IoT device fingerprint using deep learning

Published in 2018 IEEE international conference on internet of things and intelligence system (IOTAIS), 2018

Recommended citation: Sandhya Aneja, Nagender Aneja, Md Islam "IoT device fingerprint using deep learning." 2018 IEEE international conference on internet of things and intelligence system (IOTAIS), 2018. pp. 174--179 doi: 10.1109/IOTAIS.2018.8600824

(Conference Publication)

Access paper here

Abstract: Device Fingerprinting (DFP) is the identification of a device without using its network or other assigned identities, including IP address, Medium Access Control (MAC) address, or International Mobile Equipment Identity (IMEI) number. DFP identifies a device using information from the packets that the device uses to communicate over the network. Packets are received at a router and processed to extract the information. In this paper, we worked on the DFP using Inter Arrival Time (IAT). IAT is the time interval between the two consecutive packets received. It has been observed that the IAT is unique for a device because of the different hardware and software used for the device. The existing work on the DFP uses statistical techniques to analyze the IAT and to generate further information using which a device can be identified uniquely. This work presents a novel idea of DFP by plotting graphs of IAT for packets, with each graph plotting 100 IATs and subsequently processing the resulting graphs to identify the device. This approach improves the efficiency of identifying a device DFP due to the benchmark achieved by deep learning libraries in image processing. We configured Raspberry Pi to work as a router and installed our packet sniffer application on the Raspberry Pi. The packet sniffer application captured the packet information from the connected devices in a log file. We connected two Apple devices, iPad4 and iPhone 7 Plus, to the router and created IAT graphs for these two devices. We used Convolution Neural Network (CNN) to identify the devices and observed the accuracy of 86.7%.