Unknown Attack Detection via Hierarchical Feature Fusion in Mobile Robot Systems

With the development of mobile robots, the anomaly detection of robot sensor data has become a research hot spot. Anomaly detection is crucial for the normal operation of robots because it can timely detect possible attacks during the movement and prevent sensor-dependent robots from being induced to take wrong actions. However, traditional method can only classify known attack types and are unable to identify unknown attack types. In that case, this paper proposes a Hierarchical Feature Fusion-based Unknown Attack Detection (HFF-UAD) method for unknown attack detection in mobile robotic systems. A multilayer convolutional neural network cascaded with a self-attention module is used, extracting hierarchical features effectively from the raw data. Meanwhile, unknown attack detection is achieved through the Nearest Class Mean (NCM) classifier based on Mahalanobis distance. The implementation using real robotic sensor data verifies the effectiveness of HFF-UAD and confirms that it can detect both known and unknown attacks.