Threat Modelling with Domain-Adapted Language Models: Small Models Match Large on STRIDE
arXiv·medium signal
Systematic evaluation of domain-adapted language models of varying sizes against general-purpose LLMs for structured threat modelling (STRIDE framework). Key finding: small language models (SLMs) fine-tuned on cybersecurity data match or exceed general-purpose LLMs on threat identification tasks, offering a practical path for organizations that cannot send sensitive architecture details to external APIs.