Deep Learning Prediction Models for the Detection of Cyber-Attacks on Image Sequences

With the introduction of Cyber Physical Systems and Industrial Internet of Things within Industry 4.0, vision systems, as indispensable element for robot cognition, become smart devices integrated into the control system using different communication links. In this control framework image streams are transferred between elements of distributed control system opening the possibility for various cyber-attacks that can cause changes in certain parts of images eventually triggering wrong decisions and negative consequences to the system performance. Timely detection of the attacks on communicated image streams is necessary to mitigate or completely avoid their negative effects. In this paper we propose a method for the prediction of the next image in the sequence which can be utilized for the development of anomaly-based cyber-attack detection mechanisms. For the model generation, we have explored the application of several deep learning architectures based on two-dimensional Convolutional Neural Networks and Convolutional Long Short-Term Memory Recurrent Neural Networks. Images obtained from the real-world experimental installation were utilized for model design. Our deep learning models proved to be effective in predicting the next frames according to the criteria of a discrepancy between pixels of the real and estimated images.