Authors - Jason Elroy Martis, Ronith, Anvitha Rao, Vignesh Salian, Apoorva Shetty, Philomina Princiya Mascarenhas Abstract - The task of recovering high-level architectures from embedded software systems is error-prone and difficult, and state-of-the-art methods still rely on static analysis or heuristics and lack explainability. To address these challenges, an explainable and automated method for recovering high-level architectural diagrams directly from source code is suggested. Specifically, this method begins with the generation of function call graphs at the function level via static analysis and functions grouping into domain-agnostic component classes, generating a component graph. Components are then augmented with semantic attributes learned via CodeBERT embeddings, facilitating a light graph convolutional network (light GCN) model for learning-component interactions reflecting structure and semantics. Methods for explainability via gradients are incorporated for emphasizing prominent components and edges, helping in developer understanding, validation, and tuning of predicted architectures. The performance of this method on several embedded projects showed accuracy as high as 91.87%, precision of 96.48%, recall of 86.90%, and an F1-score of 91.44%. Use cases have shown successful extraction and interpretation of critical paths, bottlenecks, and unusual architectures and highlight explainable insights that enable efficient analysis and thus make it a highly significant progress in explainable AI for embedded software.
Open Zoom