On the Challenges of Holistic Intrusion Detection in ICS
Stefan Lenz 1, Julia Raab 1, Benedikt Holzbach 1, Deniz Köller 1, Sotiris Michaelides 1, Martin Henze 1,2 1 Security and Privacy in Industrial Cooperation, RWTH Aachen University • 2 Cyber Analysis & Defense, Fraunhofer FKIE • · 2026
Read On the Challenges of Holistic Intrusion Detection in ICS by Stefan Lenz 1, Julia Raab 1, Benedikt Holzbach 1, Deniz Köller 1, Sotiris Michaelides 1, Martin Henze 1,2 1 Security and Privacy in Industrial Cooperation, RWTH Aachen University • 2 Cyber Analysis & Defense, Fraunhofer FKIE • (2026) 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-2604-21626.txt.
Introduction
Stefan Lenz and five colleagues present challenges in research toward a complete intrusion detection system for industrial control systems, meant to cover every part of an ICS. Past attacks show foes often target both the ICS network and the real process, risking major harm. Intrusion detection promises quick finding of foes, but current tools usually focus on single traits such as packet timings, forcing multiple parallel systems that tangle work. The paper lists these challenges to spur debate and further research.
Catalog record
| CREATOR | Stefan Lenz 1, Julia Raab 1, Benedikt Holzbach 1, Deniz Köller 1, Sotiris Michaelides 1, Martin Henze 1,2 1 Security and Privacy in Industrial Cooperation, RWTH Aachen University • 2 Cyber Analysis & Defense, Fraunhofer FKIE • |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 10177 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.21626
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
On the Challenges of Holistic Intrusion Detection in ICS Stefan Lenz 1, Julia Raab 1, Benedikt Holzbach 1, Deniz Köller 1, Sotiris Michaelides 1, Martin Henze 1,2 1 Security and Privacy in Industrial Cooperation, RWTH Aachen University • 2 Cyber Analysis & Defense, Fraunhofer FKIE • Abstract Past attacks against industrial control systems (ICS) show that adversaries often target both the ICS network and the physical process to achieve potential catastrophic impact. To secure ICS, intrusion detection systems promise timely uncovering of such adversaries. However, as these detection mechanisms typically focus on isolated characteristics of ICS (e.g., packet timings), multiple detection systems have to be deployed in parallel, complicating their operation in practice. In this work, to spur discussion and further research, we present challenges encountered during our research towards a holistic intrusion detection system aiming to cover all dimensions of an ICS. \springyear 2026 \springdate April 21–22 \springlocation Heidelberg, Germany 1 Holistic Intrusion Detection in ICS By connecting industrial control systems (ICS) to the Internet, these safety critical systems are exposed to sophisticated attackers [ 7 , 6 , 3 ] . As ICS consist of low-resource legacy devices that are difficult to replace and not capable of complex security operations, retro-fitting attack detection into ICS has been a major focus of recent research [ 19 , 8 ] . Since the behavior of ICS is tightly coupled with a deterministic physical process, typical industrial intrusion detection systems (IIDS) model benign behavior (e.g., packet timings [ 10 ] or process states [ 17 ] ) to later identify anomalies. However, as attacks materialize in the network and the physical dimension of ICS, comprehensive detection must cover all aspects of ICS [ 20 ] . Thus, we argue that IIDS should aim for a holistic detection approach, covering multiple dimensions simultaneously as depicted in Fig. 1 . One method to achieve holistic detection is to use multiple one-dimensional detectors in parallel [ 20 ] . Although such ensembles promise great detection performance in theory , they are difficult to deploy, since each detector requires sufficient training data as well parameter optimization. Furthermore, individual detectors risk to miss interplay between different dimensions and it is unclear how to combine alerts of individual detectors in a sensible manner [ 20 ] . Figure 1: Holistic industrial intrusion detection systems (IIDS) aim to monitor both network behavior and physical process state, while traditional IIDS focus on one. To overcome these challenges, we set out to address the need for a practical holistic detection method. Instead of combining different detectors specializing in individual aspects, we strive to cover the complete benign behavior of an ICS in one holistic model. During this pursuit, we encountered several challenges while exploring different approaches. In this paper, we discuss these challenges regarding process state discretization (§ 2 ), IIDS parameterization (§ 3 ), and collecting sufficiently good training data (§ 4 ). Metric Name Accuracy Precision Recall F0.1 F1 eTaP eTaR eTaF0.1 eTaF1 Jenkspy 0.68 0.23 0.15 0.23 0.21 0.23 0.52 0.24 0.32 kMeans 0.68 0.24 0.16 0.23 0.21 0.24 0.52 0.24 0.32 Quantiles 0.77 0.71 0.00 0.03 0.00 0.69 0.14 0.66 0.23 Inter-Arrival Times Medium L1 [ms] L2 [ms] L3 [ms] Wired 46.06 (0.37) 46.06 (0.37) 100.00 (0.02) Wireless + 46.06 (0.37) 46.06 (0.37) 100 (0.12) Wireless - 45.97 (1.47) 45.97 (1.56) 100 (0.67) Table 1: The best algorithm ( green ) to discretize process data not only depends on data distribution but also on the detection performance metric. Figure 2: Alerts ( black ) of LLM detection on SWaT ( red ) are sensitive to process data variance. Monitoring subprocesses reduces processing overhead. Table 2: Differences in inter-arrival times in wired and good ( + ) / bad ( - ) wireless channels show the need for detection to adapt to dynamic behavior. 2 Challenge 1: Discretization To create a holistic IIDS capturing the complete ICS in a single model, we explored the use of process mining , which has shown promising results in modeling physical process state [ 12 , 2 ] as well as network traffic [ 15 ] . Indeed, initial experiments showed that such a detector can detect attacks targeting both these dimensions, while performing on par with other IIDSs that capture only one dimension. However, as process mining relies on discrete labels (e.g., bins) to incorporate the physical state into the model, the fundamental challenge of discretization arises to assign such labels to continuous ICS data (e.g., temperatures or fill levels). Although many discretization mechanisms, such as statistical methods or unsupervised learning approaches (e.g., clustering), solve this problem, the quality of the resulting labels/clusters is highly dependent on the distribution of the underlying data [ 1 , 4 , 16 ] . To choose the optimal discretization method for an IIDS using process mining, we compared multiple discretization methods based on their resulting detection performance as summarized in Tb. 1 . Our experiments show that, even while utilizing benchmarking datasets (with posterior knowledge that is not achievable in any practical scenario), this task is not trivial, as the best discretization method ( green in Tb. 1 ) depends on the considered performance metric. Take Away While discretization mechanisms promise well-defined labels for process data, in the context of intrusion detection, identifying the “best” approach depends on the considered detection metric. 3 Challenge 2: Parameterization A holistic IIDS, e.g., based on process mining, requires multiple additional parameters (e.g., length of the ICS cycle in packets) that must be set manually, resulting in a need for hard-to-come-by expert knowledge. To address the resulting challenge of parameterization , we explored different “closed-box” modeling approaches (e.g., based on LLMs), which promise no parameterization at all. The goal of such approaches is to gain “adequate” detection performance while utilizing the reasoning capabilities of LLMs. Exemplary experiments on the SWaT [ 5 ] dataset (cf. Fig. 2 ) indicate that an LLM-based IIDS struggles to monitor the complete process in a holistic manner, due to excessive resource demands, even exceeded the capabilities of the high-performance cluster node (96GB VRAM) we utilized. Therefore, we focused on individual subprocesses in further experiments. While the smaller scope reduces computational complexity, the detector provided nearly-constant alerts, rendering attacks unrecognizable (except for Process 5 which has comparatively low variance). Still, due to its closed-box nature, it is impossible to comprehend the detector’s reasoning behind alerts [ 14 ] , making discerning between true and false alarm challenging. Together with the high resource demands, these issue prevent a practical application of a parameterless approach as a holistic IIDS. Take Away Although addressing the challenge of parameterization, monolithic closed-box detectors face challenges in deployment, resource usage, and interpretability, preventing their use for holistic IIDS. 4 Challenge 3: Good Training Data To meet increasing demands for flexibility and efficiency, ICS adopt new technologies such as wireless communication [ 11 ] . These more dynamic systems exaggerate the challenge of gathering good training data in IIDS research [ 18 ] . To investigate this issue, we built an ICS simulation enabling us to change the communication medium without altering the process behavior [ 9 ] . There, we analyze a popular timing-based IIDS [ 10 ]