Hierarchical Testing with Rabbit Optimization for Industrial Cyber-Physical Systems
Jinwei Hu, Zezhi Tang, Xin Jin, Benyuan Zhang, Yi Dong, Xiaowei Huang · 2025
Read Hierarchical Testing with Rabbit Optimization for Industrial Cyber-Physical Systems by Jinwei Hu, Zezhi Tang, Xin Jin, Benyuan Zhang, Yi Dong, Xiaowei Huang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-04100.txt.
Introduction
Jinwei Hu, Zezhi Tang, Xin Jin, Benyuan Zhang, Yi Dong, and Xiaowei Huang published a paper in 2025 on attacking the health-monitoring brains of industrial systems to find their weak points. Factories increasingly trust deep learning models to predict failures in coupled physical and digital systems. If small, realistic data tweaks can fool those models, maintenance decisions become dangerous.
Their framework, HERO, tests such models in layers, using an optimization method modeled on rabbit foraging to craft deceptive inputs. The fake data respects physical limits and matches how real readings spread. The testbed is a hydrogen fuel cell system with fast-changing conditions and complex wear. Even top monitoring models showed vulnerabilities under the crafted attacks, a warning the authors extend to resilient design across industrial settings.
Catalog record
| CREATOR | Jinwei Hu, Zezhi Tang, Xin Jin, Benyuan Zhang, Yi Dong, Xiaowei Huang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-04100
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Hierarchical Testing with Rabbit Optimization for Industrial Cyber-Physical Systems Jinwei Hu, Zezhi Tang, Xin Jin, Benyuan Zhang, Yi Dong, Xiaowei Huang Abstract This paper presents HERO (Hierarchical Testing with Rabbit Optimization), a novel black-box adversarial testing framework for evaluating the robustness of deep learning-based Prognostics and Health Management systems in Industrial Cyber-Physical Systems. Leveraging Artificial Rabbit Optimization, HERO generates physically constrained adversarial examples that align with real-world data distributions via global and local perspective. Its generalizability ensures applicability across diverse ICPS scenarios. This study specifically focuses on the Proton Exchange Membrane Fuel Cell system, chosen for its highly dynamic operational conditions, complex degradation mechanisms, and increasing integration into ICPS as a sustainable and efficient energy solution. Experimental results highlight HERO’s ability to uncover vulnerabilities in even state-of-the-art PHM models, underscoring the critical need for enhanced robustness in real-world applications. By addressing these challenges, HERO demonstrates its potential to advance more resilient PHM systems across a wide range of ICPS domains. I Introduction With the rapid development of net zero, there is a need for advanced predictive models and system integration plays a crucial role in the field of renewable energy technologies, particularly in the deployment and management of Proton Exchange Membrane Fuel Cells (PEMFC). Regarded as an integral part of future energy conversion technologies, PEMFC boast high energy conversion efficiency, low operating temperature, low emissions, and rapid startup capabilities [ 1 ] . Their application range, from portable power sources to transportation and stationary power generation systems, displays their exceptional flexibility and environmental friendliness, indicating their pivotal role in future energy systems [ 2 ] . However, the widespread deployment of PEMFC across various sectors brings significant challenges in ensuring their long-term stable operation and accurately approximating their lifespan. This highlights the critical role of Prognostics and Health Management (PHM) strategies, which employ predictive models to forecast system degradation and estimate Remaining Useful Life (RUL), thereby enabling preventive maintenance in critical industrial applications [ 3 ] . Machine Learning (ML) and Deep Learning (DL) models are always employed in enhancing the accuracy and robustness of detections in PHM systems, capitalizing on their ability to discern complex data patterns from extensive datasets [ 4 ] . Despite the significant advantages these models provide in refining predictions and operational efficiencies, they inherently exhibit vulnerabilities that can undermine their effectiveness in critical applications such as PEMFC lifespan predictions. One significant challenge is the “black-box” nature of many deep learning algorithms. This complexity complicates the interpretability and validation of the models, raising concerns in critical applications like PEMFC operation monitoring [ 5 , 6 ] . While ML and DL models always perform well under typical conditions in ICPS, they often lack robustness against diverse threats, such as cyber attacks [ 7 ] and adversarial attacks [ 8 ] . For example, subtle, intentionally crafted disturbances in PHM system inputs can cause significant prediction errors for PEMFC health and performance, potentially leading to undetected system failures or suboptimal operations until severe issues arise [ 9 , 10 ] . Adversarial testing, a technique widely applied across various fields, is focusing on identifying performance-degrading examples that can impact DL models. Identified adversarial examples can also facilitate system improvements through adaptation or retraining [ 11 , 12 ] . Mainstream white-box testing methods, such as DLFuzz [ 13 ] and DeepXplore [ 14 ] , predominantly employ coverage-guided testing to actively mutate inputs, maximizing neuron coverage and exposing prediction discrepancies. However, these methods are primarily designed for the image domain and do not cater to the nuances of time-series DL models. Transitioning to time-series DL models, TESTRNN [ 15 ] is also a coverage-guided testing approach that designs metrics to assess both values and temporal relationships in LSTM inputs. Despite its innovative approach, TESTRNN primarily focuses on image data and is not suitable for the tabular data typically associated with PHM systems. Moreover, because of its model-specific nature, coverage-guided testing does not adapt well to applications involving transformer-based DL tools. In black-box adversarial testing, the Hierarchical Distribution-Aware (HDA) [ 16 ] approach employs a dual-distribution strategy to capture global and local data characteristics. Despite being originally developed for image tasks, the theoretical foundation and inherent nature of black-box testing enable the HDA approach to transcend its initial scope, making it versatile across various domains, including transformer-based PHM systems involving time-series data. However, its reliance on Variational Autoencoders for latent space approximations fails to account for the temporal dependencies and dynamic variability crucial to time-series data in ICPS, limiting its ability to generate representative adversarial examples. Furthermore, HDA lacks efficient optimization tailored to the time-series domain and does not incorporate physical constraints, reducing its effectiveness in producing realistic and operationally plausible adversarial examples for industrial applications. Therefore, while previous methods are often effective in their specific fields, they are generally time-consuming and frequently overlook the stringent physical constraints associated with industrial time-series data. This limitation leads to low sampling efficiency, as seen with approaches like genetic algorithms, which are inadequate for applications such as PHM systems for PEMFC. To address this, we incorporate the state-of-the-art Artificial Rabbits Optimization (ARO) algorithm to meet the unique requirements of these applications. ARO, which balances exploration and exploitation strategies inspired by natural foraging behavior, efficiently generates physically constrained adversarial examples while significantly outperforming traditional methods in computational efficiency. In contrast, our approach integrates global and local data insights within a four-tiered hierarchical framework. It begins with Kernel Density Estimation (KDE) combined with a Long Short-Term Memory Variational Autoencoder (LSTMVAE) encoder, which captures temporal relationships in the latent space. This facilitates accurate approximation of global data distributions in time-series data, enabling the effective identification of potential test seeds. Following this, a local robustness indicator is calculated to refine the selection of test seeds by leveraging insights from the global distribution and local characteristics of DL-based PHM system. Subsequent steps involve applying physical constraints to these test seeds based on the features of industrial application, such as PEMFC, ensuring the resulting Adversarial Examples (AEs) are realistic and compliant with physical principles. The generation and optimization of these AEs are then executed through the ARO algorithm, focusing on identifying the most effective adversarial instances. This strategy ensures the discovery of high-quality adversarial samples by harmonizing overall data distributions with specific local information, thereby testing the model’s adversarial robustness in a trustworthy way. The contributions of this paper are summarized below: 1. This paper presents Hierarchical Testing with Rabbit Optimization (HERO), a novel black-box adversarial testing framework which combines hierarchical data analysis with efficient optimization techniques to identify vulnerabilities in industrial deep learning models through physically constrained examples. 2. The proposed method integrates the accuracy of hierarchical testing with the efficiency of the state-of-the-art ARO algorithm, optimizing the identification and generation of high-quality time series adversarial samples for robustness evaluation. 3. Empirical validation of the HERO on real-world dataset, demonstrating its effectiveness in enhancing the resilience of PHM systems and the broader field of ICPS against threats posed by black-box models. II Preliminaries and Related Work Fig. 1 : Schematic Overview of the TSTransformer Architecture. II-A PHM Systems for Energy RUL Prediction in ICPS With the rapid development of ICPS, PEMFCs have become a pivotal