Backdoor Suppression in Neural Networks using Input Fuzzing and Majority Voting

Abstract

While inference is needed at the edge, training is typically done at the cloud. Therefore, data necessary for training a model, as well as the trained model, have to be transmitted back and forth between the edge and the cloud training infrastructure. This creates significant security issues, including the inclusion of a backdoor sent to the user without the user's knowledge. This article presents an approach where a trained model can still operate as expected, irrespective of the presence of such a backdoor.

Topics & Concepts