generated anomalies.
Limitations
- Source code release is deferred until publication; implementation details are in the appendix.
- Theoretical results rely on strong assumptions (Gaussian clusters, equal priors, well-converged generator, Lipschitz constant via layer normalization).
Experimental setup
Datasets: ADBench
Benchmarks: ADBench
Metrics: AUC-ROC total inference time (TIF)
Baselines: DTE Rejex ADGym SMOTE Borderline-SMOTE ADASYN SVM-SMOTE CBO Oversampling GAN VAE-Geometry
Builds on: Conditional VAE Sparsely-Gated Mixture-of-Experts ReST Proximal Policy Optimization
Closest papers by meaning
- MambaAD: Exploring State Space Models for Multi-class Unsupervised Anomaly Detection (NeurIPS 2024 · similarity 0.84)
- Scalable, Explainable and Provably Robust Anomaly Detection with One-Step Flow Matching (NeurIPS 2025 · similarity 0.83)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.83)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.83)
- A Comprehensive Augmentation Framework for Anomaly Detection (AAAI 2024 · similarity 0.83)
- Single-Model Attribution of Generative Models Through Final-Layer Inversion (ICML 2024 · 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.