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

Zhaoteng Yan, Nan Yu, Hui Wen, Zhi Li, Hongsong Zhu, Limin Sun

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 $$92\%$$ . 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.