In this work, we consider two real autonomous robotic systems as case studies: a robotic boat for water quality monitoring and a social robot. We propose a solution for identifying possible attacks by analysing the robot’s behaviour. The proposed approach is based on the analysis of system logs and on the development of an analyser module, based on autoencoders, to identify abnormal behaviours. In more details, the proposed approach aims at detecting attacks to a cyber-physical system, in real-time, by processing system logs containing data related to the activities of the system (e.g., GPS position, heading, and velocity). Data are recorded at short time interval (e.g., every 200 ms) hence the logs provide a rich representation of the behaviour of the system that is extremely valuable to identify faults and cyber-physical attacks. In particular, we propose a transformation process to convert the different types of variables included in the log records into either grey-scale or black-and-white pixel array to handle sparse and non-linear features. This process allows to learn specific patterns of the log data related to normal behaviours and to detect possible abnormal situations.

From this thesis, we have extracted a paper that has been published at IROS 2019, in Macau, China.