Abstract

Deception technologies like honeypots generate large volumes of log data, which include illegal Unix shell commands used by latent intruders. Several prior works have reported promising results in overcoming the weaknesses of network-level and program-level Intrusion Detection Systems (IDSs) by fussing network traffic with data from honeypots. However, because honeypots lack the plug-in infrastructure to enable real-time parsing of log outputs, it remains technically challenging to feed illegal Unix commands into downstream predictive analytics. As a result, advances on honeypot-based user-level IDSs remain greatly hindered. This article presents a run-time system (GPT-2C) that leverages a large pre-trained language model (GPT-2) to parse dynamic logs generated by a live Cowrie SSH honeypot instance. After fine-tuning the GPT-2 model on an existing corpus of illegal Unix commands, the model achieved 89% inference accuracy in parsing Unix commands with acceptable execution latency.