GLHAD: A Group Lasso-based Hybrid Attack Detection and Localization Framework for Multistage Manufacturing Systems

Ahmad Kokhahi, Dan Li

Abstract

As Industry 4.0 and digitalization continue to advance, the reliance on information technology increases, making the world more vulnerable to cyber-attacks, especially cyber-physical attacks that can manipulate physical systems and compromise operational data integrity. Detecting cyber-attacks in multistage manufacturing systems (MMS) is crucial due to the growing sophistication of attacks and the complexity of MMS. Attacks can propagate throughout the system, affecting subsequent stages and making detection more challenging than in single-stage systems. Localization is also critical due to the complex interactions in MMS. To address these challenges, a group lasso regression-based framework is proposed to detect and localize attacks in MMS. The proposed algorithm outperforms traditional hypothesis testing-based methods in expected detection delay and localization accuracy, as demonstrated in a linear multistage manufacturing system.

Introduction

In the era of Industry 4.0, manufacturing facilities have evolved into highly connected cyber-physical systems. While this connectivity enhances efficiency, it also exposes factories to "False Data Injection" (FDI) attacks. In these scenarios, adversaries manipulate sensor data to induce physical damage or subtle quality defects that can go unnoticed by traditional monitoring systems.

Multistage Manufacturing Systems (MMS) present two primary challenges for security:

  1. Error Propagation: In a sequential process, an anomaly introduced at an early stage ripples through subsequent operations. This propagation makes it difficult to distinguish whether a deviation is local or a downstream effect of an earlier compromise.
  2. Data Scarcity: Modern AI-based security tools typically require extensive datasets of historical attacks for training. However, manufacturers rarely possess a comprehensive library of successful cyber-attacks to facilitate supervised learning.

System Architecture and Methodology

The researchers introduced GLHAD (Group Lasso-based Hybrid Attack Detection), a framework that leverages the inherent physics of the manufacturing process to create unique "signatures" for potential attacks. Unlike purely data-driven methods, GLHAD functions as an anomaly detector that understands the mathematical relationships between stages.

Architecture of a multistage manufacturing system under sensor spoofing attacks.

The technical workflow involves several key components:

  • State-Space Modeling: The factory is represented using mathematical equations describing product transitions between stages.
  • Kalman Filtering: A standard Kalman Filter estimates the "healthy" state of the system, providing a baseline for normal operation.
  • Theoretical Residual Analysis: The authors derived how an attack at any specific stage uniquely distorts sensor residuals across the entire system. This distortion serves as a mathematical fingerprint.
  • Group Lasso Regression: Because an attack typically targets a specific stage, it affects a "group" of associated sensors. Group Lasso regression is employed to identify which specific group (stage) is responsible for the observed noisy signals. This technique effectively "zeroes out" non-culprit stages, facilitating simultaneous detection and localization.

Experimental Evaluation

The framework was evaluated using both numerical simulations of a 4-stage linear system and a real-world case study based on established assembly process literature. The performance of GLHAD was compared against a standard benchmark, the Hotelling $T^2$ test.

Key Results

GLHAD demonstrated superior performance across several metrics:

  • Detection Speed: The framework showed a significantly lower Average Run Length (ARL), indicating faster detection of attacks, particularly when the attack signal was weak.
  • Localization Accuracy: As shown in the experimental results, GLHAD achieved near-perfect localization accuracy at higher Signal-to-Noise Ratios (SNR). Even under low SNR conditions, it consistently outperformed the benchmark, which often failed to distinguish between the attacked stage and subsequent stages affected by propagation.
Localization accuracy of the GLHAD framework (F) vs. benchmark (b) across different stages and signal-to-noise ratios.

Conclusion and Future Work

GLHAD provides a robust solution for securing MMS by treating the system's physics as a security signature. By identifying the specific stage of an attack during the detection process, it allows for rapid response and mitigation.

While the current framework is optimized for linear systems and single-point attacks, future research will focus on expanding GLHAD to handle non-linear manufacturing dynamics and multiple simultaneous attacks across different parts of the factory. This evolution is essential for protecting increasingly complex and interconnected industrial environments.

Summary updated

Abstract

As Industry 4.0 and digitalization continue to advance, the reliance on information technology increases, making the world more vulnerable to cyber-attacks, especially cyber-physical attacks that can manipulate physical systems and compromise operational data integrity. Detecting cyber-attacks in multistage manufacturing systems (MMS) is crucial due to the growing sophistication of attacks and the complexity of MMS. Attacks can propagate throughout the system, affecting subsequent stages and making detection more challenging than in single-stage systems. Localization is also critical due to the complex interactions in MMS. To address these challenges, a group lasso regression-based framework is proposed to detect and localize attacks in MMS. The proposed algorithm outperforms traditional hypothesis testing-based methods in expected detection delay and localization accuracy, as demonstrated in a linear multistage manufacturing system.
BibTeX
@article{kokhahi2023glhad,
  author = {Ahmad Kokhahi and Dan Li},
  title = {{GLHAD: A Group Lasso-based Hybrid Attack Detection and Localization Framework for Multistage Manufacturing Systems}},
  journal = {Journal of Computing and Information Science in Engineering},
  year = {2023},
  doi = {10.1115/1.4063862},
}
Figure