variational anomaly detection

In today's digital world, the generation of vast amounts of streaming data in various domains has become ubiquitous. However, many of these data are unlabeled, making it challenging to identify events, particularly anomalies. This task becomes even more formidable in nonstationary environments where model performance can deteriorate over time due to concept drift. To address these challenges, this paper presents a novel method, VAE++ESDD, which employs incremental learning and two-level ensembling: an ensemble of Variational AutoEncoder(VAEs) for anomaly prediction, along with an ensemble of concept drift detectors. Each drift detector utilizes a statistical-based concept drift mechanism. To evaluate the effectiveness of VAE++ESDD, we conduct a comprehensive experimental study using real-world and synthetic datasets characterized by severely or extremely low anomalous rates and various drift characteristics. Our study reveals that the proposed method significantly outperforms both strong baselines and state-of-the-art methods.

To address insufficient sensitivity to unseen anomalies in unsupervised anomaly detection, this paper proposes a two-stage light-supervision enhancement framework. First, a variational autoencoder (VAE) is pretrained unsupervisedly on normal data; second, it undergoes supervised fine-tuning using only a small number of known anomaly samples, guided by reconstruction error minimization. Crucially, this work introduces minimal anomaly supervision—requiring labels *only* for anomalies, with no normal-class labels—directly into the VAE reconstruction objective, thereby significantly improving generalization to previously unseen anomalies. Extensive evaluation across four heterogeneous benchmarks—MNIST, CICIDS, LHCO2020, and SMEFT—demonstrates superior normal/anomalous separation. Notably, on the Higgs event momentum-shift detection task, sensitivity improves markedly, validating the efficacy of few-shot anomaly supervision in enhancing unsupervised models. The approach is particularly suited for high-precision, sensitivity-critical domains such as particle physics detection and cybersecurity.

Addressing the challenge of online anomaly detection in automotive testing—characterized by multivariate, state-varying time-series data, scarcity of labeled samples, stringent requirements for low false positive rate (FPR), real-time responsiveness, and root-cause interpretability—this paper proposes the Temporal Variational Autoencoder (TeVAE). TeVAE mitigates the “bypass” problem in standard VAEs via a tailored architectural design; introduces a novel sliding-window-to-continuous-time remapping mechanism that decouples anomaly localization from detection latency; and defines a new evaluation metric jointly optimizing detection delay and root-cause localization accuracy. The method is fully unsupervised and supports few-shot training. Evaluated on real-world industrial datasets, TeVAE achieves a 6% FPR and 65% detection rate—substantially outperforming established baselines—demonstrating strong engineering applicability and promising capability for root-cause analysis.

This work addresses the strong reliance on manual hyperparameter tuning and poor adaptability to novel dynamic anomalies in time-series anomaly detection. We propose the first deep reinforcement learning (DRL)-driven collaborative framework integrating variational autoencoders (VAEs) and active learning. Methodologically, the framework jointly leverages LSTM networks to capture temporal dependencies, VAEs to model latent data distributions, DRL to optimize sample selection policies, and active learning to prioritize high-information unlabeled instances—enabling adaptive discovery and detection of previously unseen anomalies under low-labeling budgets. Extensive experiments on multiple real-world datasets demonstrate significant improvements in F1-score and recall; notably, detection accuracy increases by over 23% under constrained annotation budgets. These results validate the framework’s superior generalization capability and practical applicability for real-world time-series monitoring scenarios.

To address low detection accuracy and poor robustness in anomaly detection under non-i.i.d. heterogeneous data, this paper proposes Multi-Input Variational Autoencoders (MIVAE) and Multi-Input Autoencoders for Anomaly Detection (MIAEAD). The method introduces three key innovations: (1) a novel parallel sub-encoder architecture that independently models distinct feature subsets; (2) a theoretical proof demonstrating superior anomaly discriminability compared to standard VAEs; and (3) the first heterogeneity-adaptive assessment mechanism based on the coefficient of variation (CV). By jointly leveraging latent-space regularization and reconstruction-error-driven anomaly scoring—enhanced with an AUC-maximization strategy—the approach achieves an average 6% AUC improvement over state-of-the-art unsupervised methods across eight real-world datasets, with particularly notable gains on low-heterogeneity subsets.

This work addresses the challenge of integrating variational autoencoders (VAEs) as trainable layers within neural networks. It proposes a general framework for flexibly embedding VAEs into arbitrary network architectures, accompanied by an end-to-end training strategy that leverages the reparameterization trick and probabilistic modeling to ensure full differentiability throughout the pipeline. For the first time, this approach enables VAEs to function as plug-and-play modules akin to standard neural network layers, substantially enhancing their compatibility and representational capacity within complex models. Experimental results demonstrate that the proposed VAE layer consistently achieves stable performance across diverse tasks and outperforms conventional standalone VAE models, thereby significantly expanding the applicability of VAEs in deep learning systems.

This work addresses the degradation of unsupervised and out-of-distribution (OOD) anomaly detection performance in high-dimensional variational autoencoders (VAEs), which arises from the exponential growth of latent space hypervolume and the concentration of latent variables near the equatorial region of the hypersphere. To mitigate this issue, the study introduces hyperspherical coordinates into the modeling of the VAE latent space for the first time, constraining latent vectors to align with specific directions on the hypersphere. This yields a more expressive approximate posterior distribution that effectively alleviates high-dimensional degeneracy. The proposed method demonstrates significantly improved anomaly detection sensitivity across diverse real-world datasets—including Martian terrain and galaxy images—as well as standard benchmarks such as CIFAR-10 and ImageNet subsets, achieving state-of-the-art performance.

This work addresses the challenge of anomaly detection in unlabeled network traffic within IT/OT convergence scenarios by proposing an unsupervised approach based on β-variational autoencoders (β-VAE). It presents the first systematic comparison between two detection mechanisms: latent space distance and reconstruction error. Experimental evaluation on the NSL-KDD dataset demonstrates that measuring the distance from test samples to the training data distribution in the latent space significantly outperforms conventional reconstruction error–based methods, yielding notably higher detection accuracy under fully unsupervised conditions. The findings highlight the critical advantage of leveraging latent space structure for unsupervised network anomaly detection and offer a novel direction for future research in this domain.

This work proposes a novel approach to time series anomaly detection that addresses the limitations of traditional observation-likelihood-based methods, which often fail to capture structured temporal dynamics and misclassify anomalies as normal patterns. By introducing inductive biases into the latent space of conditional normalizing flows, the method models time series as discrete-time state-space systems, enforcing latent trajectories to conform to prescribed dynamical laws. Anomalies are then defined as deviations from these expected dynamics. The approach frames anomaly detection as a goodness-of-fit test for dynamic consistency—a formulation introduced here for the first time—and evaluates compliance of latent trajectories accordingly. Experiments on both synthetic and real-world datasets demonstrate its effectiveness in detecting anomalies in frequency, amplitude, and noise characteristics, achieving high detection performance alongside strong interpretability.

This work addresses the challenge of anomaly detection in multivariate time series where anomalies are unlabeled, extremely rare, yet highly costly. The authors propose a self-supervised method that explicitly shapes the geometry of the embedding space through a novel “velocity-consistency” objective, which enforces local smoothness and directional coherence of normal trajectories without requiring negative samples or synthetic anomalies. The approach integrates a channel-aware encoder with two complementary scoring mechanisms: a Mahalanobis-distance-based positional score and a velocity-bank directional score, enabling precise discrimination of anomalous instances. Evaluated on the TSB-AD-M benchmark, the method achieves state-of-the-art performance, significantly outperforming existing approaches that are often more complex and computationally expensive.