Privacy-Preserving SACAA-Net: A Scale-Aware Consistency and Anomaly-Attentive Network for Fine-Grained Industrial Anomaly Detection in Secure Distributed Edge-IIoT Systems

Abstract

Fine-grained industrial anomaly detection has become a core component of intelligent inspection in Industrial Internet of Things (IIoT) scenarios, particularly when visual data are continuously acquired and processed through distributed sensors together with edge devices. Yet once industrial visual data and model updates are distributed in this way, a series of security risks becomes difficult to avoid, including data leakage, unauthorized access, gradient inversion attacks, model poisoning, and privacy inference. What may be exposed through these risks is not only raw data itself, but also sensitive information related to equipment structure, production status, and enterprise operations. Under such conditions, privacy preservation is no longer optional; it is an essential requirement for real-world industrial deployment. Existing methods still concentrate mainly on representation learning under heterogeneous data distributions, while security constraints in distributed environments are often insufficiently considered. This can easily result in inconsistent feature representations and inadequate protection of privacy-sensitive information during both training and transmission. To address these issues, we propose Privacy-Preserving SACAA-Net for fine-grained industrial anomaly detection in secure distributed IIoT systems. The proposed model brings together cross-scale semantic consistency learning, anomaly-aware attention, and structural dependency modeling, with the goal of improving detection accuracy while retaining fine-grained details and overall structural information. For privacy-preserving deployment, SACAA-Net supports local training on edge devices, reduces the need for raw data sharing, and remains compatible with secure aggregation, differential privacy, and encrypted transmission, thereby helping mitigate the risks of information leakage and model inversion attacks.Extensive experiments on six industrial datasets demonstrate that SACAA-Net achieves superior performance in both accuracy and robustness, achieving average image-level and pixel-level AUROC scores of 96.4\% and 92.3\%, respectively, and outperforming the strongest baseline by 1.5\% and 1.8\% on these two metrics, providing an effective and secure solution for distributed industrial anomaly detection.