What is Your Robot Doing? Fingerprinting Legged Robot Movements from Encrypted Traffic
Year: 2026
Type: conference-paper
Abstract: Modern commercial legged robots can be remotely operated by forwarding commands over an encrypted Wi-Fi link. However, encryption might not be enough to conceal the commands and, in turn, the robot's resulting physical movement. In this paper, we present a passive traffic-analysis attack that infers such physical movement from the encrypted communication between two commercial robots (Unitree's Go2, and Unitree's G1) and the manufacturer's companion mobile application. Our approach relies only on observable metadata, packet sizes, inter-arrival times, and direction, without decrypting any payload. For each robot, we recorded 700 labelled captures of five movements and turned them into three complementary representations: windowed, per-capture, and per-packet. We used the resulting five datasets to train a set of 11 models, spanning classical tabular classifiers and deep sequence models, all under a leakage-safe, capture-level evaluation. Our attack recovers Go2's movements with up to 0.97 macro-F1 and G1's movements with up to 0.74, both clearing the 0.08 majority-class baseline. Our results are consistent across model families and not tied to a single method. A SHAP analysis shows that the two platforms leak through different features: packet timing on the Go2 and packet size on the G1. Our findings show that encryption protects the content of robots' communication, but not their behaviour: using only a commodity laptop, an observer on the same network can infer the robot's physical activity from traffic patterns alone, creating a concrete privacy and operational-security risk for deployed robotic systems.
Institutions: Örebro University, University of Padua
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7220700107