Automating Data Flow Diagram Generation from User Stories Using Large Language Models

Visual modeling, particularly Data Flow Diagrams (DFDs), plays an essential role in modern software development, aiding in the design, understanding, and communication of system structures and potential security and privacy threats. Despite their importance, the manual creation of visual models is time-consuming highlighting the need for automation in the generation of DFDs from user requirements. Automating the generation of DFDs presents a significant challenge, especially in accurately interpreting user requirements and abstracting them into correct and complete diagram elements. The complexity of this task is compounded by the need for semantic accuracy and the ability to facilitate visual editing for human intervention. This study explores the use of Large Language Models (LLMs) to automate DFD generation, utilizing GPT-3.5, GPT-4, Llama2, and Mixtral models. This study emphasizes human oversight and employs an open-source diagramming tool to ensure that diagrams are accurate, complete, and editable. The findings reveal GPT-4’s superior capability in generating complete DFDs, with significant progress from open-source models like Mixtral, indicating a viable path toward automated visual modeling. This approach advances scalable automation in creating visual software models, with broader implications for automating other diagram types.