An Experimental Analysis on Mitigating the Effects of Malicious Nodes in a Federated Learning System

This paper describes how deep learning can be used to provide security for IoT devices by analyzing the data packets that arrive at an IoT device and classifying them as packets part of the normal operation of the device or packets sent with a malicious intent. An experimental analysis is performed to check the effectiveness of such an approach with the help of the data present in the MQTT dataset. Federated learning approach is suitable for the IoT platform as IoT devices tend to contain less computing power. But a consequence of this is that the networks can contain malicious nodes which send wrong updates to the model decreasing its accuracy. We propose the introduction of verifier nodes into the system which verify the given updates sent by a node and check if it actually increases the accuracy of the model before appending it to the global model. The extent to which the malicious nodes impact the accuracy of the model and the remedy provided by the introduction of verifier nodes is also studied in this paper.