Abstract
With the advent of lightweight cryptography in edge devices, attackers can hide malicious code under encrypted network communications to perform malware attacks. This makes IoT botnet attacks extremely challenging to detect using traditional signature-based techniques. In this paper, we propose a novel IoT botnet detection framework that uses vision transformers to detect malicious communications captured in encrypted network flow images. Our approach achieved ∼98% accuracy and around 94% reduced inference latency compared to network flow header analysis, the popular network-based botnet detection technique. Further, we have validated the practicality of our approach by testing it on Jetson Orin Nano acting as an edge gateway and achieved reduced inference latency of 25.16 ms and area overhead of 88.13 MB.
Funding Information
  • New York University Abu Dhabi (NYUAD) Center for Cyber Security (CCS), funded by Tamkeen under the NYUAD Research Institute Award (G1104)
  • C3ihub IIT Kanpur (NTIHAC/2025/01/3)

This publication has 23 references indexed in Scilit: