Deep-Radial Basis Function (RBF) networks provide arejection class alongside the class predictions, which can be utilized for detecting anomalies at runtime . In this paper, we show how the deep-RBF network can be used for detectinganomalies in CPS regression tasks such as continuous steering predictions . We also use the resulting rejection class for detectingadversarial attacks such as a physical attack and data poison attack . We hypothesize that a single network that can perform controller predictions and anomaly detection is necessary to reduce the resource requirements . Our results show that the deep RBF networks can robustly detect these attacks in a short time without additional resource requirements. We also evaluate these attacks using a hardwareCPS testbed called DeepNNCar and a real-world German Traffic Sign Benchmark(GTSB) dataset. Finally,we evaluated these attacks and the trained deep-NRF networks using the trainedDeepNNCar’s ‘Data Poison Attack’. Our results showed that theDeep-RBf networks can be robustly

Author(s) : Matthew Burruss, Shreyas Ramakrishna, Abhishek Dubey

Links : PDF - Abstract

Code :

Keywords : deep - rbf - networks - attacks - attack -

Leave a Reply

Your email address will not be published. Required fields are marked *