Adaptive Patch Management: Predicting Security Requirements in Evolving Software Environments

Abstract

The integration of Machine Learning (ML) into software maintenance processes represents a paradigm shift in identifying, predicting, and remediating security vulnerabilities and defects within evolving codebases. This survey examines peer-reviewed literature published between 2020 and 2025, synthesizing empirical findings from fifteen major methodological approaches in ML-driven software maintenance. Our critical analysis encompasses deep learning architectures (CNNs, LSTMs, Transformers, and Graph Neural Networks) applied to vulnerability detection, defect prediction, security patch identification, and code clone detection. The survey reveals that contemporary approaches achieve accuracy rates spanning 79.6% to 95.1%, yet persistent challenges remain regarding model generalizability across projects, dataset imbalance, adversarial robustness, and interpretability of security-critical predictions. Notably, transformer-based models, coupled with Code Property Graphs (CPGs) and graph neural networks, demonstrate superior scalability and semantic preservation compared with recurrent approaches. This work identifies critical gaps between laboratory performance and real-world deployment scenarios, emphasizing the necessity for principled data collection methodologies, standardized evaluation protocols, and hybrid human-in-the-loop frameworks. Ethical and legal implications surrounding automated security decision-making and privacy-preserving model deployment are discussed. Concrete future research directions address cross-project transferability, explainable AI integration, and resilience against adversarial manipulation—all imperative for trustworthy, production-grade ML systems in security-critical software maintenance contexts.