TLDR: This research introduces a new method for detecting attacks in Cyber-Physical Systems (CPS) that use ROS2. It addresses the challenge of limited labeled attack data in CPS by using “domain adaptation” to transfer attack knowledge from a well-labeled network-only dataset to an unlabeled, multi-layer CPS environment. The model combines contrastive learning, representation learning, clustering, and adversarial training to effectively identify anomalies across different system layers (network, OS, ROS) and outperforms other anomaly detection methods.
Cyber-physical systems (CPS), which integrate computing and physical components, are becoming increasingly vital for critical applications like unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs). These systems offer significant benefits but also present unique security challenges due to their complex, multi-layer designs involving networking, computational, and physical interfaces. Unlike traditional IT systems, attacks on CPS can have substantial real-world physical and economic impacts.
Intelligent intrusion detection systems (IDS) are crucial for securing CPS. However, many existing IDS solutions are trained and validated primarily on network traffic data, often overlooking attacks that can occur on other system layers, such as the operating system or the Robot Operating System (ROS) layer. This gap means that CPS could be vulnerable to distinct attack types that network-only IDS might miss.
The Robot Operating System (ROS) and its successor, ROS2, are fundamental frameworks for developing robotic and automation applications. ROS2, in particular, is designed with features relevant to CPS, including real-time functionality and security measures based on the Data Distribution Service (DDS) via Secure ROS2 (SROS2). Despite these advancements, ROS2 systems are not inherently immune to vulnerabilities, and there’s a recognized need to augment existing security with intelligent intrusion detection.
This research paper, titled “Securing Swarms: Cross-Domain Adaptation for ROS2-based CPS Anomaly Detection” by Julia Boone and Fatemeh Afghah, addresses this critical need. The authors propose an adaptable CPS anomaly detection model that can identify attacks within CPS without relying on previously labeled attack data in the target CPS environment. This is a significant advantage, as obtaining labeled attack data for complex CPS is often resource-intensive and impractical in real-world scenarios.
The core of their approach lies in utilizing domain adaptation techniques. This allows the model to transfer known attack knowledge from a well-labeled, network traffic-only environment (the source domain) to an unlabeled, multi-layer CPS environment (the target domain). This transfer of knowledge helps the system detect anomalies across various layers of a CPS, including network, operating system (OS), and ROS data, which is crucial for comprehensive security.
The proposed model is a novel domain adaptation framework that combines several advanced machine learning techniques. It uses dual feature encoders to independently process data from the source and target domains, transforming them into lower-dimensional representations. Contrastive learning is employed to learn features that are both distinct for different attack classes and robust enough to generalize across different domains. For the labeled source domain, a supervised triplet generation method is used, while for the unlabeled target domain, an unsupervised, time-based sampling technique helps create triplets for contrastive learning.
A domain discriminator is also integrated into the model. This component is trained adversarially to ensure that the feature encoders produce representations that are generalizable across both the source and target domains, making the model effective even when the data characteristics shift. Additionally, the model incorporates a Dunn Index-based center loss, which helps to enforce stricter regularization on the latent space, guiding benign and attack samples into tightly formed, well-separated clusters.
For anomaly detection, the model applies a K-means clustering algorithm to the generated latent representations. By configuring K-means with two clusters (representing benign and anomalous cases), the system can classify new, unlabeled test samples based on their proximity to these learned clusters.
The researchers validated their approach using two datasets: WUSTL-2021, a network-only dataset, as the source domain, and ROSPaCe, a state-of-the-art CPS intrusion dataset that combines network, OS, and ROS data, as the target domain. The WUSTL-2021 dataset includes various network attack types, while ROSPaCe features six different attacks, including discovery and Denial-of-Service (DoS) attacks across multiple layers.
The results demonstrate the effectiveness of their model. On the labeled, network-only source data (WUSTL-2021), the method achieved an impressive 98% attack detection accuracy. More importantly, on the unlabeled, multi-layer CPS target data (ROSPaCe), it achieved 87% accuracy. This performance is comparable to, and in many cases, outperforms other baseline anomaly detection methods, particularly when those baselines are trained on mixed benign and anomaly data. A key advantage highlighted by the authors is the use of sequenced system metric flows, which allows for aggregated metrics and periodic detection, potentially saving system resources.
Also Read:
- Securing Space Networks: A New Approach to Understanding Cyber-Physical Threats
- Securing Visual AI: Why Robustness is Key for Foundation Models
In conclusion, this work represents a significant step towards developing comprehensive, multi-layer IDS specifically designed for CPS. By bridging the gap between network-only IDS and the complex architectures of CPS through novel domain adaptation techniques, the research offers a promising solution for securing critical cyber-physical systems against evolving threats. You can read the full paper here.
