ROS-based experimental platform for network attack and detection in networked multi-robots

Networked robot cooperative control systems operating over open networks are vulnerable to cyber-attacks such as false data injection, replay, and denial-of-service, which can undermine leader–follower formation stability and control performance and complicate systematic security experiments. To address this issue, this paper designs and implements a ROS-based experimental platform for network attacks and threat detection in networked robots, integrating cooperative control, attack injection, and intelligent detection into a unified framework for teaching and research.

Built on the ROS communication framework and the Gazebo simulation environment, this platform includes a leader–follower formation control subsystem, a network attack injection subsystem, and a network attack detection subsystem. The leader–follower formation control subsystem facilitates the exchange of position and velocity information between a leader robot and multiple followers through ROS topics, supports common motion tasks like straight-line and circular trajectory tracking and various formation patterns, and offers a visualization interface that displays robot poses, inter-robot distances, and formation errors in real time. This provides an intuitive way to observe how cooperative behavior degrades under malicious interference. The network attack injection subsystem utilizes standard network analysis tools to identify host and traffic characteristics, creates a man-in-the-middle environment between the control workstation and the robots, and employs script-based configuration of packet payloads, sending rates, and target ports to simulate representative attack scenarios, including falsified state information, replay of historical data, and flooding-based denial-of-service, with adjustable intensity and duration. On the defense side, the detection subsystem constructs temporal samples by jointly analyzing network traffic statistics and the physical states of the robots, aligns and segments these heterogeneous data streams using timestamps and sliding time windows, and performs deep multimodal data fusion through feature-level concatenation and normalization before inputting them into a proposed CNN–BiLSTM–Attention model. In this model, convolutional layers extract local spatiotemporal patterns from the fused features, the bidirectional LSTM captures long-term dependencies between network conditions and cooperative movement, and an attention mechanism highlights time segments more indicative of abnormal behavior. The system then classifies attacks through a softmax output layer. Training involves weighted loss functions and regularization to address class imbalance and enhance generalization.

Experimental results from datasets collected under normal operation and multiple attack scenarios demonstrate that the platform can reliably support multi-robot leader–follower formation simulations and accurately reproduce phenomena such as formation divergence, follower lag, increased communication delay, and higher packet loss. Compared to baseline models that only use network traffic or physical states as inputs, the proposed CNN–BiLSTM–Attention detector with deep multimodal data fusion achieves higher overall accuracy, precision, and F1-scores, and exhibits more balanced performance across attack types, especially those that are hard to distinguish using a single modality. These results show that combining attack injection experiments with multimodal learning significantly improves attack detection sensitivity and robustness.

Overall, the platform provides a comprehensive experimental chain—from theoretical explanation and attack simulation to intelligent defense—for network security education in cooperative robotic systems. Additionally, it offers a scalable and easily extendable testbed for investigating complex attack scenarios and developing cyber-physical defense strategies in realistic networked robot environments.