Detecting Internet-Scale NATs for IoT Devices Based on Tri-Net

Abstract

Due to the lack of available labeled Network Address Translation (NAT) samples, it is still difficult to actively detect the large-scale NATs on the Internet. In this paper, we propose an novel method to identify NATs for online Internet of Things (IoT) devices based on Tri-net (a semi-supervised deep neural network). By learning the features on three layers (network, transport and application layer) in the small labeled data set (with thousands of instances), the Tri-net can automatically identify millions of online NATs. We evaluate this approach on the real-world dataset with more than 8 million online IoT devices, and the performance shows the precision and recall can be both up to. Moreover, we found 2, 511, 499 IoT devices connecting to the Internet via NAT, which account for one-third of the total. To our knowledge, this is the first successful attempt to automatically identify Internet-scale NATs.