ABA-IDS: Attention-Based Autoencoder for Intrusion Detection in Assistive Mobility Robotic Network

Innovative mobility aids like smart wheelchairs and walkers are transforming the lives of individuals with mobility challenges. These devices use sensors and in-vehicle network technologies, such as Controller Area Network (CAN), to enable real-time data exchange. While CAN is cost-effective, robust, and low-latency, it is vulnerable to cyber-attacks due to issues like message transmission on a shared bus, ID-based priority arbitration, and lack of encryption or authentication mechanisms. To mitigate these vulnerabilities, researchers have developed machine and deep learning-based intrusion detection systems (IDS). However, these systems often struggle with low feature learning and high model complexity, making them unsuitable for resource-constrained CAN networks. This paper introduces ABA-IDS, a low-footprint IDS that uses a ConvLSTM autoencoder with an attention mechanism to detect anomalies in assistive mobility networks. The model’s design allows it to efficiently capture key spatial and temporal features of the network. By preprocessing CAN data based on CAN ID and transmission frequency, ABA-IDS combines frequency-based and content-based approaches to enhance anomaly detection. Anomaly scores are generated from reconstruction errors, and testing on a benchmark dataset showed that ABA-IDS outperformed existing models with fewer parameters. This makes ABA-IDS effective for deployment in resource-constrained in-vehicle networks, significantly improving the security of assistive mobility devices against cyber threats. The link to the project repository can be found here.