Developing Synthetic Data or Cybersecurity Policies

Data scarcity – driven by under-detection and under-reporting of incidents, legal and competitive disincentives to share, and vendors’ reluctance to expose product weaknesses – impedes the development of data-driven cybersecurity policies. We investigate large language model (LLM)–based synthetic tabular data generation as a pragmatic remedy. Our approach follows a GReaT-style pipeline: (i) text-based serialization of heterogeneous tables to preserve schema semantics; and (ii) fine-tuning a pretrained decoder LLM (Unsloth/Llama-3.2-1B) on 3388 publicly reported cyber-attack records. Preliminary results that LLM-generated synthetic data can approximate the statistical and structural properties of scarce cybersecurity data without exposing sensitive information, thereby enabling data augmentation and supporting the design of data-driven cyber policies