“Out-of-the-Box” Monitoring of VM-Based High-Interaction Honeypots

Abstract

Honeypot has been an invaluable tool for the detection and analysis of network-based attacks by either human intruders or automated malware in the wild. The insights obtained by deploying honeypots, especially high-interaction ones, largely rely on the monitoring capability on the honeypots. In practice, based on the location of sensors, honeypots can be monitored either internally or externally. Being deployed inside the monitored honeypots, internal sensors are able to provide a semantic-rich view on various aspects of system dynamics (e.g., system calls). However, their very internal existence makes them visible, tangible, and even subvertible to attackers after break-ins. From another perspective, existing external honeypot sensors (e.g., network sniffers) could be made invisible to the monitored honeypot. However, they are not able to capture any internal system events such as system calls executed. It is desirable to have a honeypot monitoring system that is invisible, tamper-resistant and yet is capable of recording and understanding the honeypot's system internal events such as system calls. In this paper, we present a virtualization-based system called VMscope which allows us to view the system internal events of virtual machine (VM)-based honeypots from outside the honeypots. Particularly, by observing and interpreting VM-internal system call events at the virtual machine monitor (VMM) layer, VMscope is able to provide the same deep inspection capability as that of traditional inside-the-honeypot monitoring tools (e.g., Sebek) while still obtaining similar tamper-resistance and invisibility as other external monitoring tools. We have built a proof-of-concept prototype by leveraging and extending one key virtualization technique called binary translation. Our experiments with real-world honeypots show that VMscope is robust against advanced countermeasures that can defeat existing internally-deployed honeypot monitors, and it only incurs moderate run-time overhead.