Robust and Energy Efficient Malware Detection for Robotic Cyber-Physical Systems

Cyber-Physical Systems (CPS) increasingly use multiple robots as edge devices to enhance their functionalities. However, this introduces new security vulnerabilities such as control channel attacks and false data injection that an adversary can exploit to put the users and environment at risk. In this paper, we build a robust malware detection system strengthened by carefully crafted adversarial samples. We generate adversarial samples within the bounds of domain constraints and integrate them into model training to improve the model's robustness. Additionally, we formulate an objective function to distribute the computation of malware detection to multiple edges, making optimal use of the robot mesh network to reduce power consumption. In the adjoining poster, we show the details of the dataset and the models, and illustrate the specifics of our contributions.