Graph Similarity-Aware Collaborative Federated Knowledge Distillation for Anomaly Detection in Lights-Out Manufacturing
Year: 2026
Type: conference-paper
Abstract: The heterogeneity of multimodal data in Lights-out Manufacturing (LoM) poses a significant challenge to anomaly detection based on Federated Knowledge Distillation (FedKD) due to the substantial differences in graph structure. Existing FedKD methods treat all clients equally, ignoring graph similarity and leading to negative transfer; existing graph similarity calculations either focus only on the global topology or only on the local distribution, making it difficult to comprehensively evaluate; existing collaborative methods have high communication overhead or poor model adaptability. To address this, this paper proposes a graph similarity-aware FedKD collaborative anomaly detection framework. First, a dual-strategy graph similarity metric is designed: Laplace spectral analysis is used to capture the global topology, and Kullback- Leibler Divergence (KLD) is used to quantify the local distribution of nodes. Second, the two similarity metrics are adaptively fused using a dynamic gating network. Finally, based on similarity filtering, knowledge collaboration is achieved only between structurally similar client-side mentor models (Teacher) through soft label sharing. Experiments on three real-world industrial datasets-SWaT, WADI, and MVTec AD-show that compared to the uncoordinated baseline, our method improves the F1 score and accuracy by$\mathbf{0.28} \boldsymbol{\%}$and$\mathbf{2.73} \boldsymbol{\%}$, respectively, with an AUC of 0.91 (SWaT) and 0.88 (WADI), outperforming several existing coordinated methods. Ablation experiments validated the effectiveness of each component.
Authors: Manshan Mo, Weijian Zhong, Zhongwei Huang
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204209914