DeCoFlow: Structural Decomposition of Normalizing Flows for Continual Anomaly Detection

Overview

DeCoFlow is a framework designed for continual anomaly detection in industrial environments where new product categories arrive sequentially. Standard Normalizing Flows (NFs) are powerful for density estimation but struggle with continual learning because parameter updates across tasks distort the density manifold, leading to catastrophic forgetting. DeCoFlow addresses this by exploiting the structural independence of affine coupling layers, allowing for parameter isolation without compromising the invertibility or Jacobian validity required by flow-based models.

The core innovation is the Decomposed Coupling Layer (DCL), which partitions each subnet into a frozen universal base and task-specific low-rank adapters. By freezing the base after the first task, the model prevents interference between sequential tasks. To overcome the rigidity of a frozen base, the authors introduce three compensatory mechanisms:

• Task-Specific Alignment (TSA): Standardizes input features using task-specific affine transformations to mitigate distribution shifts.
• Auxiliary Coupling Layers (ACL): Provides additional residual density correction to bridge statistical gaps that low-rank adapters might miss.
• Tail-Aware Loss (TAL): A specialized loss function that reweights high-loss patches to ensure the model focuses on under-represented tail distributions, which are critical for accurate anomaly detection.

DeCoFlow achieves state-of-the-art performance on standard benchmarks (MVTec-AD and VisA) while maintaining zero parameter-level forgetting. By using a prototype-based routing mechanism, it enables task-agnostic inference in a single forward pass, making it highly efficient for real-world manufacturing lines where memory and privacy constraints prevent the storage of past data. This approach demonstrates that structural decomposition is a viable strategy for adapting density-based models to evolving data distributions.