Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Diffuse to Detect: A Generalizable Framework for Anomaly Detection with Diffusion Models Applications to UAVs and Beyond

Mingze Gong, Juan Du, Jianbang You

TL;DR

Diffuse to Detect (DTD) tackles anomaly detection in high-dimensional, multimodal UAV data by using a single-step diffusion-based noise predictor tied to the data score function. It combines dynamic sensor graphs, a two-branch scoring system (parametric DM-P with an Energy-Based Model and nonparametric DM-NP), and EVT-based labeling to achieve fast, interpretable, and accurate detection across time series, graphs, and images. Empirical results on ALFA, BASiC, SMD, and CIFAR-10 show superior performance over strong baselines, with real-time inference enabled by the one-step diffusion strategy. The work broadens diffusion-model applications to safety-critical, multimodal anomaly detection and suggests wide applicability in industrial monitoring and beyond.

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.

Diffuse to Detect: A Generalizable Framework for Anomaly Detection with Diffusion Models Applications to UAVs and Beyond

TL;DR

Diffuse to Detect (DTD) tackles anomaly detection in high-dimensional, multimodal UAV data by using a single-step diffusion-based noise predictor tied to the data score function. It combines dynamic sensor graphs, a two-branch scoring system (parametric DM-P with an Energy-Based Model and nonparametric DM-NP), and EVT-based labeling to achieve fast, interpretable, and accurate detection across time series, graphs, and images. Empirical results on ALFA, BASiC, SMD, and CIFAR-10 show superior performance over strong baselines, with real-time inference enabled by the one-step diffusion strategy. The work broadens diffusion-model applications to safety-critical, multimodal anomaly detection and suggests wide applicability in industrial monitoring and beyond.

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.
Paper Structure (37 sections, 2 theorems, 40 equations, 7 figures, 6 tables, 2 algorithms)

This paper contains 37 sections, 2 theorems, 40 equations, 7 figures, 6 tables, 2 algorithms.

Table of Contents

  1. Limitations of Existing Methods
  2. Contributions of DTD
  3. Related Works
  4. Statistical Methods for UAV Anomaly Detection
  5. Deep Learning Methods for UAV Anomaly Detection
  6. Methodology
  7. Preliminaries
  8. Overview of the DTD Framework
  9. Diffusion Model Formulation
  10. Noise-Based Anomaly Detection
  11. Scoring Methodology
  12. Nonparametric Scoring
  13. Parametric EBM Scoring
  14. Anomaly Labeling with EVT
  15. Experimental Setup
  16. ...and 22 more sections

Key Result

Proposition 1

For a diffusion model with forward process $x_k=\sqrt{\bar{\alpha}_k}\,x_0+\sqrt{1-\bar{\alpha}_k}\,\epsilon$, $\epsilon\sim\mathcal{N}(0,I)$, the noise predictor satisfies where $p_k(x_k)$ is the distribution of $x_k$. When $k$ is small, such as $k=1$, $\epsilon_\theta(x_0, 1, x_{hist}) \approx -\sqrt{1 - \bar{\alpha}_0} \nabla_{x_0} \log p_{\text{data}}(x_0 | x_{hist})$ encodes the local geomet

Figures (7)

  • Figure 1: An Overview of the Framework. The Diffuse to Detect (DTD) framework first transforms UAV sensor data into a graph structure, where nodes represent sensors and edges initialized first and learnt through stochastic gradient descent. The framework then trains a diffusion model to learn the anomaly-free data distribution, enabling the prediction of noise patterns. For anomaly detection, the framework applies a single diffusion step to a test sample, generating a perturbed sample and predicting its noise. Two scoring branches (parametric and nonparametric) evaluate the predicted noise against the learned distribution, providing robust anomaly scores. Finally, the framework uses Extreme Value Theory (EVT) to compare scores against a threshold, enabling the detection of anomalies.
  • Figure 2: Anomaly scores and predictions for ALFA dataset. Left column (a–b): scores; right column (c-d): predictions. Ground truth (fault) regions are shaded in red.
  • Figure 3: Visualization of UAV nodes connection in the DM-P (KDE) for graph-structured ALFA sensor data.
  • Figure 4: 3D surface plots of DM outputs, diffused data and normalized energy values for both branches on ALFA dataset: (a) DM-NP (KDE) and (b) DM-P. The x-axis represents noise prediction outputs, the y-axis represents diffused values (0 to 1), and the z-axis represents energy outputs.
  • Figure 5: BASiC: anomaly scores (left) and predictions (right). Ground-truth fault regions are shaded in red.
  • ...and 2 more figures

Theorems & Definitions (2)

  • Proposition 1
  • Proposition 2