Time to Protection in Federated Intrusion Detection on Jetson Orin Nano Clusters
TL;DR
The findings show that federation delivers recognition of an attack class quickly when much of the fleet already holds it, and slowly or not within the authors' horizon when only a few devices have been hit, which is the case a defender is most concerned with.
Abstract
Federated Learning (FL) allows IoT devices to improve a shared intrusion detector without exchanging raw data. A device, however, can only learn from attacks that reach it, and because federated detection is evaluated as a whole for the fleet of IoT devices or averaged over attack classes, it is not reported how long a device that has never observed an attack class remains exposed to it without protection. The aim of this paper is to measure that exposure window, round by round, and to determine how it depends on the share of the fleet that already holds the class. We deployed a federated system with nine clients on two Jetson Orin Nano clusters connected over a WAN. As the main dataset, we used N-BaIoT, which covers two devices that were never infected by the mirai attack. It was tested with one to seven attacked clients and validated on CWRU and PAMAP2 datasets. The results showed that new class arrival depends on the share of the fleet that already knows about it, reaching 0.827 on mirai after one round when trained from the start with mirai data on seven out of nine clients, and reaching protection thresholds after 4, 17, and 39 rounds for seven, three, and two of nine clients attacked, when the model was initially trained without knowledge of mirai. Our findings show that federation delivers recognition of an attack class quickly when much of the fleet already holds it, and slowly or not within our horizon when only a few devices have been hit, which is the case a defender is most concerned with.
