Federated Machine Learning Architecture for Privacy Preserving Software Quality Assessment

Abstract

Software organizations often need shared quality models but cannot exchange private source code, defect records, or development histories. This study presents a federated machine learning architecture for privacy-preserving software quality assessment across distributed repositories. Six participating organizations trained local models using complexity, code churn, testing history, defect frequency, ownership, and maintainability indicators. Protected model updates were clipped, masked, and securely aggregated to create a global model without transferring raw repository data. The architecture was evaluated across repeated communication rounds using accuracy, precision, recall, F1- score, and AUC. All measures improved steadily during early training and became stable after convergence. The final global model reached 94.4% accuracy, 93.7% precision, 92.4% recall, 93.0% F1-score, and 96.3% AUC. These results show that distributed organizations can obtain reliable software quality predictions while maintaining local control of sensitive information. The architecture supports secure collaboration, defect-risk assessment, maintenance planning, and responsible use of industrial software data effectively.