Work

ShortcutBreaker: Low-Rank Noisy Bottleneck and Frequency Filtering Block for Multi-Class Unsupervised Anomaly Detection

Year: 2026
Type: conference-paper
Abstract: Multi-class Unsupervised Anomaly Detection (MUAD) has attracted increasing research interest for its ability to develop a unified model across multiple classes, significantly reducing the computational cost of training separate models for distinct objects. However, despite the performance gains offered by advanced Transformer-based architectures, the identity shortcut issue persists. These models tend to directly copy inputs to outputs, narrowing the reconstruction error gap between normal and abnormal cases and hindering distinguishability. To address this, we propose ShortcutBreaker, a framework designed to disrupt such shortcuts through two core innovations. First, we introduce a Low-Rank Noisy Bottleneck (LRNB) derived from matrix rank inequality, which theoretically prevents trivial identity mapping by perturbing features across two distinct reconstruction groups. Second, we design a Frequency-Filter Unit (FFU) that integrates frequency spectral filtering and global-local masking. This mechanism further enlarges the anomaly score gap, with less computational cost than standard self-attention. Extensive evaluations on four benchmarks demonstrate the superiority of ShortcutBreaker.
Language: English
Open Access status: hybrid
OpenAlex ID: W4415967321
FWCI: 1.281
Cites: 40
Cited by: 2
Related to: 10