Anomaly Detection Unleashed: From Edge AI to Quantum Circuits, What’s Next?

Latest 50 papers on anomaly detection: Nov. 30, 2025

Anomaly detection is the bedrock of robust AI systems, crucial for everything from securing smart grids to flagging medical irregularities. It’s a field constantly evolving, grappling with challenges like contaminated data, the need for real-time performance, and interpretability. Recent research is pushing the boundaries, offering groundbreaking solutions across diverse domains by leveraging cutting-edge AI paradigms.

The Big Idea(s) & Core Innovations

One of the most compelling trends is the drive towards smarter, more resilient systems. For instance, securing critical infrastructure is a major theme. The paper “An AI-Enabled Hybrid Cyber-Physical Framework for Adaptive Control in Smart Grids” by Muhammad Siddique and Sohaib Zafar from NFC IET and LUMS, introduces a hybrid framework integrating agent-based modeling, reinforcement learning, and game theory. This aims to create self-healing smart grids capable of adapting to and recovering from cyberattacks, demonstrating superior performance in control cost and system stability.

Complementing this, “Federated Anomaly Detection and Mitigation for EV Charging Forecasting Under Cyberattacks” explores federated learning for secure EV charging forecasting, emphasizing privacy-preserving, robust detection and mitigation against adversarial attacks. The framework developed by Author Name 1 and Author Name 2 from Institution A and B shows that integrating real-time detection with mitigation improves forecast accuracy under hostile conditions.

Another significant innovation tackles the perennial problem of contaminated training data. “Anomaly Detection with Adaptive and Aggressive Rejection for Contaminated Training Data” by Jungi Lee et al. from ELROILAB Inc. introduces AAR, a novel method that dynamically estimates contamination and aggressively rejects anomalies using statistical thresholds and Gaussian mixture models. This approach significantly boosts AUROC on contaminated datasets, proving that cleaner data directly leads to better models.

In the realm of explainability and interpretation, several papers shine. “Explainable Visual Anomaly Detection via Concept Bottleneck Models” by T. Liu et al. from University of Technology, Shenzhen, uses concept bottleneck models to make visual anomaly detection more interpretable. This allows systems not just to detect anomalies, but to explain why something is anomalous, fostering trust in AI for industrial applications. Similarly, “EVA-Net: Interpretable Brain Age Prediction via Continuous Aging Prototypes from EEG” by Kunyu Zhang et al. introduces EVA-Net, an interpretable framework that uses continuous aging prototypes from EEG data to not only predict brain age but also identify neurological anomalies through a novel Prototype Alignment Error (PAE) metric.

Large Language Models (LLMs) are also finding surprising applications beyond text. “Evaluation of Large Language Models for Numeric Anomaly Detection in Power Systems” by Wang, et al. benchmarks LLMs against traditional methods for numeric anomaly detection in power systems, showing their promise for energy grid monitoring. Furthermore, “LLM-Powered Text-Attributed Graph Anomaly Detection via Retrieval-Augmented Reasoning” by Haoyan Xu et al. from University of Southern California and Capital One introduces TAG-AD, a benchmark for text-attributed graphs that uses retrieval-augmented generation to enable zero-shot anomaly detection, proving LLMs excel at contextual anomalies while GNNs handle structural ones.

Under the Hood: Models, Datasets, & Benchmarks

Recent advancements are underpinned by robust new models, innovative architectures, and comprehensive benchmarks:

• AAR (Adaptive and Aggressive Rejection): Introduced in “Anomaly Detection with Adaptive and Aggressive Rejection for Contaminated Training Data”, this method dynamically estimates contamination and aggressively rejects outliers using statistical thresholds and Gaussian Mixture Models, improving AUROC by 0.041.
• I-GLIDE: Proposed in “I-GLIDE: Input Groups for Latent Health Indicators in Degradation Estimation” by Thil, L. et al. from NASA Ames Research Center and University of California, Berkeley. This framework uses uncertainty quantification and multi-head autoencoder architectures to enhance health indicators for degradation estimation, improving RUL prediction accuracy on datasets like NASA C-MAPSS and MILL NASA. Code available at https://github.com/LucasStill/I-GLIDE.
• ALADAEN: Featured in “Ranking-Enhanced Anomaly Detection Using Active Learning-Assisted Attention Adversarial Dual AutoEncoders” by Sidahmed Benabderrahmane et al. from NYU and University of Edinburgh. This framework detects Advanced Persistent Threats (APTs) using unsupervised anomaly detection with active learning and adversarial dual autoencoders, demonstrating superior results on DARPA Transparent Computing datasets. Code: https://gitlab.com/adaptdata.
• ADNet: Introduced by Hai Ling et al. from Communication University of China and UNSW Sydney in “ADNet: A Large-Scale and Extensible Multi-Domain Benchmark for Anomaly Detection Across 380 Real-World Categories”. This is the largest multi-domain anomaly detection dataset with 380 categories and over 196k images, providing standardized MVTec-style annotations. Visit: https://grainnet.github.io/ADNet.
• Pistachio: A novel benchmark for video anomaly detection (VAD) and understanding (VAU), introduced in “Pistachio: Towards Synthetic, Balanced, and Long-Form Video Anomaly Benchmarks” by Jie Li et al. from University of Science & Technology Beijing and Universiti Malaya, offering synthetic, balanced, and long-form videos with diverse scenes and anomaly types.
• VLM4TS & ViT4TS: In “Harnessing Vision-Language Models for Time Series Anomaly Detection” by Zelin He et al. from Pennsylvania State University and MIT. This two-stage framework uses VLMs (VLM4TS) and lightweight vision encoders (ViT4TS) for time series anomaly detection, achieving state-of-the-art F1-max scores while being 36× more efficient in token usage. Code: https://github.com/ZLHe0/VLM4TS.
• A2Seek: A large-scale, reasoning-centric benchmark for aerial anomaly understanding, presented in “A2Seek: Towards Reasoning-Centric Benchmark for Aerial Anomaly Understanding” by Mengjingcheng Mo et al. from Chongqing University of Posts and Telecommunications. It includes high-resolution videos and the A2Seek-R1 framework for enhancing multimodal foundation models. Code: https://2-mo.github.io/A2Seek/.
• VSAD (ViewSense-AD): From “Unsupervised Multi-View Visual Anomaly Detection via Progressive Homography-Guided Alignment” by Xintao Chen et al. from ShanghaiTech University. This framework performs unsupervised multi-view visual anomaly detection using a Multi-View Alignment Module (MVAM) and a Fusion Refiner Module (FRM) for robust cross-view feature consistency, achieving state-of-the-art results on RealIAD and MANTA datasets. Code: https://github.com/huggingface/diffusers.
• AutoGraphAD: Introduced in “AutoGraphAD: A novel approach using Variational Graph Autoencoders for anomalous network flow detection” by E. Caville et al. from University of Lorraine and ETH Zurich. This unsupervised method uses Variational Graph Autoencoders (VGAEs) to detect network attacks faster and more efficiently than traditional methods like Anomal-E. Code: https://github.com/Lightning-AI/lightning.
• ReCoVAD: Featured in “Sparse Reasoning is Enough: Biological-Inspired Framework for Video Anomaly Detection with Large Pre-trained Models” by He Huang et al. from Peking University. This biological-inspired VAD framework mimics the human nervous system’s reflex and conscious pathways, achieving state-of-the-art performance with 80% reduced computational cost by processing only a fraction of frames.
• CroTad: A contrastive reinforcement learning framework for online trajectory anomaly detection, introduced in “CroTad: A Contrastive Reinforcement Learning Framework for Online Trajectory Anomaly Detection” by Rui Xue from University of Technology Sydney and Uber Technologies Inc. It enables fine-grained anomaly localization and achieves up to 110.29% performance gains over existing methods.
• PersonaDrift: A new benchmark for temporal anomaly detection in language-based dementia monitoring, introduced in “PersonaDrift: A Benchmark for Temporal Anomaly Detection in Language-Based Dementia Monitoring”. This dataset addresses the need for robust models that can detect subtle shifts in patient behavior over time.
• AquaSentinel: A physics-informed AI system for urban underground water pipeline anomaly detection, presented in “AquaSentinel: Next-Generation AI System Integrating Sensor Networks for Urban Underground Water Pipeline Anomaly Detection via Collaborative MoE-LLM Agent Architecture” by Qiming Guo et al. from Texas A&M University – Corpus Christi. It combines sparse sensor deployment, the RTCA algorithm, Mixture of Experts (MoE) graph neural networks, and causal flow-based leak localization to achieve 100% detection accuracy. Code: https://github.com/VV123/STEPS.
• FKM-AD (Fourier-KAN-Mamba): Introduced in “Fourier-KAN-Mamba: A Novel State-Space Equation Approach for Time-Series Anomaly Detection” by Xiancheng Wang et al. from Harbin Institute of Technology. This model integrates Fourier KAN networks and Mamba structures with an energy anomaly score and a gated sharpening temperature mechanism for state-of-the-art time-series anomaly detection.
• GiBy: A “Giant-Step Baby-Step Classifier For Anomaly Detection In Industrial Control Systems” by Author 1 et al. from the University of Bristol in “GiBy: A Giant-Step Baby-Step Classifier For Anomaly Detection In Industrial Control Systems”. GiBy offers fast, near real-time anomaly detection with integrated explainability.
• LogPurge: A cost-aware, rule-enhanced framework for log data purification for anomaly detection, proposed in “LogPurge: Log Data Purification for Anomaly Detection via Rule-Enhanced Filtering” by Shenglin Zhang et al. from Nankai University. It uses LLM-enhanced filtering and a divide-and-conquer strategy to significantly improve F-1 scores for downstream models. It achieves 98.74% anomaly removal while retaining 82.39% of normal sequences.
• SAE-MCVT: A real-time and scalable multi-camera vehicle tracking framework, introduced in “SAE-MCVT: A Real-Time and Scalable Multi-Camera Vehicle Tracking Framework Powered by Edge Computing” by Yuqiang Lin et al. from the University of Bath. It includes the RoundaboutHD dataset, a comprehensive, high-resolution multi-camera vehicle tracking benchmark. Code: https://github.com/mikel-brostrom/boxmot.
• AGPNet (Attention-Guided Perturbation Network): Presented in “Not All Regions Are Equal: Attention-Guided Perturbation Network for Industrial Anomaly Detection” by Tingfeng Huang et al. from Huazhong Agricultural University. This reconstruction framework enhances industrial anomaly detection through attention-guided perturbations, achieving superior performance on MVTec-AD.

Impact & The Road Ahead

These advancements herald a new era for anomaly detection, moving beyond simple outlier identification to nuanced, context-aware, and often real-time reasoning. The integration of LLMs opens doors to more interpretable and adaptable systems, especially for contextual anomalies in domains like power systems and network security as seen in “From Topology to Behavioral Semantics: Enhancing BGP Security by Understanding BGP’s Language with LLMs”. The emphasis on explainability, as demonstrated by EVA-Net and Concept Bottleneck Models, will build greater trust in AI-driven decisions, crucial for sensitive applications like healthcare and critical infrastructure.

The drive towards energy-efficient and lightweight solutions, exemplified by “Lightweight Autoencoder-Isolation Forest Anomaly Detection for Green IoT Edge Gateways” and “Energy-Aware Pattern Disentanglement: A Generalizable Pattern Assisted Architecture for Multi-task Time Series Analysis” will enable broader deployment in resource-constrained environments, from IoT edge devices to scalable industrial control systems. New benchmarks like ADNet, Pistachio, and A2Seek are vital for fostering robust, generalizable models that can handle the complexity of real-world multi-domain data and diverse anomaly types.

Beyond current frontiers, the field is even looking into quantum computing, with “Neural Architecture Search for Quantum Autoencoders” exploring how Neural Architecture Search can optimize quantum circuits for data compression, potentially revolutionizing future quantum anomaly detection. The paradigm shift highlighted by “Labels Matter More Than Models: Quantifying the Benefit of Supervised Time Series Anomaly Detection” also reminds us that while models advance, data quality and clever use of even limited labels can yield profound improvements. The future of anomaly detection is bright, promising more intelligent, resilient, and insightful systems across every aspect of our technologically driven world.