Data injection detection in ROS 2 robotic systems using entropy metrics and an autoencoder+LSTM pipeline

Abstract

Distributed robotic systems based on ROS 2 require lightweight, real-time mechanisms for the early detection of anomalous behavior (Blázquez-García et al., 2021). Data injection or irregularities in topic publication compromise security, robustness, and operational continuity. Although ROS 2 incorporates security measures (e.g., DDS Security), its overhead can be non-trivial; therefore, complementary monitoring strategies that operate on internal telemetry, without network instrumentation or heavy cryptography, are valued (Zhang et al., 2022; Fernández et al., 2018). This work presents a hybrid approach that combines entropy metrics calculated in sliding windows with an Autoencoder (AE) + LSTM pipeline trained exclusively on nominal robot data. Three complementary entropic measures are employed: Shannon (Shannon, 1948), transitions (based on successive differences; Nardone, 2014), and KDE (a nonparametric approximation of density; Myers et al., 2025). All are calculated over windows of W = 100 samples to capture recent behavior. Stationarity is also verified using ADF on nominal segments, enabling statistical calibration of thresholds (Dickey and Fuller, 1979; Wang et al., 2023). The use of AE and LSTM is supported by robust representation and sequential modeling (Vincent et al., 2008; Hochreiter and Schmidhuber, 1997; Malhotra et al., 2016; Hundman et al., 2018). The system relies on internal telemetry (e.g., /cmd_vel, /odom, /laser_scan, /imu/data; and /movement_label, /attack_type tags) and produces a binary decision with K-persistence and alerts (/ads/alert). In offline evaluation, with explicit selection of the operating point under FPR constraint, high detection rates and low latency are achieved, indicating feasibility for online deployment over ROS 2 (Abokhdair and Baig, 2025).