Tianhao Niu, Ziyu Han, Qiguang Chen, Shiqi Zhou, Baocai Shan, Hengjie Fang, Qingfu Zhu, Wanxiang Che. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

While the interpretability of deep learning (DL)-based models has been extensively explored in academia, applying existing interpretation methods to anomaly detection in industrial control systems (ICSs) poses challenges for two primary reasons. First, security experts in ICS have distinct interpretive priorities, emphasizing the need for stability and readability. Second, there are various types of device components in ICS, and the potential interactions between sensors and actuators are yet to be explored. To tackle the above challenges, we propose DeepINT, an interpreter for anomaly detection in ICS. In DeepINT, we adopt a search optimization algorithm to find the reference and capture feature importance by the backpropagation gradient to improve interpretation performance and reliability. In addition, we construct a finite difference-based interaction detection, which tests the interaction of different device components, in order to address the problem that actuators in ICS are not easily interpreted, meanwhile improving the comprehensiveness and accuracy of the interpretation results. In comprehensive experiments on two real water treatment datasets [secure water treatment (SWaT) and water distribution (WADI)], DeepINT shows excellent interpretation performance compared to the six state-of-the-art baseline methods, especially on the SWaT dataset, with a 60% improvement in interpretation accuracy. In addition, our method significantly improves the efficiency of interaction detection, which balances interpretation performance and time efficiency.

The structural complexity, material diversity, and defect concealment in industrial detection scenes pose challenges of robustness, multi-information, and effectiveness to optical imaging systems. Partially blurred images due to the limited depth of field (DoF) of industrial imaging systems, shadow occlusions due to simple illumination conditions, and material and texture interference due to multiple compositions have become key issues affecting imaging quality in complex scenes. This paper proposes a systematic scheme fusing the DoF expansion approach, light source optimization, and polarization information (DLP-Fusion) to comprehensively improve imaging quality. Herein, a DoF fusion algorithm and a liquid zoom lens are used to increase the DoF from 2.5 mm to 40 mm. Moreover, a combination of ring light and freely rotatable strip light sources is introduced to improve the uniformity and robustness of the illumination, resulting in an average enhancement of 56.46% in the contrast of the target features. Furthermore, a polarization selection fusion network (PSFNet) is constructed to achieve flare suppression and complex material characterization, with the image naturalness improving by 32.05%. The experimental results with diverse scenes demonstrate that DLP-Fusion considerably improves the DoF range, image uniformity, and target feature contrast. DLP-Fusion exhibits remarkable robustness in various environments and was seamlessly deployed in real-world industrial settings with good performance. This paradigm may open a path toward intelligent imaging systems for sophisticated applications, including multimaterial detection and target recognition under harsh conditions.

Anomaly detection plays a vital role as a crucial security measure for edge devices in Artificial Intelligence and Internet of Things (AIoT). With the rapid development of IoT ( Internet of Things), changes in system configurations and the introduction of new devices can lead to significant alterations in device relationships and data flows within the IoT, thereby triggering concept drift. Previously trained anomaly detection models fail to adapt to the changed distribution of streaming data, resulting in a high number of false positive events. This paper aims to address the issue of concept drift in IoT anomaly detection by proposing a comprehensive Concept Drift Detection, Interpretation, and Adaptation framework (CDDIA). We focus on accurately capturing the concept drift of normal data in unsupervised scenarios. To interpret drift samples, we integrate a search optimization algorithm and the SHAP method, providing a comprehensive interpretation of drift samples at both the sample and feature levels. Simultaneously, by utilizing the sample-level interpretation results for filtering new and old samples, we retrain the anomaly detection model to mitigate the impact of concept drift and reduce the false positive rate. This integrated strategy demonstrates significant advantages in maintaining model stability and reliability. The experimental results indicate that our method outperforms five baseline methods in adaptability across three datasets and provides interpretability for samples experiencing concept drift.