Research on Automated Software Defect Detection Using Deep Learning Techniques

Abstract

Aligned with the prevailing research trajectory that emphasizes the integration of semantic comprehension and structural modeling within deep learning frameworks, this paper investigates an automated system for software defect detection. The study elaborates on the system’s architecture, which is designed to jointly encode code semantics and syntactic structure through a hybrid representation approach combining graph neural networks and transformer-based language models. We detail the model formulation, which fuses program dependency graphs with contextual token embeddings, and describe a multi-stage training strategy that optimizes for both accuracy and inference efficiency. Furthermore, the paper introduces a seamless integration mechanism that embeds the detection pipeline into continuous integration (CI) environments, enabling real-time static analysis during development workflows. The system is rigorously evaluated on the CodeXGLUE and Devign benchmarks, where it achieves state-of-the-art performance, surpassing existing methods in key metrics including F1-score (by 3.2 %) and detection latency (by 18 %). These results substantiate the system's strong practical applicability and its potential for scalable engineering deployment in modern software development pipelines.