DPDA: A dual-path domain adaptation method for cross-terrain real-time attack detection in mobile robots

Deep learning-based attack detection methods for mobile robots achieve high accuracy within a single operating environment, but exhibit significant limitations when deployed in real-world cross-terrain scenarios. The core issue stems from the inability of single-domain models to adapt to data distribution shift across terrains, resulting in the misclassification of terrain-induced variations as potential attacks. To overcome this challenge, a dual-path domain adaptation (DPDA) network is proposed for cross-terrain real-time attack detection in mobile robots. The temporal dynamics and spectral characteristics of sensor data are extracted through the dual-path architecture in DPDA, overcoming the informational constraints of unimodal approaches. Furthermore, a time-frequency correlator (TFC) and a bidirectional cross attention (BCA) mechanism are introduced to enable interactive fusion of bimodal features, capturing stable attack features across domains. To tackle the dual-target domain alignment problem, a reciprocal prototype drift (RPD) mechanism is designed to bridge disparate target domains. By guiding feature prototypes from different target domains to converge mutually in the embedding space, distribution shift induced by terrain variations is effectively mitigated. Experiments conducted on a real mobile robot platform demonstrate that DPDA surpasses existing methods in cross-terrain attack detection tasks. Cross-dataset validation further confirms the model's robust cross-domain generalization capability.