Diffuse to Detect: A Generalizable Framework for Anomaly Detection with Diffusion Models – Applications to UAVs and Beyond
Mingze Gong, Juan Du, Jianbang You · 2025
Read Diffuse to Detect: A Generalizable Framework for Anomaly Detection with Diffusion Models – Applications to UAVs and Beyond by Mingze Gong, Juan Du, Jianbang You (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-22928.txt.
Introduction
Mingze Gong, Juan Du, and Jianbang You address the hard task of spotting faults in complex sensor data, such as readings from drones. Odd readings can warn of danger, but older methods miss subtle faults and scale poorly. Their framework, called Diffuse to Detect, borrows diffusion models in a new way.
Rather than rebuild data slowly, the method predicts noise in one step to flag values that do not fit. A graph network links sensors that affect each other, which sharpens the search. The authors link the method to the shape of the data itself and test it on drone data and beyond.
Catalog record
| CREATOR | Mingze Gong, Juan Du, Jianbang You |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 79 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-22928
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Diffuse to Detect: A Generalizable Framework for Anomaly Detection with Diffusion Models – Applications to UAVs and Beyond Mingze Gong, Juan Du, Jianbang You Abstract Anomaly detection in complex, high-dimensional data, such as UAV sensor readings, is essential for operational safety but challenging for existing methods due to their limited sensitivity, scalability, and inability to capture intricate dependencies. We propose the Diffuse to Detect (DTD) framework, a novel approach that innovatively adapts diffusion models for anomaly detection, diverging from their conventional use in generative tasks with high inference time. By comparison, DTD employs a single-step diffusion process to predict noise patterns, enabling rapid and precise identification of anomalies without reconstruction errors. This approach is grounded in robust theoretical foundations that link noise prediction to the data distribution’s score function, ensuring reliable deviation detection. By integrating Graph Neural Networks to model sensor relationships as dynamic graphs, DTD effectively captures spatial (inter-sensor) and temporal anomalies. Its two-branch architecture, with parametric neural network-based energy scoring for scalability and nonparametric statistical methods for interpretability, provides flexible trade-offs between computational efficiency and transparency. Extensive evaluations on UAV sensor data, multivariate time series, and images demonstrate DTD’s superior performance over existing methods, underscoring its generality across diverse data modalities. This versatility, combined with its adaptability, positions DTD as a transformative solution for safety-critical applications, including industrial monitoring and beyond. \RRHFirstLine \RRHSecondLine \LRHFirstLine \LRHSecondLine keywords Anomaly Detection, Diffusion Models, Unmanned Aerial Vehicles (UAVs), H igh-dimensional Sensor Data, Graph Neural Networks † † runningauthor: Mingze, Juan and Jianbang † † runningtitle: Diffuse to Detect: A Generalizable Framework for Anomaly Detection with Diffusion Models – Applications to UAVs and Beyond † † authors: The Hong Kong University of Science and Technology (Guangzhou), [email protected] The Hong Kong University of Science and Technology (Guangzhou), The Hong Kong University of Science and Technology, [email protected] The Hong Kong University of Science and Technology (Guangzhou), The Hong Kong Polytechnic University, [email protected] † † affiliation: † † affiliation: † † affiliation: † † affiliation: † † affiliation: † † affiliation: 1 Introduction Unmanned Aerial Vehicles (UAVs) have become indispensable across diverse applications, including aerial surveillance, package delivery, disaster response, and environmental monitoring, due to their flexibility, cost-effectiveness, and ability to operate in challenging environments ( Shakhatreh et al. 2019 ) . As UAV deployment grows, ensuring operational reliability and safety becomes paramount, especially in high-stakes contexts like urban air mobility and infrastructure inspection. A key aspect of this is monitoring sensor data (encompassing metrics such as altitude, speed, battery levels, and engine performance) to gain insights into a UAV’s operational health. High-dimensional anomaly detection is essential for identifying critical issues, such as sudden engine failures, security breaches, or environmental disruptions, could compromise mission success or lead to accidents ( Yang et al. 2024 ) . UAV sensor data, however, poses significant challenges for effective anomaly detection. These data are high-dimensional, collected by numerous sensors with high sampling rates, and exhibit complex multivariate dependencies that shift dynamically based on flight conditions ( Huang et al. 2025 ) . Anomalies, such as engine failures, often manifest as abrupt change in sensor readings (e.g., simultaneous drops in power and altitude). Furthermore, UAVs produce diverse data modalities, including structured time series, graphs of sensor interactions, and occasionally images, necessitating detection frameworks capable of handling heterogeneous inputs ( Wang et al. 2019 ) . Conventional statistical and deep learning methods struggle to address these complexities, underscoring the need for tailored approaches to achieve rapid and reliable anomaly detection for UAV safety. 1.1 Limitations of Existing Methods Current anomaly detection techniques, spanning statistical and deep learning paradigms, exhibit significant limitations when tasked with detecting anomalies in UAV sensor data. These limitations include: • Difficulty Detecting Subtle and Abrupt Anomalies in High-Dimensional Data : Methods like autoencoders or statistical tests often fail to detect anomalies in high-dimensional UAV data, where noise and complex patterns can obscure critical deviations. For instance, abrupt changes following an engine failure may resemble normal pattern locally, leading to low reconstruction errors or insufficient sensitivity, resulting in detection failures ( Huang et al. 2025 ) . • Inadequate Modeling of Complex Sensor Dependencies : UAVs produce multivariate time series with intricate inter-sensor relationships. Traditional methods often treat sensors independently or use simplistic correlation models, resulting in detection failures of anomalies arising from coordinated deviations, such as simultaneous drops in power and altitude post-engine failure, which are critical for accurate detection ( Lin et al. 2010 ) . • Limited Use and Adaptation of Generative Models for Structured Data : Generative models such as GANs and VAEs are predominantly applied to image-based anomaly detection, with limited use for time series or other structed data, as their high computational cost hinders anomaly detection in UAV operational monitoring, where such data is commonly collected. Their reliance on reconstruction errors often fails to capture complex anomalies, and advanced generative models are rarely adapted for real-world engineering challenges involving noisy UAV data ( Liu et al. 2023 ) . • Trade-offs Between Scalability and Interpretability : Neural network-based approaches scale well for large datasets but lack transparency, obscuring why anomalies are flagged. Conversely, interpretable methods like Kernel Density Estimation (KDE) struggle with high-dimensional data, hindering real-time detection in UAV monitoring. This trade-off complicates diagnosis in safety-critical scenarios ( Hu et al. 2020 ) . • Limited Generalization Across Diverse Data Modalities : Most anomaly detection frameworks are designed for specific data types, such as time series or images, and struggle to generalize to others, such as graphs, without significant adjustments. This specialization hinders their application in real-world UAV systems, where heterogeneous data (e.g., sensor readings, images, network graphs) are collected, requiring multiple models and complicating deployment for comprehensive monitoring ( Wang et al. 2019 ) . These challenges are intensified by the curse of dimensionality, where high-dimensional data becomes sparse because the data points grow increasingly spread out and isolated across an exponentially expanding volume of space, undermining distance-based or density-based detection methods ( Suboh et al. 2023 ) . Moreover, the rarity and diversity of anomalies in UAV operations, coupled with the scarcity of labeled data, necessitate unsupervised anomaly detection approaches ( Lin et al. 2024 ) . However, developing robust unsupervised methods that can effectively handle the high-dimensional and noisy UAV sensor data remains a significant challenge. Together, these limitations highlight the inadequacy of existing approaches for rapidly and reliably detecting critical anomalies in UAV operations without labeled data. 1.2 Contributions of DTD The identified limitations underscore the need for an anomaly detection framework that is sensitive, scalable, interpretable, and adaptable to the complex, multimodal data from UAVs. Diffusion models (DMs) predict noise patterns in a manner aligned with detecting anomalous deviations in UAV data, with a detailed theoretical derivation provided later. Inspired by recent advancements in generative modeling for anomaly detection ( Livernoche et al. 2023 , Wang et al. 2023 ) , we propose the Diffuse to Detect (DTD) framework, a customized generative approach for rapid and effective anomaly detection. Unlike conventional methods relying on reconstruction errors to identify anomalies, DTD directly evaluates data deviations using the learned patterns from diffusion models, enhancing efficiency and supported by theoretical validation that validates its efficacy for anomaly detection. It provides two