A Survey on Deep Anomaly Detection: Weak Supervision, Automation, and Pre-training

A professionally curated list of awesome resources (paper, code, data, etc.) on Resource-Constrained Deep Anomaly Detection (RCDAD), which is the first work to comprehensively and systematically summarize the recent advances of deep anomaly detection under resource constraints (data, label, and expertise) from the methodology design to the best of our knowledge.

We will continue to update this list with the latest resources. If you find any missed resources (paper/code) or errors, please feel free to open an issue or make a pull request.

Note: This repository is currently under construction. Detailed resources (including code and full paper) will be fully supplemented once the paper is officially published.

Taxonomy

Papers & Methods

1. Weak Supervision

Addresses Limited Annotation Resources via Incomplete, Inexact, and Inaccurate Supervision.

Incomplete Weak Supervision (Limited Labels)

Method	Venue	Backbone	Modalities	Key Idea	Code
OE	KDD'14	-	Tabular	Anomaly feature representation learning	No
XGBOD	IJCNN'18	-	Tabular	Anomaly feature representation learning	Yes
DeepSAD	ICLR'20	MLP	Tabular	Anomaly feature representation learning	Yes
ESAD	arXiv'20	MLP	Tabular	Anomaly feature representation learning	No
DSSAD	ICASSP'21	CNN	Image/Video	Anomaly feature representation learning	No
REPEN	KDD'18	MLP	Tabular	Anomaly feature representation learning	No
AA-BiGAN	IJCAI'22	GAN	Tabular	Anomaly feature representation learning	Yes
Dual-MGAN	TKDD'22	GAN	Tabular	Anomaly feature representation learning	Yes
WAKE	TKDE'23	AE	Time series	Anomaly feature representation learning	No
DevNet	KDD'19	MLP	Tabular	Anomaly score learning	Yes
PReNet	KDD'23	MLP	Tabular	Anomaly score learning	No
FEAWAD	TNNLS'21	AE	Tabular	Anomaly score learning	Yes
Overlap	KDD'23	-	Tabular	Anomaly score learning	Yes
SNPAD	TKDE'23	MLP	Tabular	Probabilistic Modeling	No
SNARE	KDD'09	-	Graph	Graph learning and label propagation	No
AESOP	KDD'14	-	Graph	Graph learning and label propagation	No
SemiGNN	ICDM'19	MLP+Attention	Graph	Graph learning and label propagation	No
SemiGAD	IJCNN'21	GNN	Graph	Graph learning and label propagation	No
Meta-GDN	WWW'21	GNN	Graph	Graph learning and label propagation	Yes
SemiADC	IS Journal'21	GAN	Graph	Graph learning and label propagation	No
SSAD	JAIR'13	-	Tabular	Active learning	No
AAD	ICDM'16	-	Tabular	Active learning	Yes
SLA-VAE	WWW'22	VAE	Time series	Active learning	No
Meta-AAD	ICDM'20	MLP	Tabular	Reinforcement learning	Yes
DPLAN	KDD'21	MLP	Tabular	Reinforcement learning	No
GraphUCB	WSDM'19	-	Graph	Reinforcement learning	Yes
SR-CNN	KDD'19	TCN	Time series	Data Augmentation and Frequency Processing	Yes
RobustTAD	KDDW'20	U-Net	Time series	Data Augmentation and Frequency Processing	No
TFAD	CIKM'22	TCN	Time series	Data Augmentation and Frequency Processing	Yes
NCAD	IJCAI'22	TCN	Time series	Data Augmentation	Yes
RealNet	CVPR'24	CNN	Image	Anomaly feature representation learning	Yes

Inexact Weak Supervision (Coarse-grained Labels)

Method	Venue	Backbone	Modalities	Key Idea	Code
MIL	CVPR'18	MLP	Video	Multiple Instance Learning	Yes
TCN-IBL	ICIP'19	CNN	Video	Multiple Instance Learning	No
AR-Net	ICME'20	MLP	Video	Multiple Instance Learning	Yes
RTFM	ICCV'21	CNN+Attention	Video	Multiple Instance Learning	Yes
Motion-Aware	BMVC'19	AE+Attention	Video	Multiple Instance Learning	No
CRF-Attention	ICCV'21	TRN+Attention	Video	Multiple Instance Learning	No
MPRF	IJCAI'21	MLP+Attention	Video	Multiple Instance Learning	No
MCR	ICME'22	MLP+Attention	Video	Multiple Instance Learning	No
CoMo	CVPR'23	GCN	Video	Multiple Instance Learning	No
MGFN	AAAI'23	CNN+Attention	Video	Multiple Instance Learning	Yes
CNL	TCAS II'22	AE+Attention	Video	Multiple Instance Learning	No
UMIL	CVPR'23	MLP	Video	Multiple Instance Learning	Yes
XEL	SPL'21	MLP	Video	Cross-epoch Learning	Yes
MIST	CVPR'21	MLP+Attention	Video	Multiple Instance Learning	Yes
MSLNet	AAAI'22	Transformer	Video	Multiple Instance Learning	Yes
SRF	SPL'20	MLP	Video	Self Reasoning	No
WETAS	ICCV'21	MLP	Time-series/Video	Dynamic Time Warping	No
Inexact AUC	ML Journal'20	AE	Tabular	AUC maximization	No
Isudra	TIST'21	-	Time-series	Bayesian optimization	Yes
VADCLIP	AAAI'24	CLIP	Video	Multiple Instance Learning	Yes
PEMIL	CVPR'24	CLIP/ViT	Video	Multiple Instance Learning	Yes
Fed-WSVAD	AAAI'25	CLIP	Video	Multiple Instance Learning	Yes
PLOVAD	TCSVT'25	CLIP	Video	Multiple Instance Learning	Yes
OVVAD	CVPR'24	VLM	Video	Multiple Instance Learning	No
TPWNG	CVPR'24	Transformer	Video	Multiple Instance Learning	No

Inaccurate Weak Supervision (Noisy Labels)

Method	Venue	Backbone	Modalities	Key Idea	Code
LAC	CIKM'21	MLP/GBDT	Tabular	Ensemble learning	No
ADMoE	AAAI'23	Agnostic	Tabular	Ensemble learning	Yes
BGPAD	ICNP'21	LSTM+Attention	Time series	Denoising network	Yes
SemiADC	IS Journal'21	GAN	Graph	Denoising network	No
TSN	CVPR'19	GCN	Video	GCN	Yes
Unity	SIGMOD'25	MLP/DNN	Tabular	Ensemble learning	Yes
RHGL	IJCAI'24	GNN	Graph	Graph learning and label propagation	No
M3DM-NR	TPAMI'25	CNN/Transformer	Image/3D	Denoising network	No

2. Automation

Addresses Limited Expertise Resources via Optimization, Meta-Learning, and LLMs.

Automation via Optimization Algorithm

Method	Venue	Backbone	Modalities	Key Idea	Code
PyODDS	WWW'20	NN-based	Tabular	Iterative Optimization Algorithm	Yes
TODS	AAAI'21	NN-based	Time series	Iterative Optimization Algorithm	Yes
AutoOD	ICDE'21	Auto-Encoder	Tabular	RL-based NAS	Yes
AutoPatch	AutoML'23	CNN	Image	NAS for Visual Anomaly Segmentation	Yes
PASTA	T-ETCI'24	RNN	Time series	NAS for Time Series AD	Yes
RLNAS	IoT-J'24	Auto-Encoder	Time series	RL-based NAS	Yes
TSAP	SDM'25	CNN	Time series	Self-tuning Augmentation	Yes

Automation via Meta-Learning

Method	Venue	Backbone	Modalities	Key Idea	Code
MetaOD	NeurIPS'21	ML-based	Tabular	Performance Matrix Completion	Yes
ELECT	ICDM'22	ML-based	Tabular	Internal Performance Measures	Yes
Hydra	AutoML'23	NN-based	Time series	Meta-Recommender for Model Selection	Yes
UMSTAD	ICLR'23	NN-based	Time series	Surrogate Metrics of Model Performance	No
ADGym	NeurIPS'23	NN-based	Tabular	Model Components Benchmark	Yes
HYPER	KDD'24	Auto-Encoder	Tabular	Hypernetwork for Generating Optimal Weights	Yes
LogCraft	ASE'24	Auto-Encoder	Log	AutoML for Log AD	Yes
ADecimo	ICDE'24	ML-based	Time series	Model Selection	Yes
MetaUAS	NeurIPS'24	VFM	Image	One-Prompt Meta-Learning	Yes
MetaCAN	CIKM'25	NN-based	Tabular	Few-shot Meta-Learning	Yes

Automation via Large Language Models

Method	Venue	Backbone	Modalities	Key Idea	Code
AD-LLM	ACL'25	GPT-4 / Llama	Tabular	Zero-Shot Model Recommendation	Yes
AD-AGENT	arXiv'25	GPT-4 (Agent)	Tabular	Multi-Agent Code Generation	Yes

3. Pre-training

Addresses Limited Task-Data Resources via Specialized and Foundation Models.

Specialized Pre-trained Models

Method	Venue	Backbone	Modalities	Key Idea	Code
DeepSVDD	ICML'18	AutoEncoder	Tabular	Reconstruction	Yes
DeepSAD	ICLR'20	AutoEncoder	Tabular	Reconstruction	Yes
UP2ME	ICML'24	Transformer	Time-series	Masked Modeling	Yes
RealNet	CVPR'24	Diffusion	Img/TS	Anomaly Synthesis	Yes
TSAP	SDM'25	CNN	Time-series	Anomaly Synthesis	Yes
Panda	CVPR'21	ResNet (ImageNet)	Image	Feature Adaptation	Yes
SimpleNet	CVPR'23	ResNet (ImageNet)	Image	Feature Adaptation	Yes
MSC-AD	AAAI'23	ResNet (ImageNet)	Image	Contrastive Learning	Yes
SPD	ECCV'22	ResNet (ImageNet)	Image	Anomaly Synthesis	Yes
CVDD	ACL'19	GloVe	Text	Feature Adaptation	Yes
DATE	NAACL'21	BERT	Text	Masked Modeling	Yes
MRONet	CVPR'21	LSTM	Time-series	Forecasting	No
WAKE	TKDE'23	GRU-AE	Time-series	Reconstruction	Yes
GUDI	ICDE'24	GNN	Graph	Domain-Agnostic Transfer	Yes
DIAD	KDD'23	GAM	Tabular	Data-Efficient Transfer	Yes

Large Language Models for AD

Method	Venue	Backbone	Modalities	Key Idea	Code
TAD-Bench	arXiv'25	BERT / GPT / Llama	Tabular / Text	Feature Adaptation	Yes
AD-LLM	ACL'25	GPT-4	Tabular	Prompting / Agent	Yes
AnoLLM	ICLR'25	SmolLM	Tabular	Generative Modeling	Yes
OFA	NeurIPS'23	GPT-2	Time-series	Prompting / Reprogramming	Yes
AnomalyLLM	arXiv'24	LLM	Graph	Prompting / In-Context	Yes
AnomalyGPT	AAAI'24	VLM	Image	Cross-Modal	Yes
Any-Anomaly	CVPR'24	VLM	Image	Prompt Regularization	No
InContext-AD	CVPR'24	Transformer	Image	In-Context Learning	Yes
WinCLIP	CVPR'23	CLIP	Image	Cross-Modal Alignment	Yes
VadCLIP	AAAI'24	CLIP	Video	Cross-Modal Alignment	Yes

Foundation Models for AD

Method	Venue	Backbone	Modalities	Key Idea	Code
CM2	WWW'24	BERT-based Encoder	Tabular	Masked Modeling	Yes
TabPFN	Nature'25	Transformer	Tabular	Anomaly Synthesis	Yes
TimeGPT	arXiv'23	Transformer	Time-series	Forecasting	Yes
MOMENT	ICML'24	T5-Encoder	Time-series	Masked Modeling	Yes
Timer	ICML'24	GPT-style Decoder	Time-series	Generative Modeling	Yes
DADA	arXiv'24	Transformer / AE	Time-series	Anomaly Synthesis	Yes
UniTS	NeurIPS'24	Transformer	Time series	Unified Time Series Model	Yes
GCCAD	TKDE'22	GNN	Graph	Contrastive Learning	Yes